Visual training system and method for stomach tube

Abstract

The invention relates to the technical field of medical instruments and medical simulation training, discloses a stomach tube visual training system and method, and aims to solve the problems that an existing training system is low in visual precision, lacks neurology special pathological scene modeling, and cannot intelligently recognize operation errors and provide personalized feedback. The method comprises the following steps: constructing a high-fidelity anatomical digital twin integrating hyperpharyngeal reflex and esophageal spasm simulation; through fusion sensing of the fiber bragg grating and the inertial measurement unit, the pose and mechanical parameters of the stomach tube are collected in real time, and advanced training tasks are dynamically recommended. The system comprises a high-fidelity anatomical model module, a multi-modal sensing fusion module, a visual rendering module, an intelligent error recognition module and a personalized training feedback module. According to the technical scheme, the authenticity of training, the accuracy of operation feedback and the individuation level of skill improvement are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical devices and medical simulation training, and particularly relates to a gastric tube visualization training system and method. BACKGROUND

[0002] With the continuous development of clinical medical education and skill training, the demand for simulation training of neurology-related operations is increasing. Gastric tube placement, as a basic but technically demanding clinical operation, is widely used in the nursing of neurology patients with dysphagia, consciousness disorders and nutritional support. This operation requires accurate judgment of the anatomical path and avoidance of risks such as misentry into the airway, which puts high demands on the spatial perception and hand-eye coordination of the operator. Traditional training methods mainly rely on theoretical instruction combined with simple models or live demonstrations, and lack real-time visual feedback and dynamic path guidance during the operation process, making it difficult to effectively improve the comprehensive judgment ability of beginners in terms of gastric tube trajectory, anatomical landmark recognition and resistance feedback.

[0003] Among them, the gastric tube visualization training technology aims to simulate the real human anatomical structure and dynamically present the tube placement process through image or sensing means to assist students in mastering the operation essentials. Existing training systems mostly use static anatomical models combined with external cameras or simple sensors, which can only provide limited visual reference or rough position prompts, and cannot achieve high-precision, low-delay visualization of the gastric tube's running state in complex anatomical cavities, nor can they provide intelligent identification and immediate intervention mechanisms for operation errors such as misentry into the trachea or excessive bending.

[0004] In the prior art, some high-end simulation systems attempt to introduce endoscopes or electromagnetic positioning technology to enhance visualization, but generally have problems such as high equipment cost, low system integration, high operation complexity, etc., making it difficult to popularize in basic teaching units. At the same time, most systems do not model the specific pathological features of neurology patients (such as hyperactive pharyngeal reflex and esophageal spasm), resulting in a disconnect between training content and real clinical scenarios. In addition, existing solutions generally lack the ability to structure the collection and analysis of student operation data, and cannot form personalized skill assessment and advanced training paths. Therefore, under the premise of ensuring system practicality and cost control, there is an urgent need for a gastric tube visualization training system and method that can deeply integrate anatomical visualization, intelligent error identification, and personalized training feedback. SUMMARY

[0005] The present application aims to provide a gastric tube visualization training system and method, which can effectively solve the problems in the background art. The existing gastric tube placement training system generally has low visualization accuracy, lacks modeling of neurology-specific pathological scenarios, cannot achieve intelligent identification and real-time intervention of operation errors, and lacks a personalized training feedback mechanism, resulting in a training effect that is disconnected from real clinical scenarios and difficulty in effectively improving the spatial perception ability, anatomical landmark recognition ability, and operation fault tolerance of trainees.

[0006] The technical solution of the present application is a gastric tube visualization training system, which comprises the following components: a high-fidelity anatomical model module for constructing complete upper digestive tract and respiratory tract anatomical structures including the pharynx, larynx, trachea, esophagus, and stomach, and integrating typical neurology pathological features, including a pharyngeal reflex hyperactivity simulation unit and an esophageal spasm simulation unit; a multi-modal sensor fusion module for real-time acquisition of position, posture, bending degree, and contact force data of the gastric tube during its travel in the model; a visualization rendering module for generating a three-dimensional dynamic travel trajectory of the gastric tube in the anatomical cavity based on the data of the multi-modal sensor fusion module, and superimposing anatomical landmarks and risk area identification; an intelligent error identification module for real-time detection and classification of operation errors such as gastric tube misentry into the trachea, excessive bending, and abnormal pushing resistance based on pre-set clinical operation specifications and anatomical constraints; a personalized training feedback module for recording and structuring the operation data of trainees, generating a skill evaluation report based on operation error types, completion time, and path deviation degree indicators, and dynamically recommending advanced training tasks.

[0007] The high-fidelity anatomical model module is formed by layer-by-layer casting of flexible silicone material, with a micro piezoelectric ceramic array embedded in the pharynx and esophagus wall for generating a local force feedback of 0.5 to 2 Newton to simulate the muscle contraction resistance when the pharyngeal reflex is hyperactive, and a shape memory alloy wire embedded in the middle esophagus with a phase transition temperature set to 42℃, which generates a 15% radial contraction rate within 0.5 seconds after being energized to simulate esophageal spasm, and the radial contraction rate is gradient-distributed to match the real muscle contraction pattern.

[0008] The multi-modal sensor fusion module includes a composite sensing structure of a fiber Bragg grating sensor array and an inertial measurement unit, the fiber Bragg grating sensors are distributed at an interval of 5 mm along the axial direction of the gastric tube for measuring local strain and inverting bending curvature, and the inertial measurement unit outputs three-axis acceleration and angular velocity data at a frequency of 100 Hz, the data of the fiber Bragg grating sensor and the inertial measurement unit are fused through an extended Kalman filtering algorithm to solve the six-degree-of-freedom pose of the gastric tube tip in the global coordinate system, and the pose solving error is controlled within ±1.2 mm and ±2.5 degrees.

[0009] The visualization rendering module employs volume rendering technology accelerated by a graphics processing unit to convert computed tomography or magnetic resonance imaging segmentation data of the anatomical model into a three-dimensional texture volume. The gastric tube trajectory is presented by Bézier curve fitting, and its control points are determined by least squares optimization of discrete pose points output by the multimodal sensing fusion module. The root mean square of the fitting residual is less than 0.8 mm. The positions of key anatomical landmarks are automatically extracted by semantic segmentation labels of the digital twin and mapped to the three-dimensional rendering scene.

[0010] The intelligent error recognition module incorporates a dual-channel discrimination mechanism based on rule-based and machine learning principles. The rule-based channel sets rigid constraints based on prior anatomical knowledge, including: "If the Z-coordinate of the gastric tube tip is lower than the glottic plane and the X-coordinate deviates from the midline by more than 10 mm, it is determined as tracheal intubation," and "If the local curvature exceeds the threshold k=0.15..." If the input is "excessive bending" and "the push resistance suddenly increases by more than twice the baseline value, it is judged as abnormal resistance", the machine learning channel uses a lightweight convolutional neural network, the input is a multi-dimensional physical parameter time window of 500 milliseconds, and the output is the error type probability distribution.

[0011] The lightweight convolutional neural network contains three convolutional layers and one fully connected layer, with the number of parameters controlled within 128 kilobytes. Its input multidimensional physical parameter time window includes the rate of change of spatial coordinates of the gastric tube tip, the rate of change of Euler angles, the local curvature sequence, and the tube wall contact pressure sequence. The convolutional layers are used to extract time-series features related to specific error patterns, and the fully connected layer maps the extracted features to the probability distribution of tracheal mis-entry, excessive bending, or tissue damage.

[0012] The personalized training feedback module defines the path deviation index as the Hausdorff distance between the trainee's actual gastric tube trajectory and the standard safe path, calculated using the following formula:

[0013] Where P is the set of points on the trainee's operation trajectory, and Q is the set of points on the standard path; the operation time efficiency index is defined as the ratio of the standard operation time to the actual operation time, and the value ranges from 0 to 1; the skill assessment report integrates the path Hausdorff distance deviation, the operation time efficiency index and the frequency of error occurrence, and uses a weighted scoring model to generate a quantitative score of 0 to 100.

[0014] The formula for calculating the total score in the weighted scoring model is: Total Score = ×(1-Path Deviation Normalization)+ ×Operation Time Efficiency Index+ ×(1-error frequency normalization), where , , The weighting coefficients are set according to clinical importance. The frequency of errors is statistically weighted according to the severity of the error, with one instance of tracheal intrusion having a higher weight than one instance of slight excessive bending.

[0015] A method for visualizing gastric tube training includes: Step S110: Initialize the high-fidelity anatomical model module and load the three-dimensional anatomical digital twin of the upper gastrointestinal and respiratory tracts containing typical pathological features of neurology. Step S120: The multi-dimensional physical parameters of the gastric tube during its movement within the model are collected in real time through the multi-modal sensing fusion module. The multi-dimensional physical parameters include spatial coordinates, Euler angles, local curvature, and tube wall contact pressure. In step S130, the visualization rendering module uses the ray projection and depth buffer fusion algorithm based on the multi-dimensional physical parameters to render the three-dimensional dynamic trajectory of the gastric tube in the anatomical cavity in real time, and highlights key anatomical landmarks such as the piriform fossa, epiglottic valley, esophageal inlet and tracheal carina in the display interface, while marking the high-risk area of ​​the trachea entering the high-risk area with red semi-transparent areas. Step S140: The intelligent error recognition module compares the multidimensional physical parameters with preset clinical operation safety thresholds. If it detects that the gastric tube tip has entered the tracheal region or that the local curvature exceeds the threshold, the module will detect the error. If the pushing resistance suddenly increases by more than twice the baseline value, it is judged as an operation error and a graded alarm is triggered. In step S150, the personalized training feedback module stores the data of the entire process of this operation in a structured manner, calculates the path Hausdorff distance deviation, operation time efficiency index and error frequency, generates a quantitative skill assessment report, and dynamically adjusts the pathological complexity and feedback prompt intensity of the next training session based on the trainee's historical performance.

[0016] The graded alarms trigger different levels of intervention based on the type and severity of the error. Accidental tracheal intubation triggers the highest level alarm, which includes a red flashing screen, a red gastric tube, a high-risk area warning, a rapid alarm sound, and simulated gastric tube vibration feedback, and suspends the training process. Excessive bending or slight abnormal resistance triggers a medium level alarm, which includes a yellow warning bar at the edge of the screen and prompts, allowing the trainee to continue operating for self-correction.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. By integrating neurological pathological features (such as hyperpharyngeal reflex and esophageal spasm) into a high-fidelity anatomical model, the consistency between the training scenario and the real clinical situation is significantly improved, enabling trainees to effectively cope with the special operational challenges of neurological patients in a simulated environment.

[0018] 2. By adopting FBG and IMU multimodal sensing fusion technology, high-precision (pose error ≤ ±1.2mm) and low-latency (refresh rate ≥100Hz) dynamic capture of the gastric tube's movement in complex cavities was achieved, providing a reliable data foundation for visualization and error identification.

[0019] 3. An intelligent error recognition mechanism integrating rules and machine learning was constructed, which can detect and classify key operational errors such as tracheal intrusion and excessive bending in real time, with an accuracy rate of over 95%, and supports hierarchical alarm intervention.

[0020] 4. A personalized training feedback system based on quantitative indicators such as Hausdorff distance deviation and time efficiency index has been established, which can automatically generate structured skill assessment reports and dynamically recommend advanced tasks, effectively supporting students' gradual skill improvement.

[0021] 5. While ensuring high precision and intelligence, the entire system significantly reduces hardware costs and operational complexity through modular design and embedded algorithm optimization, making it feasible for large-scale application in medical colleges and clinical training centers at all levels. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical architecture of the gastric tube visualization training system and method proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the intelligent error recognition module in this invention; Figure 3 This is a logical flow diagram of the multimodal sensing fusion module and the visualization rendering module in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the high-fidelity anatomical model module and the personalized training feedback module in this invention; Detailed Implementation

[0023] Please refer to Figures 1-4 This system aims to address the structural deficiencies of existing gastric tube placement training systems, such as low visualization accuracy, lack of modeling for neurological pathological scenarios, inability to intelligently identify and intervene in operational errors in real time, and lack of personalized training feedback mechanisms. These deficiencies lead to a disconnect between training effectiveness and real clinical situations, hindering the improvement of trainees' spatial perception, anatomical landmark recognition, and operational error tolerance. This system simulates high-risk gastric tube placement scenarios, requiring resident physicians to perform the procedure safely and accurately under complex pathological conditions. The proposed gastric tube visualization training system and method play a central role in this scenario, with its components working closely together to form a closed-loop training and evaluation system.

[0024] At the start of this training, the system first executes step S110 to initialize the high-fidelity anatomical model module and load a three-dimensional digital twin of the upper gastrointestinal and respiratory tracts containing typical pathological features of neurology. Specifically, the high-fidelity anatomical model module is a physical simulation device manufactured using a precise layered casting process with flexible silicone material, accurately replicating the real anatomical structure and tactile elasticity of the human pharynx, larynx, trachea, esophagus, and stomach. This model not only closely resembles the real human body in its macroscopic morphology but also meticulously simulates its microscopic biomechanical properties. For example, micro-piezoelectric ceramic arrays are strategically embedded within the walls of the pharynx and esophagus. These piezoelectric ceramic elements, through precise voltage control, can simulate the hyperreflexia symptoms commonly seen in neurological patients. When the gastric tube touches the pharyngeal wall, the piezoelectric ceramic array instantly undergoes local deformation, generating a force feedback of 0.5 to 2 Newtons, simulating the contraction resistance of the pharyngeal muscles, allowing the trainee to experience a realistic sense of resistance when the gastric tube is advanced. Furthermore, shape memory alloy wires are cleverly embedded within the model in the middle of the esophagus. These alloy wires, controlled precisely by a microcontroller integrated within the model, undergo a phase transition and contraction to simulate the pathological state of esophageal spasm. The phase transition temperature of this shape memory alloy wire is set at 42 degrees Celsius. When a preset current is applied, it can generate a radial contraction rate of up to 15% within an extremely short time of 0.5 seconds, simulating a sudden narrowing of the esophageal lumen. This radial contraction rate change is not uniform but exhibits a gradient distribution, simulating real muscle contraction patterns and ensuring the simulation's realism and clinical relevance.

[0025] Simultaneously, the system loads a corresponding three-dimensional anatomical digital twin of the upper gastrointestinal and respiratory tracts at the digital level. This digital twin is constructed based on high-resolution computed tomography and magnetic resonance imaging data, containing high-precision three-dimensional mesh models of all key anatomical structures, such as the pyriform fossa, vallecula epiglottis, esophageal inlet, and tracheal carina. Each anatomical structure is assigned a unique semantic segmentation label and its precise spatial coordinates and geometric attributes in the global coordinate system are stored. The digital twin also contains parametric models of typical neurological pathological features. For example, the pharyngeal muscle contraction model corresponding to the hyperpharyngeal reflex simulation unit has parameters including contraction intensity coefficient and duration; the esophageal lumen narrowing model corresponding to the esophageal spasm simulation unit has parameters including narrowing initiation location, narrowing length, maximum radial contraction rate of narrowing, and duration. During initialization, the system selects a digital twin configuration file containing specific pathological features from a preset training case library, such as a patient model with moderate hyperpharyngeal reflex and mild esophageal spasm. The loading process involves importing these digital models and their associated pathological parameters into the memory of the visualization rendering module and initializing the rendering pipeline. This ensures the integrity of the geometric data structure of all anatomical landmarks and risk areas, providing a foundation for subsequent real-time rendering and error identification. After initialization, the system performs a self-check to confirm the consistency between the physical model and the digital twin, such as verifying the matching degree between the physical model's sensor feedback and the expected response of the digital model, ensuring the accuracy of the simulation environment. If deviations are found, the system will trigger a calibration procedure, such as fine-tuning the driving current of the piezoelectric ceramic array and shape memory alloy wire, to ensure the accuracy of the simulation.

[0026] Next, the training enters the core operation phase. The system executes step S120, which uses a multimodal sensor fusion module to collect multidimensional physical parameters of the gastric tube in real time as it moves within the model. The core component of this module is a composite sensing structure combining a fiber Bragg grating sensor array and an inertial measurement unit. These two elements work together to provide comprehensive physical state information of the gastric tube. The fiber Bragg grating sensor array is precisely integrated inside the flexible wall of the simulated gastric tube, uniformly distributed along the tube's axis at fixed intervals of 5 mm, forming a linear array. Each fiber Bragg grating sensor can sense local strain. When the gastric tube bends, the grating period inside the fiber changes slightly, causing a shift in the center wavelength of the reflected spectrum. By monitoring this wavelength shift, the system can accurately invert the local strain of the gastric tube at various points, thereby calculating the curvature of the tube. The raw spectral data collected by the sensors is demodulated at a high speed of 200 times per second using a fiber optic demodulator, converting it into wavelength shift values.

[0027] Meanwhile, a miniature inertial measurement unit (IMU) is integrated near the tip of the gastric tube, outputting triaxial acceleration and angular velocity data at a frequency of 100 times per second. These data reflect the linear and rotational motion of the gastric tube tip in space. The IMU data undergoes preliminary filtering and noise suppression by a microcontroller to eliminate high-frequency noise and environmental interference. To overcome the limitations of a single sensor type and improve overall measurement accuracy, the system employs an extended Kalman filter (EPF) algorithm to fuse the local curvature data output from the fiber Bragg grating sensor array with the triaxial acceleration and angular velocity data output from the IMU. The EPF algorithm uses the kinematic model of the gastric tube and the sensor noise model to predict the current state of the gastric tube, and then corrects the prediction by incorporating the latest sensor measurements. This algorithm effectively handles the nonlinear characteristics of the sensor data and provides optimal pose estimation in noisy environments. The fused data is used to calculate the six-degree-of-freedom pose of the gastric tube tip in the global coordinate system, namely its precise three-dimensional spatial coordinates (X, Y, Z axes) and three Euler angle attitudes (pitch, roll, and yaw). By carefully designing algorithm parameters and models, the pose calculation error is strictly controlled within ±1.2 mm and ±2.5 degrees, ensuring high-precision perception of the spatial position and orientation of the gastric tube.

[0028] In addition, thin-film pressure sensors are integrated on the surface of the gastric tube to collect real-time contact pressure data between the tube and the lumen wall. These pressure sensors sample 100 times per second and are synchronized with pose data. The collected multidimensional physical parameters, including the spatial coordinates of the gastric tube tip, Euler angle orientation, local curvature data distributed along the gastric tube axis, and tube wall contact pressure data, are transmitted to the main processing unit via a high-speed serial communication interface. The data transmission protocol uses a custom binary format to maximize transmission efficiency and reduce latency. After receiving the data, the main processing unit performs timestamp alignment to ensure consistency of all sensor data on the timeline, laying a solid foundation for subsequent visualization rendering and error identification. The data flow is from the built-in sensors in the gastric tube to the local signal processing unit, and then transmitted via wired or wireless means to the main controller of the multimodal sensor fusion module. The main controller performs extended Kalman filter fusion and finally outputs a multidimensional physical parameter stream in a unified format.

[0029] After acquiring high-precision multidimensional physical parameters of the gastric tube, the system immediately executes step S130. Based on these multidimensional physical parameters, the visualization rendering module uses a ray casting and depth buffer fusion algorithm to render the three-dimensional dynamic trajectory of the gastric tube within the anatomical cavity in real time. Key anatomical landmarks such as the piriform fossa, vallecula epiglottis, esophageal inlet, and tracheal carina are highlighted on the display interface, while high-risk areas for tracheal misalignment are marked with red semi-transparent areas. Specifically, the visualization rendering module employs volume rendering technology accelerated by a graphics processing unit. First, the original computed tomography or magnetic resonance imaging segmentation data of the upper gastrointestinal tract-respiratory tract three-dimensional anatomical digital twin loaded during the initialization phase is preprocessed and converted into optimized three-dimensional texture volumes. These texture volumes store density information and color attributes of different tissue types (such as bone, soft tissue, and air cavities), enabling the high-fidelity presentation of the internal details of anatomical structures through volume rendering technology.

[0030] The rendering of the 3D dynamic trajectory of the gastric tube is a multi-stage process. The discrete gastric tube pose points output by the multimodal sensing fusion module, including spatial coordinates and orientation information, are first processed using a Bézier curve fitting algorithm. Specifically, these discrete points are used as control points for the Bézier curve, or the control points are determined through least-squares optimization, ensuring that the fitted curve smoothly and accurately represents the actual shape and curvature of the gastric tube. The root mean square of the residuals in this fitting process is strictly controlled to be less than 0.8 mm, ensuring a high degree of consistency between the rendered trajectory and the actual physical shape of the gastric tube. The fitted Bézier curve, as a geometric model, is then fed into the rendering pipeline.

[0031] The fusion algorithm of ray casting and depth buffering is the core of achieving accurate interactive rendering of the gastric tube and anatomical cavity. Ray casting is used to volumetrically render the anatomical cavity, simulating light penetration through volumetric data and calculating the final color of each pixel based on tissue density and color attributes. Depth buffering is used to resolve the occlusion relationship between the gastric tube and the cavity, ensuring that the gastric tube is correctly rendered inside the cavity without any penetration. The geometry of the gastric tube is rendered as a semi-transparent object with specular reflectivity to clearly observe its position and curvature within the cavity. The rendering refresh rate is optimized to over 60 frames per second to ensure that the operator can perceive smooth, lag-free real-time dynamic feedback.

[0032] In the display interface, the location information of key anatomical landmarks such as the piriform fossa, vallecula, esophageal inlet, and tracheal carina is automatically extracted from the semantic segmentation tags of the digital twin and precisely mapped onto the 3D rendered scene. These landmarks are highlighted, for example, through outlines, the use of contrasting colors, or flashing effects, to guide trainees' attention to key anatomical reference points during gastric tube placement. The color and brightness of the highlighted markers can be dynamically adjusted according to the trainee's focus or training stage; for example, as the gastric tube approaches a key area, the landmarks in that area become more prominent.

[0033] More importantly, to warn trainees of potentially dangerous operations, the system dynamically marks high-risk areas for tracheal tube ingress with a red semi-transparent area. The geometric boundaries of this high-risk area are pre-defined in the digital twin, based on the anatomical critical judgment area for tracheal tube ingress. When the tip of the gastric tube approaches or enters this area, the red semi-transparent display of the area gradually intensifies, possibly accompanied by a pulse effect, thus providing an immediate visual alert to the operator. This visual feedback mechanism greatly enhances trainees' perception of operational risks, helping them establish correct spatial judgment and operating habits. The data flow is as follows: the gastric tube pose data output from the multimodal sensor fusion module is fitted with a Bézier curve and then sent to the graphics processing unit along with pre-loaded anatomical digital twin data. The graphics processing unit performs volume rendering and geometric rendering, and finally outputs the rendering results to a high-resolution display.

[0034] Subsequently, the system enters the intelligent judgment and early warning stage, executing step S140. The intelligent error recognition module compares the multidimensional physical parameters with the preset clinical operation safety threshold. If it detects that the gastric tube tip has entered the tracheal region or the local curvature exceeds the threshold, the system will detect the error. If the pushing resistance suddenly increases by more than twice the baseline value, it is judged as an operational error and a tiered alarm is triggered. The intelligent error recognition module has a built-in dual-channel discrimination mechanism based on rule-based and machine learning integration to ensure the accuracy, real-time performance, and robustness of error recognition.

[0035] The rule-based approach primarily relies on strict prior anatomical knowledge and clinical operating procedures to set rigid constraints. For example, regarding the assessment of tracheal aspiration, the system establishes a clear geometric rule: "If the Z-coordinate of the gastric tube tip in the global coordinate system is lower than the glottic plane and the X-coordinate deviates from the midline by more than 10 mm, it is considered tracheal aspiration." The glottic plane and midline positions are precise geometric references extracted beforehand from the anatomical data of the digital twin. The glottic plane is typically defined as the horizontal plane at the lower edge of the cricoid cartilage, while the midline is defined as the midsagittal plane between the trachea and esophagus. This 10 mm deviation threshold is determined based on clinical practice and biomechanical studies, taking into account the actual dimensions of the gastric tube diameter and the tracheal inlet. Another key rule concerns the detection of excessive gastric tube curvature; the system sets a local curvature threshold k=0.15. This threshold indicates that if the bending radius of the gastric tube at any point is less than approximately 6.67 mm (i.e., the reciprocal of the radius is greater than 0.15), it is considered excessively bent, which may lead to tangling, tissue damage, or difficulty in advancement. This threshold is determined based on the material properties of the simulated gastric tube and the clinically safe bending limits of the tube. Furthermore, for the detection of abnormal pushing resistance, the system has a rule that triggers a "sudden increase in pushing resistance exceeding twice the baseline value." The baseline value is determined by analyzing the average contact pressure and tip thrust when the gastric tube is advanced in a normal, resistance-free path. When the data from the pressure sensor feedback or physical thrust sensor on the simulated gastric tube suddenly increases to more than twice the baseline value within a short period (e.g., 200 milliseconds), it indicates that the gastric tube may have encountered anatomical obstacles, luminal narrowing, or is resisting the pharyngeal reflex, thus triggering an abnormal resistance alarm. These rules are calibrated and validated based on expert knowledge and extensive clinical data.

[0036] The machine learning channel employs a lightweight convolutional neural network (CNN) model to handle more complex and subtle operational error patterns. The input to this CNN is a 500-millisecond time window of multidimensional physical parameters, including the rate of change of spatial coordinates at the gastric tube tip, the rate of change of Euler angles, the local curvature sequence, and the tube wall contact pressure sequence. The network structure comprises three convolutional layers and one fully connected layer, with the number of parameters strictly controlled to within 128 kilobytes, enabling real-time and efficient operation on embedded processors without relying on high-performance computing servers. The convolutional layers extract features associated with specific error patterns from the time-series data, such as the specific posture change sequence of the gastric tube before tracheal aspiration and the cumulative pattern of contact pressure before excessive bending. The fully connected layer maps the extracted features to probability distributions for different error types, such as the probability of tracheal aspiration, excessive bending, and tissue damage. This model, trained on a large amount of simulated operational data, learns and identifies subtle operational error patterns that are difficult for the regular channel to capture.

[0037] When any channel detects an error, the system triggers tiered alarms based on the error type and severity. For example, tracheal intubation is classified as a top-level error, immediately triggering strong visual (flashing red screen, red gastric tube, high-risk area warning) and audible (rapid alarm sound) alarms, possibly accompanied by simulated gastric tube vibration feedback. The training process is paused, forcing the trainee to review the error. Excessive bending or slight resistance abnormalities may trigger medium-level alarms, such as a yellow warning bar appearing at the screen edge with a prompt message, but allowing the trainee to continue operating, aiming to guide self-correction. All error events, including error type, occurrence time, gastric tube pose, physical parameters, and alarm level, are recorded in real time, forming a detailed error log to provide data for subsequent feedback evaluation. The data flow is a multi-dimensional physical parameter stream output from the multimodal sensor fusion module, concurrently input into the rule engine and convolutional neural network model. The judgment results from the two channels are fused by the error decision logic module, and tiered alarm signals are output to the rendering and feedback modules according to preset alarm strategies.

[0038] Finally, to ensure continuous improvement and personalized development of training effectiveness, the system executes step S150. The personalized training feedback module stores the entire process data of this operation in a structured manner, calculates the path Hausdorff distance deviation, operation time efficiency index, and error frequency, generates a quantitative skill assessment report, and dynamically adjusts the pathological complexity and feedback intensity of the next training session based on the trainee's historical performance. The training feedback module is responsible for a comprehensive analysis and evaluation of each of the trainee's operations.

[0039] First, all data from this operation, including the raw multidimensional physical parameters collected in step S120, the error identification log generated in step S140, and the start and end timestamps of the operation, are structured and stored in a training database located on a local server or in the cloud. The data storage employs an efficient columnar storage format to ensure the performance of data querying and analysis. Stored data fields include a unique student identifier, training session identifier, date and time, a sequence of 3D trajectory points at the gastric tube tip, a sequence of gastric tube postures, a sequence of local curvature, a sequence of tube wall contact pressure, error type, error occurrence time, and alarm level. This data is anonymized before storage to protect student privacy.

[0040] Next, the system calculates the core skills assessment indicators.

[0041] Hausdorf distance deviation is a key indicator measuring the degree of deviation between the trainee's actual gastric tube trajectory and the standard safe path. Its calculation formula is:

[0042] in, For Hausdorf distance, For point It is a set One of the points represents a specific location on the student's operational trajectory. For point It is a set One of the points represents a location on the standard path. Euclidean distance (geometric distance) between two points, that is, the distance between two points in space. and The straight-line distance between them For a fixed point Find it on the standard path The distance from the nearest point, i.e., "from the student's trajectory point" The shortest distance to the nearest point on the standard path. This represents the "shortest distance" from all student trajectory points to the standard path. It can be understood as: "Which point in the student trajectory is furthest from the standard path?" This value reflects the maximum deviation of the student trajectory from the standard. For all standard path points, calculate the furthest "nearest distance" to the student's trajectory in reverse: Which point on the standard path is furthest from the student's trajectory? This indicates whether the "standard path is completely covered". To take the larger of the two values, the final Hausdorff distance is the maximum of the two directional deviations mentioned above, ensuring bidirectional comparison; This represents the set of trajectory points of the gastric tube tip during this procedure performed by the trainees. This represents a predefined set of standard safe pathway points generated by experienced experts for the current pathological scenario. The standard safe pathway is determined by comprehensively considering optimal anatomical path, minimum risk of tissue damage, and highest operational efficiency. This indicator quantifies the accuracy and smoothness of the trainee's operation; a smaller value indicates that the trajectory is closer to the standard pathway. Path deviation is not only calculated based on geometric distance but also takes into account the posture of the gastric tube in different anatomical regions. For example, trajectory deviation in the vallecula region and trajectory deviation at the tracheal inlet are assigned different weights to reflect their clinical importance.

[0043] The operation time efficiency index is defined as the ratio of the standard operation time to the trainee's actual operation time, ranging from 0 to 1. The standard operation time is determined based on the average operation time of senior experts in the same pathological scenarios. If the trainee's actual operation time is less than or equal to the standard operation time, the efficiency index approaches 1; if the actual operation time significantly exceeds the standard time, the index will decrease accordingly. This index assesses the trainee's proficiency and efficiency; excessively low time efficiency may indicate a lack of fluency, hesitation, or unfamiliarity with anatomical structures.

[0044] Error frequency counts the number of times various errors (such as tracheal intubation, excessive bending, and abnormal pushing resistance) occur during the operation, and weights them according to their severity. For example, the weight of a single instance of tracheal intubation is much higher than that of a single instance of slight excessive bending.

[0045] Based on the above indicators, the system generates a comprehensive quantitative skills assessment report. This report uses a weighted scoring model to comprehensively calculate indicators such as path Hausdorff distance deviation, operation time efficiency index, and error frequency, generating a total score from 0 to 100. For example, the total score calculation formula might be: Total Score = ×(1-Path Deviation Normalization)+ ×Operation Time Efficiency Index+ ×(1-error frequency normalization), where , , The weighting coefficients are set according to clinical importance. The report not only includes the total score, but also details the specific values ​​of each indicator, the detailed time and type of error, a comparison chart of the visualized trajectory and the standard path, and targeted textual feedback. For example, if a trainee makes multiple improper operations in the pyriform fossa, the report will indicate "Hesitation in the operation in the pyriform fossa area; it is recommended to strengthen the understanding of pharyngeal anatomy and the control of gastric tube posture."

[0046] To provide a truly personalized training experience, the feedback module dynamically adjusts the pathological complexity and feedback intensity of the next training session based on the trainee's historical performance. For example, if a trainee consistently performs poorly in a specific pathological scenario (such as esophageal spasm) in recent training sessions, the system will automatically recommend a training scenario with a higher severity of esophageal spasm and may increase the force feedback intensity of the simulated gastric tube. Simultaneously, the feedback intensity is also dynamically adjusted: for beginners, the system provides more direct and frequent warnings and prompts in the early stages; while for trainees with some foundation but still progressing, the system gradually reduces direct prompts, encouraging them to identify and correct errors through self-awareness and judgment, thereby cultivating their ability to solve problems independently. This adjustment is based on the clustering analysis results of the trainee's historical operation data. The system can identify the trainee's weaknesses and strengths, thus recommending targeted training scenarios that best promote their skill improvement. For example, if the error pattern clustering results show that the trainee's main problem is "tracheal aspiration," the system will recommend a series of training scenarios specifically targeting precise pharyngeal guidance and airway protection.

[0047] The data flow involves receiving operational data from the intelligent error detection module and the multimodal sensor fusion module, cleaning and structuring the data, and storing it in the database. Then, the data analysis algorithm module calculates various evaluation indicators and integrates the results into the skills assessment report generation module. After the report is generated, it is displayed to the trainee through the user interface, and the evaluation results are simultaneously input into the training scenario dynamic adjustment algorithm to select appropriate pathological complexity and feedback intensity parameters for the next training session.

[0048] The gastric tube visualization training system and method of this invention significantly improves the realism, effectiveness and relevance of gastric tube placement training through its closed-loop mechanism of high-fidelity simulation, precise sensing, intelligent recognition and personalized feedback, providing unprecedented learning tools for neurology residents and medical students.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A gastric tube visualization training system, comprising: include: The high-fidelity anatomical model module is used to construct the complete upper digestive and respiratory tract anatomical structure, including the pharynx, larynx, trachea, esophagus and stomach, and integrates typical neurological pathological features, including a pharyngeal hyperreflexia simulation unit and an esophageal spasm simulation unit. The multimodal sensing fusion module is used to collect real-time data on the position, orientation, curvature, and contact force of the gastric tube as it travels within the model. The visualization rendering module is used to generate a three-dimensional dynamic trajectory of the gastric tube within the anatomical cavity based on the data from the multimodal sensing fusion module, and to overlay anatomical landmarks and risk area markers. The intelligent error recognition module is used to detect and classify errors such as accidental insertion of the gastric tube into the trachea, excessive bending, and abnormal pushing resistance in real time, based on preset clinical operation procedures and anatomical constraints. The personalized training feedback module records and stores trainees' operational data in a structured manner. It generates skill assessment reports based on operational error types, completion times, and path deviation indicators, and dynamically recommends advanced training tasks.

2. The gastric tube visualization training system of claim 1, wherein, The high-fidelity anatomical model module is made of flexible silicone material through layered casting. The pharynx and esophageal wall have built-in micro piezoelectric ceramic arrays to generate local force feedback of 0.5 to 2 Newtons when the gastric tube touches the pharyngeal wall to simulate the muscle contraction resistance when the pharyngeal reflex is hyperactive. A shape memory alloy wire is embedded in the middle of the esophagus. The phase transition temperature of the shape memory alloy wire is set to 42°C. After being energized, it generates a radial contraction rate of 15% within 0.5 seconds to simulate esophageal spasm. The radial contraction rate is gradient-distributed to match the real muscle contraction pattern.

3. The gastric tube visualization training system of claim 1, wherein, The multimodal sensing fusion module includes a composite sensing structure of a fiber Bragg grating sensor array and an inertial measurement unit. The fiber Bragg grating sensors are distributed along the axial direction of the gastric tube at 5 mm intervals to measure local strain and invert bending curvature. The inertial measurement unit outputs triaxial acceleration and angular velocity data at a frequency of 100 Hz. The data from the fiber Bragg grating sensors and the inertial measurement unit are fused using an extended Kalman filter algorithm to calculate the six-degree-of-freedom pose of the gastric tube tip in the global coordinate system. The pose calculation error is controlled within ±1.2 mm and ±2.5 degrees.

4. The gastric tube visualization training system of claim 1, wherein, The visualization rendering module uses volume rendering technology accelerated by graphics processing unit to convert the computed tomography or magnetic resonance imaging segmentation data of the anatomical model into a three-dimensional texture volume. The gastric tube trajectory is presented by Bézier curve fitting. Its control points are determined by the least squares method through the discrete pose points output by the multimodal sensing fusion module. The root mean square of the fitting residual is less than 0.8 mm. The positions of key anatomical landmarks are automatically extracted by the semantic segmentation labels of the digital twin and mapped to the three-dimensional rendering scene.

5. The gastric tube visualization training system of claim 1, wherein, The intelligent error recognition module is built-in based on rule and machine learning fusion dual-channel discrimination mechanism, rule channel sets hard constraint according to anatomical priori knowledge, including "gastric tube tip Z coordinate below glottis plane and X coordinate deviates from midline more than 10 millimeters, then determine as trachea misentry", "local curvature exceeds threshold k=0.15 Then determine as excessive bending" and "pushing resistance sudden increase exceeds baseline value 2 times, then determine as resistance anomaly", machine learning channel adopts light weight convolutional neural network, input is multi-dimensional physical parameter time sequence window of 500 milliseconds, output is error type probability distribution.

6. The gastric tube visualization training system of claim 5, wherein, The lightweight convolutional neural network contains three convolutional layers and one fully connected layer, with the number of parameters controlled within 128 kilobytes. Its input multidimensional physical parameter time window includes the rate of change of spatial coordinates of the gastric tube tip, the rate of change of Euler angles, the local curvature sequence, and the tube wall contact pressure sequence. The convolutional layers are used to extract time-series features related to specific error patterns, and the fully connected layer maps the extracted features to the probability distribution of tracheal mis-entry, excessive bending, or tissue damage.

7. The gastric tube visualization training system of claim 1, wherein, The personalized training feedback module defines the path deviation degree index as the Hausdorff distance between the actual gastric tube trajectory of the trainee and the standard safe path; the operation time efficiency index is defined as the ratio of the standard operation time to the actual operation time, and the value range is 0 to 1; the skill evaluation report integrates the path Hausdorff distance deviation, the operation time efficiency index and the error frequency, and generates a quantitative score of 0 to 100 by using a weighted scoring model.

8. The gastric tube visualization training system of claim 7, wherein, The total score calculation formula of the weighted scoring model is total score = 0.5 × (1-path deviation degree normalization) + 0.3 × operation time efficiency index + 0.1 × (1-error frequency normalization), wherein , , , , , is a weight coefficient set according to clinical importance, and the error occurrence frequency is weighted and counted according to the error severity, and the weight of one tracheal misentry is higher than that of one slight overbending.

9. A training method applied to the gastric tube visualization training system according to any one of claims 1 to 8, characterized in that, The method comprises the following steps: Step S110, initializing the high-fidelity anatomical model module, loading the upper digestive tract-respiratory tract three-dimensional anatomical digital twin containing typical pathological characteristics of neurology; Step S120, real-time acquisition of multi-dimensional physical parameters of the gastric tube during the running process in the model by the multi-modal sensing fusion module, wherein the multi-dimensional physical parameters include spatial coordinates, Euler angle posture, local curvature and tube wall contact pressure; Step S130, the visualization rendering module adopts the light projection and depth buffer fusion algorithm to render the three-dimensional dynamic trajectory of the gastric tube in the anatomical cavity in real time based on the multi-dimensional physical parameters, and highlights the key anatomical landmarks of the piriform fossa, the arytenoid valley, the esophageal inlet and the tracheal carina in the display interface, and identifies the tracheal misentry into the high-risk area with a red translucent area; Step S140, the intelligent error recognition module compares the multi-dimensional physical parameters with the preset clinical operation safety threshold. If it is detected that the gastric tube tip coordinate enters the tracheal region, the local curvature exceeds the threshold k=0.15 , or the pushing resistance suddenly increases by more than twice the baseline value, it is determined that an operation error has occurred, and a hierarchical alarm is triggered. Step S150, the individualized training feedback module stores the whole process data of this operation in a structured manner, calculates the path Hausdorff distance deviation, operation time efficiency index and error frequency, generates a quantitative skill evaluation report, and dynamically adjusts the pathological complexity of the next training and the intensity of the feedback prompt according to the historical performance of the student.

10. The gastric tube visualization training method of claim 9, wherein, The hierarchical alarm triggers different levels of intervention according to the error type and severity, and the tracheal misentry triggers the highest level of alarm, including screen red flashing, red gastric tube, high-risk area warning, urgent alarm sound and simulated gastric tube vibration feedback, and suspends the training process; the over-bending or slight resistance anomaly triggers the medium level alarm, including the screen edge yellow warning bar and prompt information, allowing the trainee to continue operating to correct himself.

Images