Artificial intelligence driven exoskeleton assisted multi-position radiotherapy fixation device

Abstract

The application discloses an artificial intelligence driven exoskeleton assisted multi-position radiotherapy fixing device, which comprises a bionic exoskeleton module, a multi-modal sensing positioning module, a six-degree-of-freedom driving execution module and a REAC-Net artificial intelligence control module; the exoskeleton module adopts a detachable bionic structure of a head and neck, a trunk and four limbs, supports multi-position fixing such as supine, prone, lateral recumbency, sitting and half-standing, and the like; the sensing positioning module fuses a ToF array, an IMU, a magnetic encoder and a pressure sensor, and collects position attitude and pressure data in real time; the driving execution module adopts a piezoelectric driver and a torque motor to realize sub-millimeter level adjustment; the REAC-Net module completes multi-modal data fusion, error solving, physiological motion prediction and dynamic compensation, and forms a closed loop of "initialization-sensing-decision-execution-optimization". The application can realize sub-millimeter level position fixing and intelligent dynamic compensation, significantly improves radiotherapy precision, comfort and safety, and is suitable for various radiotherapy scenes.

Description

Technical Field

[0001] This invention relates to the field of medical mechanical assistance technology, specifically to an artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device. Background Technology

[0002] In the field of tumor radiotherapy, patient positioning is a core prerequisite for ensuring treatment accuracy, avoiding accidental irradiation of normal tissues, and ensuring treatment efficacy. Currently, the mainstream clinical methods are passive fixation systems such as thermoplastic films, vacuum negative pressure pads, and expanding foam. Although these systems have been used for a long time, they are limited by outdated design concepts and inherent technical architectures, and cannot meet the stringent requirements of precision radiotherapy—especially stereotactic radiosurgery (SRS), stereotactic body radiotherapy (SBRT), and FLASH radiotherapy (FLASH-RT)—for sub-millimeter-level positional repeatability. Their core shortcomings are as follows: (1) Severe deficiency in postural adaptability and spatial adjustment ability Existing passive fixation devices are mostly standardized designs for standard supine and prone positions, lacking dedicated support and fixation structures adapted to special positions such as sitting and lateral decubitus. Their adjustment functions are limited to simple translation of the fixation base plate or knob-type fine-tuning of height, unable to achieve precise rotational adjustment around the three-dimensional X, Y, and Z axes. This necessitates the use of non-ideal irradiation angles or compromised fixation methods during non-standard positional treatments, increasing the radiation dose and risk of damage to normal tissues, limiting the optimization space of treatment plans, and failing to fully realize the advantages of precision radiotherapy.

[0003] (2) The stability of the body position decreases over time, and the repeatability between repetitions is poor. Traditional passive fixation devices suffer from unstable physicochemical properties: thermoplastic films are susceptible to irreversible shrinkage or relaxation due to temperature and humidity fluctuations, leading to a gradual decrease in their fit to the patient's body; vacuum negative pressure pads have a risk of chronic air leakage, and the microspheres inside can redistribute over time and with changes in patient position, resulting in weakened support and patient sinking. These issues cause cumulative positioning errors during fractionated treatments, failing to meet the long-term positional stability requirements of precise radiotherapy and severely impacting treatment accuracy and efficacy consistency.

[0004] (3) The fixation device interferes with the treatment radiation and image guidance. Most existing passive fixation devices contain metal fasteners, rigid support frames, or high-density polymer materials. When high-energy rays penetrate, they produce scattering, energy attenuation, and secondary electrons, resulting in uneven dose distribution in the target area and affecting the treatment effect. At the same time, they produce obvious metal artifacts in CT and CBCT image-guided images, which seriously interfere with the accurate delineation of the target area and organs at risk, reduce image registration accuracy, and further aggravate positioning errors and treatment risks.

[0005] (4) Lack of dynamic monitoring and real-time compensation capabilities for physiological movements Existing devices are rigid, passive fixation structures that can only achieve static body positioning and cannot cope with dynamic disturbances such as respiratory movements, muscle spasms, and involuntary micromovements. Clinically, auxiliary methods such as expanding the planned target volume (PTV) boundary or relying on respiratory gating systems are commonly used. The former sacrifices radiotherapy precision and increases the risk of radiation exposure to normal tissues, while the latter significantly increases treatment complexity, prolongs treatment time, and increases treatment costs. In addition, the algorithms of existing auxiliary monitoring systems are outdated, have poor adaptability to individual movement, and insufficient accuracy in predicting dynamic errors, making it impossible to achieve real-time compensation for dynamic positioning errors and meet the needs of precise radiotherapy for dynamic positioning control.

[0006] (5) The contradiction between patient comfort and fixation safety is prominent. Existing devices are designed with fixation and stability as their core objective, which can easily lead to excessive local contact pressure. This can cause pain, numbness, and even pressure injuries in patients during radiotherapy, potentially inducing involuntary movement and increasing positional errors. Furthermore, existing devices lack active monitoring and intelligent management mechanisms for contact pressure and the body's microenvironment. They cannot predict abnormal pressure distribution trends using algorithms and make adaptive adjustments in advance, making it difficult to achieve a balance between fixation stability and patient comfort, thus affecting patient treatment compliance and positional stability.

[0007] In summary, existing tumor radiotherapy positioning techniques have not escaped the limitations of static and passive design. They face significant constraints and irreconcilable contradictions in core dimensions such as positioning accuracy, adaptability to special positions, long-term stability, radiation compatibility, and patient comfort. Furthermore, they lack the core support of intelligent algorithms, making it impossible to overcome key technical bottlenecks such as dynamic positioning compensation and individualized adaptation. This has become a key shortcoming hindering a leapfrog improvement in radiotherapy accuracy, and a new positioning and control technology is urgently needed to solve this problem. Summary of the Invention

[0008] The present invention proposes an artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device, equipment and storage medium, which can at least solve one of the technical problems in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: An AI-driven exoskeleton-assisted multi-position radiotherapy fixation device includes a bionic exoskeleton module, a multimodal sensing and positioning module, a six-degree-of-freedom drive and execution module, and a REAC-Net (Radiotherapy Exoskeleton Adaptive Control Network) AI control module; the modules work together to achieve sub-millimeter-level radiotherapy positioning fixation driven by AI and assisted by the exoskeleton; The bionic exoskeleton module adopts a bionic joint topology structure with detachable and combinable head, neck, torso, and limbs, which can achieve rigid support and locking in multiple body positions such as supine, prone, lateral, sitting, and semi-standing. The multimodal sensing and positioning module includes a ToF sensor array on the treatment room ceiling, a posture sensing unit embedded in the exoskeleton joint, and a thin-film pressure sensor in the exoskeleton contact layer. Each sensor establishes a data connection with the REAC-Net artificial intelligence control module to collect real-time global spatial position of the treatment room, exoskeleton joint posture, body surface contact pressure, and patient physiological movement data, and synchronously transmits the multimodal heterogeneous data to the REAC-Net artificial intelligence control module. The six-degree-of-freedom drive execution module includes a linear piezoelectric actuator and a rotary torque motor. The six-degree-of-freedom drive execution module establishes a control connection with the REAC-Net artificial intelligence control module to receive control commands and drive the bionic exoskeleton module to complete three-dimensional translation and three-dimensional rotation adjustment, thereby achieving sub-millimeter-level body positioning and dynamic compensation. The REAC-Net artificial intelligence control module is used to perform dynamic causal fusion, six-degree-of-freedom body position error calculation, physiological motion prediction and feedforward compensation, and body surface pressure adaptive PID (Proportional-Integral-Derivative) optimization on the received multimodal data, and generate motion control commands to output to the six-degree-of-freedom drive execution module, forming a closed-loop optimization intelligent control link of "perception-decision-execution". The bionic exoskeleton module is made of carbon fiber reinforced PEEK composite material and low electron density zirconia ceramic bearings. Its metal parts meet size and spacing constraints, and the whole device follows the beam eye view (BEV) avoidance design, so that the device's radiation dose disturbance is lower than the clinical threshold.

[0010] As can be seen from the above technical solutions, in view of the systemic defects of existing tumor radiotherapy positioning techniques, such as static passivity, insufficient precision, and poor adaptability, this invention aims to provide an active, intelligent, multi-degree-of-freedom exoskeleton radiotherapy fixation device. This device integrates deep learning neural network algorithms to fundamentally solve the five core pain points of existing technologies. The specific problems solved and the corresponding technical solutions are as follows: (1) Solving the technical problem of lack of effective fixation and support in complex treatment positions: To address the shortcomings of existing fixation devices, which are only suitable for standard supine and prone positions and cannot meet the fixation requirements of special positions such as sitting and lateral decubitus, this invention adopts a multi-joint exoskeleton system designed with a human skeletal topology. It can autonomously deform and precisely lock posture according to the clinical radiotherapy positioning requirements, providing patients with multi-dimensional rigid support in multiple scenarios, including supine, prone, lateral decubitus, sitting, and even semi-standing positions. This ensures that any clinically required treatment position achieves the same fixation stability and sub-millimeter-level repeatability accuracy as the standard supine position, completely solving the technical challenge of difficult positioning during radiotherapy for lesions in special locations.

[0011] (2) Solve the technical problems of insufficient adjustment freedom and low positioning accuracy of traditional fixed devices: To address the limitations of existing fixation devices, which offer limited adjustment freedom and only allow for simple translation and minor height adjustments, resulting in insufficient positioning accuracy for precise radiotherapy, this invention employs an attitude sensing unit (integrating a 9-axis inertial measurement unit (IMU) and an absolute magnetic encoder) in conjunction with a high-precision drive system (linear piezoelectric actuator and rotary torque motor). This endows the exoskeleton system with six degrees of freedom (three translations and three rotations) in three-dimensional space, achieving sub-millimeter resolution and milliradian resolution. Furthermore, a deep learning neural network algorithm is introduced to precisely calculate and analyze multimodal sensor data, upgrading the traditional trial-and-error positioning method reliant on human experience to a data-driven, precise navigation-based intelligent positioning system. This significantly improves positioning efficiency and accuracy, meeting the stringent requirements of precision radiotherapy technologies such as SRS, SBRT, and FLASH.

[0012] (3) Solving the problem of decreased stability of body position during fractionated radiotherapy and interference with human physiological movement: To address the shortcomings of traditional fixation devices, such as the tendency for stability to decrease with each radiotherapy session and the inability to handle physiological micro-movements like respiratory movements and muscle tremors, this invention employs a collaborative technical strategy of "rigid exoskeleton frame + intelligent closed-loop control + deep learning algorithm" to ensure the stability of body positioning during radiotherapy from multiple dimensions: First, the exoskeleton body is made of carbon fiber composite material, whose excellent structural strength and fatigue resistance ensure no deformation of the mechanical structure over long-term use; second, an array of ToF sensors is deployed on the ceiling of the treatment room to collect the relative position information between the patient's body surface and the exoskeleton frame in real time, converting 3D point cloud data into a voxel network to achieve high-precision real-time monitoring of relative posture; third, relying on a deep learning network model, the system accurately predicts and dynamically compensates for the patient's respiratory movements, muscle tremors, and other physiological movement patterns, effectively suppressing endogenous positioning errors during radiotherapy and overcoming the technical bottlenecks of traditional algorithms' weak adaptability to individual movement characteristics and insufficient prediction accuracy, thus stably ensuring the accuracy of body positioning throughout the entire radiotherapy session and in each fraction.

[0013] (4) Solving the technical problem of severe physical interference of fixation devices with radiotherapy rays: To address the shortcomings of existing fixation devices, which are prone to radiation scattering, energy attenuation, and image artifacts caused by the materials used, thus interfering with radiotherapy dosage and image guidance accuracy, this invention adheres to the principle of prioritizing radiation compatibility throughout. Through multi-dimensional collaborative design involving material selection, structural optimization, and dosimetric verification, it minimizes the physical interference of the device with radiotherapy radiation. At the material level, high-quality substrates with low radiation disturbance, such as carbon fiber reinforced PEEK composite material (PEEK-CF30) and low electron density zirconia ceramic bearings, are selected to significantly reduce radiation scattering, energy attenuation, and secondary electron generation. At the structural level, a fusion attitude sensing unit (integrating a 9-axis IMU and an absolute magnetic encoder) is embedded in the exoskeleton joints to achieve high-precision positioning and attitude perception while avoiding redundant metal components. In terms of metal component control, the size of individual metal components is strictly limited to less than 3cm, and the minimum distance between components is no less than 15cm to prevent radiation scattering and dose disturbance caused by local metal accumulation. In terms of structural layout, the beam eye perspective (BEV) avoidance principle is strictly followed to optimize the installation posture and spatial arrangement of each functional component. In terms of dosimetric verification, a dual verification method combining Monte Carlo numerical simulation and actual dose measurement of physical phantoms is adopted to ensure that the additional dose error introduced by the device is controlled within 1%, fully meeting the dosimetric control standards and clinical application requirements for precision radiotherapy.

[0014] (5) Technical issues related to resolving the inherent contradiction between fixation stability and patient comfort: Existing radiotherapy fixation devices, in pursuit of precise positioning, often apply excessive pressure to the patient's local area, resulting in poor comfort and easily triggering unconscious body movements, thereby reducing the accuracy of positioning and treatment. This invention achieves a precise balance between positional stability and patient comfort through the synergistic effect of pressure sensors and deep learning algorithms: a thin-film pressure sensor is integrated into the exoskeleton contact layer to collect and dynamically report the contact pressure distribution between the patient's skin and the device in real time; relying on the PID parameter optimization module in the deep learning REAC-Net model, the device accurately predicts changes in position and the evolution trend of pressure distribution, automatically alarming when the pressure value exceeds a safe threshold and driving the support structure to adaptively adjust, thus preventing pressure-induced damage at the source; simultaneously, a micro-airflow circulation channel and a constant temperature control module are built into the contact interface to keep the patient's skin dry and maintain the temperature of the contact area within the human comfort range of 36-37℃, significantly improving the patient's tolerance to long-term radiotherapy and effectively avoiding unconscious body displacement caused by physical discomfort, further ensuring the stability and accuracy of positional fixation throughout radiotherapy.

[0015] Specifically, compared with the prior art, the present invention has the following advantages: The multi-degree-of-freedom exoskeleton radiotherapy fixation device provided by this invention integrates an artificial intelligence deep learning network algorithm model to innovatively construct a radiotherapy exoskeleton adaptive control network (REAC-Net). This network model integrates efficient multimodal data processing, dynamic causal reasoning, and online adaptive optimization capabilities, and is specifically designed for the high-precision body positioning control requirements of radiotherapy scenarios. Compared with existing passive radiotherapy fixation technologies and related products, it has the following outstanding advantages, comprehensively overcoming the shortcomings of existing technologies and fully demonstrating outstanding technological innovation and clinical application advantages: (1) Sub-millimeter level body positioning accuracy and repeatability accuracy were achieved. This invention, through the deep integration of multimodal sensors and high-precision driving devices, combined with the synergistic effect of the deep learning REAC-Net model and high-bandwidth closed-loop control strategy, can strictly control the overall positioning system error to the sub-millimeter level and maintain consistent accuracy in each fractionated treatment, completely solving the core pain points of low positioning accuracy and poor repeatability of fractionated positioning in existing technologies. Specifically, the deep learning algorithm can automatically learn the individual patient positioning characteristics and sensor data deviation patterns, continuously iteratively optimizing the data calculation accuracy, providing dual guarantees of hardware architecture and intelligent algorithms for reducing the planning target volume (PTV) boundary and improving target dose concentration. This invention is particularly suitable for ultra-high precision radiotherapy scenarios such as stereotactic radiosurgery (SRS), stereotactic body radiotherapy (SBRT), and FLASH radiotherapy (FLASH-RT), filling the gap in existing technologies that cannot meet the requirements of sub-millimeter positioning control.

[0016] (2) Possesses the ability to self-adapt to and fix body position in all positions. The exoskeleton structure of this invention completely breaks through the limitations of existing fixation devices that can only adapt to standard supine and prone positions. It can quickly switch between various treatment positions, such as supine, prone, lateral, sitting, and even semi-standing, and accurately lock postures within 5-10 minutes, flexibly adapting to the radiotherapy needs of special sites and special cases. At the same time, combined with the autonomous learning ability of the deep learning REAC-Net model to learn the human body's support needs under different positions, it can adaptively adjust the support points and support strength to achieve individualized support adaptation. This provides sufficient operational flexibility for multi-site and multi-angle combined irradiation treatment, effectively expanding the clinical application scenarios of radiotherapy technology and solving the technical problem that existing technologies cannot adapt to complex treatment positions.

[0017] (3) Active compensation is achieved by relying on the REAC-Net model to effectively suppress dynamic postural errors. This invention relies on the innovative deep learning REAC-Net model to accurately extract temporal motion features such as respiratory waveforms and muscle tremors from multimodal sensor data, achieving high-precision prediction of human physiological motion (prediction error ≤ 0.1 mm) and real-time dynamic compensation. Simultaneously, it coordinates with radiotherapy equipment to achieve synchronized beam output control, ensuring that the treatment radiation always accurately covers the target area. Compared to existing traditional motion prediction algorithms, this model can adapt to different patients' individual physiological motion patterns (such as respiratory rhythm and body movement characteristics), dynamically updating prediction parameters and compensation strategies online. It achieves precise suppression of dynamic errors without manual intervention, realizing a core technological leap from "static passive constraint" to "dynamic active stabilization" in radiotherapy positioning, completely solving the shortcomings of existing technologies in dealing with patient physiological motion interference.

[0018] (4) Coordinated optimization of materials and structure to ensure accurate execution of radiotherapy plans This invention prioritizes radiation compatibility throughout material selection and structural design. Core structural components utilize materials with minimal radiation dose disturbance and low linear attenuation coefficients (such as carbon fiber reinforced PEEK composites and low electron density zirconia ceramic bearings). Through dual verification using Monte Carlo numerical simulations and actual phantom dose measurements, the device's impact on radiotherapy dose distribution is far below the clinically permissible threshold, effectively avoiding radiation scattering, dose "cold spots," "hot spots," and image artifacts caused by metal components and high-density materials in traditional fixation devices. Simultaneously, by incorporating the deep learning REAC-Net model, the device can analyze the matching relationship between the radiotherapy gantry angle and the radiation field path in real time, automatically adjusting the exoskeleton posture to achieve radiation field avoidance, eliminating radiation field obstruction or dose abnormalities that may occur with traditional devices, and ensuring precise execution of the radiotherapy plan.

[0019] (5) Intelligent, modular and comfortable integrated design, simultaneously improving clinical efficiency and patient experience. This invention integrates multimodal sensor fusion and the deep learning REAC-Net model to construct a closed-loop positioning system that combines automatic error calculation, deviation trend prediction, and device-driven adaptive adjustment. This system can reduce manual positioning time by more than 30%, significantly improve positioning accuracy and the repeatability of treatment sessions, and effectively reduce the workload of medical staff. The model can generate personalized positioning support plans based on patient body shape characteristics and comfort feedback. It also uses a thin-film pressure sensor to monitor pressure distribution in real time, predict changes in position and pressure, and automatically alarm and adaptively adjust when thresholds are exceeded, effectively preventing pressure injuries. Simultaneously, a micro-airflow circulation and constant temperature control module stabilizes the temperature of the contact area within the human comfort range of 36–37°C. Combined with an ergonomic rigid support structure, this enhances patient comfort and psychological safety, reducing unconscious positional movements caused by anxiety. The device adopts a modular exoskeleton design for multiple parts such as the head and neck, chest and abdomen, and pelvis, which can be independently disassembled and flexibly combined to quickly adapt to different treatment sites. It has built-in standardized quality control programs and universal interfaces, and combined with the REAC-Net model, it realizes automatic identification of device abnormalities and performance degradation warnings, simplifies the whole-cycle quality control process, reduces the workload of clinical quality control, and achieves simultaneous improvement in clinical efficiency and patient treatment experience.

[0020] (6) The deep learning REAC-Net model endows the system with self-optimization and generalization capabilities. This invention's innovative deep learning model, REAC-Net, possesses online autonomous iterative learning capabilities. It can continuously optimize the system's positioning accuracy, motion prediction ability, and personalized adaptation schemes through initial positioning data and accumulated clinical case data (including body position data, motion characteristics, error data, and patient feedback). It can adapt to the treatment needs of patients with different body types and tumor types without manual intervention, significantly improving the system's clinical generalization ability. Compared to existing technologies that require manual parameter adjustment and have limited adaptability, this invention's autonomous optimization feature effectively lowers the clinical operation threshold, extends equipment lifespan, enhances long-term clinical application value, and has broad prospects for promotion.

[0021] In summary, this invention, through the deep integration of the REAC-Net model with multimodal sensing, multi-degree-of-freedom high-precision drive devices, and end-to-end radiation-compatible design, comprehensively overcomes the shortcomings of existing radiotherapy fixation technologies in core dimensions such as accuracy, patient positioning adaptability, dynamic stability, radiation compatibility, and patient comfort. The overall solution demonstrates outstanding technological innovation and strong clinical applicability, and can drive the iterative upgrade of radiotherapy positioning devices towards intelligence, precision, and humanization, possessing significant technological advantages and clinical application value. Attached Figure Description

[0022] Figure 1 Overall design drawing of an artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 2 The algorithm architecture flow for the deep learning REAC-Net model; Figure 3 Global view of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 4 Supine side view of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 5 Lateral-lying side view of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 6 A seated side view of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device. Figure 7 Side view of the head and neck module of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 8 Front view of the torso module of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 9 Front view of the limb module (arm) of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Figure 10 A frontal view of the limb module (leg) of an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device; Detailed Implementation To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0023] like Figure 1 As shown in this embodiment, the core objective of the AI-driven exoskeleton-assisted multi-position radiotherapy fixation device is to address the inherent shortcomings of existing radiotherapy fixation technologies, such as insufficient precision, poor positional adaptability, and weak dynamic stability. This invention provides an active, intelligent, and multi-degree-of-freedom exoskeleton radiotherapy fixation solution. To this end, this embodiment constructs an integrated intelligent exoskeleton radiotherapy fixation system. Using a biomimetic exoskeleton structure as a carrier, it integrates a high-precision sensing module, a deep learning-driven real-time calculation module, and an active compensation control module. This achieves adaptive adaptation to multiple treatment positions and sub-millimeter-level positional repetitive positioning accuracy. The following describes the technical solution of this invention in detail, clarifying the role, connection relationship, and specific function of each hardware device and functional module, in conjunction with the invention's objective.

[0024] I. System Overall Architecture and Workflow This system adopts a layered modular architecture, consisting of five functional layers. Each functional layer independently performs a specific function while efficiently coordinating and linking together. It balances mechanical support stability, sensor positioning accuracy, motion execution compliance, and algorithm adaptive performance, enabling personalized and precise control of radiotherapy positioning. The system strictly follows the core working logic of "initialization-sensing-decision-execution-closed-loop optimization," constructing a closed-loop control throughout the entire process to ensure continuous and stable positioning accuracy. The specific layered architecture and working steps are described in detail below: (1) Positioning and parameter initialization: This step is the core of the system's "initialization" process, which is the foundation for the system to achieve high-precision control. Its core function is to complete the initial coarse adjustment of the patient's body position and the individualized adaptation of control parameters. The specific implementation process and the collaborative logic of each component are as follows: After the patient is positioned in the exoskeleton device frame, the system first calls the preset target area position requirements in the radiotherapy treatment plan, converts the requirements into motion commands for the exoskeleton mechanical structure, drives the exoskeleton mechanical joints to complete the coarse adjustment of the body position, and ensures that the patient's body posture initially conforms to the treatment reference position, laying the foundation for subsequent high-precision fine adjustment.

[0025] Meanwhile, the deep learning REAC-Net model establishes a bidirectional data connection with the clinical case database, calling up historical clinical data stored in the database (including body position data, physiological movement characteristics, historical error data, patient adaptation feedback, etc.), and combining it with the patient's individual anatomical characteristics (such as tumor location, body shape parameters, and distribution of bone markers), and completing multi-dimensional correlation analysis through algorithms—focusing on analyzing the adaptability of the radiotherapy gantry rotation angle and the radiation field transmission path, strictly following the beam eye perspective (BEV) avoidance principle, automatically optimizing the overall posture layout of the exoskeleton, effectively avoiding the inherent defects of traditional fixation devices with fixed posture and easy to cause radiation field obstruction, and ensuring unobstructed and accurate projection of treatment rays.

[0026] Based on the results of the above-mentioned clinical data correlation analysis and shooting field adaptation, the system adaptively completes the initialization of core control parameters, including exoskeleton support points, positioning accuracy thresholds, and drive unit motion parameters, so as to achieve precise matching between positioning control parameters and individual patient characteristics and clinical treatment plans, thus laying a solid foundation for high-precision body position control throughout the entire process.

[0027] (2) Multimodal sensing data acquisition and fusion: This step is the core of the system's "sensing" process. Its core function is to collect multi-dimensional data such as patient position, exoskeleton posture, and relative position, providing comprehensive and accurate data support for subsequent error calculation and intelligent decision-making and control. The specific deployment, connection relationships, and functional principles of each sensing component are as follows: 1. Array ToF Sensors: A total of 6 groups are deployed in the ceiling of the treatment room and connected to the system's main control unit via a wired link. Their core function is to construct a global treatment coordinate system, acquire real-time 3D spatial position information of the patient's body surface and exoskeleton frame, and convert the acquired 3D point cloud data into a voxel network. This provides a unified spatial reference for the entire positioning process and enables real-time monitoring of the relative displacement between the patient's body surface and the exoskeleton frame, promptly capturing positional shift signals.

[0028] 2. Attitude Sensing Unit (Integrated 9-Axis IMU and Absolute Magnetic Encoder): Integrating an accelerometer, gyroscope, magnetometer, and absolute magnetic encoder, this unit is embedded in each joint of the exoskeleton and establishes a real-time data transmission connection with the main control unit and the REAC-Net model. Its core function is to collect real-time data on the three-dimensional rotation angle, angular velocity, and spatial attitude of each joint of the exoskeleton, synchronously feeding back the exoskeleton's motion state and providing crucial information for determining the degree of matching between the exoskeleton's attitude and the treatment reference position.

[0029] All the multi-dimensional data collected by the aforementioned sensing devices (spatial position data from the ToF sensor, attitude and angular velocity data from the 9-axis IMU, and joint position data from the absolute magnetic encoder) are uniformly processed by the main control unit to complete timestamp alignment and data noise reduction preprocessing, constructing a multi-dimensional heterogeneous sensing data matrix, and pushing it to the deep learning REAC-Net model in real time to achieve deep fusion of multi-modal information, providing a comprehensive, accurate, and time-synchronized data foundation for subsequent error calculation, motion planning, and attitude control.

[0030] (3) Error calculation and motion planning of deep learning REAC-Net network: This step is the core of the system's "decision-making" process. Its core function is to accurately calculate body position error and plan motion trajectory based on multimodal perception data, providing clear action instructions for subsequent execution steps. The connection relationships and specific functions of the REAC-Net model with each module are as follows: The deep learning model REAC-Net serves as the core decision-making module of the system, establishing bidirectional data connections with both the multimodal sensor fusion module and the system's main control unit. This model replaces the traditional thin-plate spline (TPS) deformation registration algorithm, overcoming the shortcomings of poor adaptability and insufficient calculation accuracy of traditional algorithms. Specifically, the REAC-Net model receives a multidimensional heterogeneous sensor data matrix transmitted from the multimodal sensor fusion module and simultaneously calls upon the body position reference data set during the initial positioning stage, performing high-precision registration and comparison of the two sets of data. Through the model's built-in multimodal feature extraction, data fusion, and error modeling algorithms, it accurately calculates the six-degree-of-freedom body position errors of three-dimensional translation (Δx, Δy, Δz) and three-dimensional rotation (Δθx, Δθy, Δθz), ensuring that the error calculation accuracy meets sub-millimeter requirements.

[0031] Meanwhile, the REAC-Net model can adaptively collect the temporal characteristics of the patient's respiratory movements, muscle tremors, and other physiological movements, analyze and predict the patterns of physiological movement changes in real time, and transmit the prediction results synchronously to the system's main control unit and the radiotherapy equipment control system. This provides a reliable instruction basis for the subsequent linkage control of the radiotherapy equipment to synchronously emit beams, ensuring that the radiotherapy rays are always accurately matched with the patient's physiological movement timeline and avoiding target area deviation caused by dynamic errors.

[0032] In addition, the REAC-Net model, based on error calculation results, patient physiological movement characteristics, and the dynamic characteristics of the exoskeleton mechanical structure, completes motion trajectory planning. It decomposes the six-degree-of-freedom error command into the action sequence of each joint drive unit, accurately defines the motion direction, motion amplitude, and motion speed of each drive unit, and ensures that the subsequent execution process is smooth and precise, and eliminates patient discomfort and secondary positional deviation caused by motion impact.

[0033] (4) Proactive compensation implementation: This step is the core "execution" stage of the system. Its core function is to correct postural errors and regulate pressure based on the motion commands output by the REAC-Net model, balancing postural accuracy and patient comfort. The functions, connections, and roles of the various execution hardware and auxiliary sensing devices are as follows: Drive unit: Includes high-precision drive devices (linear piezoelectric actuator and rotary torque motor), establishes a control connection with the system main control unit, receives timing action commands issued by the main control unit, and is the power source for the movement of the exoskeleton's mechanical joints; among them, the linear piezoelectric actuator realizes the adjustment of the exoskeleton's three-dimensional translational degrees of freedom, and the rotary torque motor realizes the adjustment of the exoskeleton's three-dimensional rotational degrees of freedom. The two work together to follow the motion trajectory planned by the REAC-Net model, drive the exoskeleton's joints to complete compliant and precise micro-adjustments, correct the body position offset to within the sub-millimeter threshold allowed by radiotherapy, and achieve precise compensation for body position errors.

[0034] Thin-film pressure sensor: Integrated into the contact layer between the exoskeleton and the patient's body surface, it maintains real-time data interaction with the REAC-Net model and the system's main control unit. Its core function is to collect real-time contact pressure data and simultaneously upload it to the REAC-Net model's PID parameter optimization module. Based on the measured pressure data, this module stabilizes the contact pressure within the human comfort range of 8-12 kPa, ensuring postural stability while preventing pressure injury. When the pressure monitoring value exceeds the safety threshold, the system automatically triggers an early warning and generates a posture adjustment command. The main control unit then drives the support structure to adaptively fine-tune, suppressing excessive soft tissue deformation. This design significantly improves patient comfort while maintaining high-precision postural fixation, effectively resolving the inherent technical contradiction of traditional fixation devices that struggle to balance postural stability and patient comfort.

[0035] (5) Closed-loop verification and full-process monitoring: This step is the core of the system's "closed-loop optimization." Its core function is to verify the accuracy of the patient's position, monitor the patient's position throughout the radiotherapy process, and achieve iterative optimization of the system's performance, forming a closed loop throughout the entire process. The collaborative relationships and functions of each module are as follows: After the body position is fine-tuned, the system initiates the accuracy verification process: the six array ToF sensors collect the spatial position data of the patient's body surface and the exoskeleton frame again, compare it with the initial positioning reference data, verify the consistency between the target area position and the reference posture, and ensure that the position accuracy meets the requirements of radiotherapy; if the accuracy does not meet the preset threshold, the system will re-trigger the error calculation and active compensation process until the position accuracy meets the threshold requirements.

[0036] Throughout the radiotherapy process, the system is continuously monitored: a ToF sensor, a 9-axis IMU, and an absolute magnetic encoder synchronously and continuously acquire data on the patient's respiratory movements, unconscious body movements, and exoskeleton posture; a thin-film pressure sensor continuously acquires data on the body surface contact pressure; all monitoring data are fed back to the deep learning REAC-Net model in real time; based on this validation data and error feedback information, the model updates its own parameter weights online, iteratively optimizes the accuracy of subsequent error calculation, motion prediction ability, and pressure control strategy, and achieves autonomous learning and continuous improvement of system performance.

[0037] Through the above closed-loop process, the system forms a complete closed-loop control cycle, continuously improving the stability, accuracy and individualized adaptability of body position control, ensuring that the patient's body position is always maintained within the sub-millimeter accuracy range throughout the radiotherapy process, meeting the stringent requirements of modern precision radiotherapy.

[0038] II. Detailed Description of Core Subsystems (1) Mechanical structure and materials subsystem This subsystem serves as the carrier of the entire exoskeleton radiotherapy fixation device. Its core function is to achieve multi-positional adaptation, high-precision support, and radiation compatibility. It employs a lightweight biomimetic topology and modular design, and is divided into a head and neck module, a trunk module (including chest, abdomen, and pelvic areas), and limb modules. Each functional module is detachably connected to the main frame via standardized quick-release interfaces, allowing for selective assembly according to the patient's treatment site. This design balances multi-positional adaptability, ease of operation, and structural stability. Specific parameters, installation positions, and functions of each component are shown in Table 1 below. Figures 3 to 10 As shown: Table 1

[0039] Joint Design: The exoskeleton joints utilize high-rigidity, lightweight carbon fiber connecting rods and zirconia ceramic bearings to form a high-precision rotation system. This ensures both joint flexibility and structural rigidity, while the use of low-electron-density materials mitigates the scattering and attenuation interference of therapeutic radiation from metal components. At key degrees of freedom, high-precision linear piezoelectric actuators and rotary torque motors are integrated to achieve precise control of sub-millimeter linear motion and arcsecond-level rotational motion. This provides a reliable execution basis for millimeter-level fine-tuning of body positions driven by the deep learning-based REAC-Net model, ensuring the accuracy and compliance of position adjustments.

[0040] Contact Layer Design: The contact interface between the exoskeleton and the human body adopts a medical-grade silicone composite 3D elastic mesh structure, beneath which is a composite medical-grade memory-type low-temperature thermoplastic film. Thin-film pressure sensors are uniformly integrated within the contact layer to collect real-time data on the patient's surface contact pressure distribution. This low-temperature thermoplastic film can be softened by low-temperature heating to precisely conform to the patient's body contours, and quickly solidify upon cooling, achieving individualized adaptation. Simultaneously, local pressure distribution features are extracted using a deep learning REAC-Net model. When pressure exceeds a preset safety threshold, an automatic alarm mechanism is triggered, and the support structure is driven to adaptively adjust, preventing excessive soft tissue deformation. This ensures both the fit and stability of the body position fixation, significantly improving patient comfort during prolonged radiotherapy and resolving the inherent contradiction between fixation stability and patient comfort in existing technologies.

[0041] Material list and technical requirements: 1. The main structure is made of carbon fiber reinforced PEEK (PEEK-CF30) composite material. This material has a dose perturbation of <3% for 6MV radiotherapy rays, and the linear attenuation coefficient is strictly controlled within 0.075-0.085cm. - 1. It can simultaneously ensure structural rigidity and excellent radiation transmittance; 2. The surface markers use gold nano-coated silica microspheres (1mm in diameter). 3D point cloud data is collected by a ToF sensor and converted into a voxel network for optical positioning system calibration, providing a high-precision spatial reference for spatial positioning and error calculation of the deep learning REAC-Net model algorithm. 3. All materials in contact with the therapeutic radiation strictly meet the requirements for 6MV photon irradiation, with a linear attenuation coefficient <0.1cm. - ¹; Strict control is implemented over the metal components in the system. The size of a single metal component is <3cm, and the minimum distance between metal components is ≥15cm. This completely avoids interference with X-ray transmission and image positioning caused by local metal accumulation, ensuring the accuracy of treatment dosage and image guidance.

[0042] Radiotherapy Adaptive Safety System: 1. Beam Path Optimization: A dual optimization strategy of "mechanical design + algorithm control" is adopted to ensure unobstructed radiation field. All mechanical structures of the exoskeleton are designed according to the beam-eye-perspective (BEV) principle. The deep learning REAC-Net model is used to analyze the matching relationship between the gantry rotation angle and the radiation field path, automatically adjusting the exoskeleton posture to avoid the treatment radiation field area and eliminating radiation field obstruction by the mechanical structures. 2. Dosimetric Validation: A dual validation method combining Monte Carlo numerical simulation and actual phantom measurements is used to rigorously verify and control the dose perturbation caused by the device, reducing the 6MV photon beam dose perturbation caused by the device to <1%, ensuring that the dose accuracy meets clinical requirements.

[0043] (2) High-precision positioning and sensing subsystem This subsystem is the core of achieving sub-millimeter-level body positioning and dynamic compensation. Its core function is to build a global high-precision sensor network, complete multimodal data acquisition, synchronous fusion and precise coordinate transformation, provide high-quality data input for the deep learning REAC-Net model, and achieve seamless and precise mapping of coordinate systems at all levels by optimizing coordinate transformation logic and error compensation strategy, thus ensuring the accuracy and dynamic stability of body positioning.

[0044] Global coordinate system construction in the treatment room: Six ToF sensor arrays are deployed on the ceiling of the treatment room. The sensors collect 3D point cloud data of the treatment room space in real time and convert it into a voxel network to generate a high-precision three-dimensional spatial reference for the treatment room. At the same time, the global position data of the exoskeleton base is collected to establish the initial association between the exoskeleton and the global coordinate system of the treatment room. This provides a unified spatial reference for the calibration and transformation of the coordinate system at each level, ensuring the consistency of the positioning reference throughout the process.

[0045] Precise posture perception of the exoskeleton: Each joint of the exoskeleton is equipped with a posture sensing unit (integrating a 9-axis IMU and an absolute magnetic encoder), with a joint rotation angle measurement resolution of 0.01°. It can output arcsecond-level position feedback signals, providing high-precision position data support for joint motion closed-loop control and millimeter-level posture fine-tuning. At the same time, it captures the three-dimensional acceleration and angular velocity data of the exoskeleton device in real time, dynamically feedbacks changes in exoskeleton posture, and promptly captures posture offset signals. This provides comprehensive dynamic data support for multimodal data fusion and posture error compensation, ensuring the real-time performance and accuracy of posture perception.

[0046] High-precision drive configuration: The linear drive adopts a high-precision linear piezoelectric actuator with a motion resolution of 0.1mm and a response time of <5ms, which can realize sub-millimeter level precision translational motion and meet the high-precision requirements of body position translation fine adjustment; the rotary drive is equipped with a high-torque rotary torque motor, which can output large torque and backlash-free rotary motion, ensuring the dynamic stability and positioning accuracy of body position rotation adjustment. It works in conjunction with the linear piezoelectric actuator to achieve high-precision control of the exoskeleton in six degrees of freedom.

[0047] Intelligent pressure control: The system's core control module establishes a real-time data connection with the thin-film pressure sensor integrated into the contact layer. The thin-film pressure sensor collects real-time data on the surface contact pressure distribution and transmits it synchronously to the PID parameter optimization module of the deep learning REAC-Net model. This module performs real-time analysis and adjustment based on the measured pressure data, stabilizing the surface contact pressure within the ideal comfortable range of 8–12 kPa. When the pressure exceeds this threshold, the system automatically triggers an alarm signal and drives the support structure to adaptively fine-tune, avoiding excessive deformation of soft tissue and achieving a precise balance between the accuracy of body positioning and the patient's treatment comfort.

[0048] Coordinate Transformation Mathematical Model: To achieve accurate mapping of the three-level coordinate system (treatment room - exoskeleton equipment - patient), a dedicated coordinate transformation link is constructed. The transformation matrix parameters are iteratively optimized using a deep learning REAC-Net model to effectively compensate for the effects of equipment installation errors and environmental thermal drift, reducing cumulative system errors and ensuring the accuracy of coordinate system mapping at each level. The specific transformation model is as follows:

[0049] in, This is the transformation matrix from the treatment device coordinate system to the exoskeleton base. Here is the attitude rotation matrix. This is a thermal drift compensation term. Let these be the patient's coordinates in the coordinate system of the treatment device. This model represents the coordinates of the exoskeleton joints within its own coordinate system. It enables precise calculation of the patient's coordinates, providing accurate coordinate data for postural adjustment.

[0050] (3) Deep learning REAC-Net model algorithm subsystem 1) REAC-Net structure 1. Overall Architecture REAC-Net employs a five-layer hierarchical architecture, forming a complete "initialization-perception-decision-execution-closed-loop optimization" closed loop. This ensures efficient processing of multimodal data, accurate decision-making, and dynamic iterative optimization. The deep learning REAC-Net model algorithm architecture process is as follows: Figure 2 As shown: The multimodal input layer, feature extraction layer, dynamic fusion layer, decision output layer, adaptive update layer, and closed-loop link interact sequentially, and the closed-loop link returns to the feature extraction layer. The multimodal input layer receives raw data collected by various sensors; the feature extraction layer performs preprocessing and feature extraction of each modality; the dynamic fusion layer realizes intelligent fusion of multimodal features and noise filtering; the decision output layer generates positional error calculation results, motion control commands and parameter optimization schemes; the adaptive update layer realizes online iteration and fault-tolerant adjustment of model parameters; and the closed-loop link continuously optimizes model performance to ensure adaptation to the complex physiological movements and environmental changes required during radiotherapy.

[0051] 2. Detailed Structure and Mathematical Principles 2.1 Feature Extraction Layer Core Functionality: This function performs targeted preprocessing and feature extraction on raw data from ToF sensors, 9-axis IMUs, absolute magnetic encoders, and thin-film pressure sensors. It transforms heterogeneous multimodal data into unified, fusionable feature vectors, providing high-quality input for subsequent dynamic fusion and decision output. This effectively addresses the problems of poor adaptability and low feature recognition in traditional feature extraction methods. The processing methods and mathematical expressions for each modality are as follows: ToF sensor data: The acquired 3D point cloud data is converted into a voxel network to extract spatial location features and accurately capture the relative positional relationship between the patient's body surface and the exoskeleton. The expression is:

[0052] in, This represents the number of point clouds; The point cloud is represented by its 3D spatial coordinates; the point cloud is converted into a voxel mesh through voxelization. Its three-dimensional parameters are defined as follows: height H corresponds to the number of voxels in the cranio-coccygeal direction of the human body, width W corresponds to the number of voxels in the left-right direction of the human body, and depth D corresponds to the number of voxels in the ventral-dorsal direction of the human body.

[0053] 9-axis IMU data: Temporal features are extracted using a Long Short-Term Memory (LSTM) network, adapting to the temporal characteristics of IMU data. This effectively solves the long-term dependency problem of traditional Recurrent Neural Networks (RNNs), accurately capturing the dynamic changes in exoskeleton posture. The expression is:

[0054] in, The hidden state of the LSTM at time t; : Input feature vector at time t; LSTM (Long Short-Term Memory Network): A special variant of recurrent neural network (RNN) specifically designed for processing continuous temporal data. Through a three-level gating mechanism (forget gate, input gate, and output gate), it effectively preserves valid temporal features, filters noise, and accurately captures the dynamic changes in exoskeleton posture; The 9-dimensional column vector (T denotes transpose) corresponds to the output data of the 9-axis inertial measurement unit (IMU), where... For acceleration, Angular velocity, denoted as , where is the magnetic field strength.

[0055] Magnetic encoder data: Converted into learnable embedding vectors through position encoding, this accurately captures subtle changes in joint position, providing precise feature support for pose tracking and positioning error prediction. The expression is:

[0056] in, The output position embedding vector serves as the input feature for subsequent models such as pose tracking and positioning error prediction, and can accurately capture subtle changes in joint position. Embedding functions are feature transformation modules in deep learning, used to convert raw location data into high-dimensional feature vectors. : Represents the output vector of a 17-bit absolute magnetic encoder, which collects raw data of the angle and position of the exoskeleton joints to ensure high-precision extraction of joint position features.

[0057] Pressure sensor data: Local pressure distribution features are extracted using a convolutional neural network (CNN) to accurately reflect the contact state between the patient's body surface and the exoskeleton, providing support for pressure regulation and postural stability assessment. The expression is as follows:

[0058] in, The output pressure feature vector is used for postural stability assessment, positioning deviation warning, etc. : Represents a two-dimensional pressure distribution image, corresponding to the collected pressure distribution data on the patient's body surface, used to reflect the contact state between the patient's body and the fixation device; A real matrix with M rows and N columns, corresponding to the pixel dimension of the pressure distribution image.

[0059] 2.2 Dynamic Fusion Layer Core Functionality: This innovative approach combines Transformer and Graph Neural Networks (GNNs) to overcome the limitations of traditional static fusion methods, achieving optimal fusion of multimodal data. Through dynamic confidence-aware weight allocation, causal-aware multi-head attention mechanisms, and dynamic graph structure learning, it effectively filters sensor noise and non-causal interference, improving the accuracy and robustness of fused features. It adapts to the complex physiological movements and environmental changes during radiotherapy. Its innovation lies in achieving "dynamic, causal, and structured" multimodal fusion, distinguishing it from traditional fixed-weight fusion methods. The specific implementation process is as follows: 2.2.1. Weighting of Dynamic Confidence Perception Traditional multimodal fusion methods often use fixed-weighted averaging, which cannot adapt to the dynamic changes in sensor data in radiotherapy scenarios (such as sensor noise and equipment wear). This invention, the REAC-Net model, fuses weights... It is over time The dynamically changing matrix is ​​set by the sensor to a real-time confidence vector. The decision ensures the dominant role of high-confidence sensor data, significantly improving fusion accuracy.

[0060] 2.2.1.1 Sensor Confidence Calculation For the A sensor, its confidence level at time t Taking into account two core factors: firstly, its measured value Compared with historical stable values The degree of deviation (modeled using Gaussian distribution) and the health status of the sensor itself are two factors. (Values ​​range from 0 to 1, where 1 represents perfect health), the expression is:

[0061] in: For sensors The historical average measurement value reflects the stable operating state of the sensor; For sensors The noise standard deviation is calibrated by the sensor's factory parameters and clinical measurement data. This refers to the hardware health status reported by the device driver layer, providing real-time feedback on the sensor's operating status. This expression can accurately quantify the reliability of sensor data, providing a scientific basis for dynamic weight allocation.

[0062] 2.2.1.2 Dynamic Fusion Weight Generation The confidence scores are converted into normalized fusion weights using the Softmax function. At the same time, temperature parameters are introduced. To control the smoothness of the weight distribution (smooth in the early stages of training, sharp in the later stages), the expression is:

[0063] Where N is the total number of sensors, the normalization process ensures that the sum of the weights of all sensors is 1, thus achieving a reasonable allocation of weights.

[0064] 2.2.1.3 Dynamically Weighted Fusion Output set up For the first The feature vectors extracted from each modality are then fused into a single feature vector. for:

[0065] in It is a modality-specific projection matrix used to map features of different dimensions to a unified latent space, solving the problem of inconsistent dimensions of multimodal features and ensuring the uniformity and usability of fused features.

[0066] 2.2.2. Multi-head attention mechanism of causal perception In radiotherapy settings, sensor data contains both causal factors (such as actual patient positional displacement and respiratory movements) and non-causal factors (such as sensor electromagnetic interference and environmental noise). Traditional attention mechanisms cannot distinguish between these two, easily leading to distortion of fused features. This invention introduces a causal mask into the attention mechanism. This enables an effective distinction between causal and non-causal factors, enhancing the reliability of fusion features.

[0067] 2.2.2.1 Generation of queries, keys, and values For the input sequence ( For sequence length, (as feature dimension), generate query matrix Key matrix and value matrix This enables feature mapping and association capture, expressed as:

[0068] Among them, W Q W K W V The learnable projection weight matrix consists of parameters optimized during model training, used to map input features to three different feature spaces: query (Q), key (K), and value (V). Input X is a temporal sequence of multi-sensor fusion features (such as IMU data, magnetic encoder position embeddings, and a spliced ​​sequence of pressure pad features). By generating Q, K, and V, the attention mechanism can automatically capture the dynamic correlations between features from different sensors and at different times, thereby improving the accuracy of attitude tracking and the robustness of positioning error prediction.

[0069] 2.2.2.2 Causal Similarity Calculation Add causal masking to standard dot product attention. This mask is generated in real time by the dynamic causality discovery module (based on gradient-based causality strength analysis) and is used to mask out non-causally correlated time steps or sensor nodes. The expression is:

[0070] Physical significance: In time-series modeling tasks such as posture tracking and positioning error prediction of radiotherapy exoskeletons, this mechanism can effectively filter out non-causal interference—if the IMU data is subject to electromagnetic interference (non-causal noise). This reduces its influence in the fusion process, forcing the model to focus more on high-confidence data such as ToF sensors; at the same time, it shields "posture information that has not yet occurred in the future," making the model fit the temporal logic of real clinical scenarios, significantly improving the reliability and interpretability of positioning error prediction, and solving the defect of traditional attention mechanisms being susceptible to noise interference.

[0071] 2.2.3. Learning Dynamic Graph Structures In traditional graph neural networks, the connection weights between sensor nodes are fixed, which cannot adapt to the dynamic changes in the correlation of sensor data during radiotherapy. In this invention, the connection weights between nodes (sensors) in the graph neural network are not fixed, but rather an adjacency matrix is ​​dynamically constructed based on the correlation of real-time data. This enables dynamic correlation and capture of sensor features.

[0072] 2.2.3.1 Construction of Dynamic Adjacency Matrix The real-time correlation between sensors is measured using the inner product of node features. The correlation result is compressed to the (0,1) interval using the Sigmoid function to form dynamic edge weights, expressed as:

[0073] in, , : represents the real-time feature vectors of sensor nodes i and j at time t, corresponding to the sensor monitoring data at that time; σ: is the Sigmoid activation function, which compresses the result of the inner product of features to the (0,1) interval to ensure the rationality of the edge weights. The larger the value, the stronger the real-time correlation between the two sensor nodes. 2.2.3.2 Dynamic Graph Convolution Combining dynamic adjacency matrix Sum-degree matrix A variant of GCN is used for graph convolution operations, emphasizing neighboring nodes. For the central node The influence of sensor features is considered to achieve structured fusion, expressed as:

[0074] in: :node The set of neighbors of node i, that is, the sensor nodes that are strongly correlated with node i; :No. The learnable weight matrix of the layer; Activation function: used to introduce non-linear features and enhance the model's expressive power; degree matrix. Used for normalization processing to avoid feature distortion caused by differences in node degree and ensure the rationality of graph convolution results.

[0075] 2.2.3.3 Dynamic Weight Allocation The fusion weights of graph nodes are automatically adjusted based on sensor confidence levels to further enhance the role of high-confidence sensors. The expression is as follows:

[0076] in, Let be the real-time confidence level of sensor i. This is the weight adjustment coefficient, used to control the degree of influence of confidence on node weights; the confidence level is calculated by combining historical error data and the real-time working status of the sensors to ensure the accuracy of weight allocation.

[0077] 2.2.4. Comprehensive Dynamic Fusion Function Integrating the aforementioned dynamic confidence-aware weight allocation, causal awareness multi-head attention mechanism, and dynamic graph structure learning process, the final output of the dynamic fusion layer is... This can be represented as a composite function that not only achieves efficient fusion of multimodal data, but also improves the accuracy and robustness of the fused features by filtering noise and non-causal interference through causal filtering and graph structure optimization. The expression is:

[0078] in, , , These are the feature vectors extracted from the ToF sensor, the 9-axis IMU, and the absolute magnetic encoder, respectively. All the above formulas together constitute the mathematical core of the "dynamic fusion layer" of the REAC-Net model, which distinguishes it from the traditional static fusion method. It can adapt to the complex physiological movements and environmental changes during radiotherapy and provide high-quality fusion features for subsequent decision output.

[0079] 2.3 Decision Output Layer Core Functionality: Based on the feature vectors output by the dynamic fusion layer, it accurately calculates six-degree-of-freedom positional errors, generates motion planning commands, and optimizes PID parameters online, providing precise control commands to the exoskeleton drive unit and achieving sub-millimeter-level positional adjustment and pressure stability control. The core innovation lies in the deep integration of causal analysis, motion planning, and PID optimization, improving the accuracy and adaptability of decision-making. The specific implementation of each module is as follows: Error Calculation Module: Through decoupled causal factor analysis, it distinguishes between causal components (such as actual postural deviation) and non-causal components (such as noise interference) in postural errors, improving the accuracy of error calculation. The expression is:

[0080] in, This is a causal relationship function used to extract the causal component (true body position deviation) from the error. It is a non-causal function used to extract non-causal components (such as sensor noise and environmental interference) from the error; causal and non-causal factors are determined through dynamic causal structure learning to ensure that the error calculation results can truly reflect the actual deviation of the patient's body position and provide an accurate basis for subsequent body position correction.

[0081] Motion planning module: Employing an improved bidirectional RRT algorithm, which balances the accuracy of postural adjustment, patient comfort, and fixation stability, it generates the optimal motion trajectory, expressed as:

[0082] in, The movement posture of exoskeleton joints; The distance error between the current posture and the target posture is used to ensure the accuracy of body position adjustment; The patient comfort index is fed back from pressure sensor data; As an indicator of body position stability, it is comprehensively evaluated based on multimodal sensor data; The adaptive weights are dynamically adjusted based on real-time patient feedback (such as stress data and comfort scores) to achieve a balance between accuracy, comfort, and stability.

[0083] PID parameter optimization module: This module uses deep reinforcement learning to adjust PID control parameters online, ensuring the exoskeleton contact pressure remains stable within the ideal range of 8-12 kPa, balancing fixation stability and patient comfort. The expression is:

[0084] in, , , The initial values ​​for the PID parameters are calibrated based on clinical experience and model training. , , The parameter adjustment values ​​output by the deep learning model are dynamically generated by the REAC-Net model based on real-time pressure data and error feedback, ensuring the accuracy and stability of pressure control and avoiding excessive local pressure that could cause patient discomfort or soft tissue damage.

[0085] 2.4 Adaptive Update Layer Core functionality: Enables online autonomous learning and iterative parameter updates for the REAC-Net model, while constructing a three-level fault-tolerance mechanism to ensure stable operation even under conditions such as sensor failure and data anomalies, thereby improving the model's clinical generalization ability and reliability. The core innovation lies in achieving the model's "self-optimization, self-adaptation, and self-fault tolerance," allowing it to adapt to the needs of different patients and treatment scenarios without manual intervention. Specific implementation details are as follows: Dynamic Causal Structure Learning: A Dynamic Causal Variational Autoencoder (VAE) is employed to accurately separate and model causal and non-causal factors, providing support for error calculation and feature fusion. The loss function expression is as follows:

[0086] in, Divided into causal factors Dynamic non-causal factors and static non-causal factors ; For approximate posterior distribution, To generate a distribution, It is the prior distribution; To reconstruct the loss, ensure that the model can accurately reconstruct the input data; KL divergence is used to constrain the distribution of latent variables. The constraint coefficient is used to optimize the loss function, thereby achieving accurate separation between causal and non-causal factors and improving the model's robustness against interference.

[0087] Online learning mechanism: A sliding window-based data update strategy is employed to achieve online iterative optimization of model parameters, adapting to individual differences among patients and dynamic changes during treatment. The expression is:

[0088] in, Let be the model parameters at time t+1. The model parameters are at time t; The adaptive learning rate is dynamically adjusted based on the model training error to ensure the stability and convergence of parameter updates. Let be the gradient of the loss function at time t. This includes multimodal sensing data and error feedback data within the current sliding window. Through this mechanism, the model can autonomously learn the postural characteristics and physiological movement patterns of different patients, continuously improving positioning accuracy and adaptability.

[0089] Fault Tolerance Mechanism: A three-tiered circuit breaker system is constructed. When a sensor malfunctions or data anomalies, the system automatically switches to a backup solution to ensure stable system operation and avoid treatment interruptions or positioning errors caused by equipment failure. The specific tiers are as follows: 1. Primary Circuit Breaker: When a single sensor's data is abnormal (e.g., data exceeding the normal range or excessive noise), the system automatically reduces the confidence weight of that sensor and replaces it with data from other high-confidence sensors to ensure the accuracy of the fused features; 2. Intermediate Circuit Breaker: When multiple sensor data conflicts occur (e.g., the deviation of positioning data collected by different sensors exceeds a threshold), the system initiates a data calibration process, combining historical data and causal analysis to filter out reliable data and ensure the accuracy of decision output; 3. Advanced Circuit Breaker: When system-level fault detection occurs (e.g., multiple core sensors fail simultaneously or the drive unit malfunctions), the system automatically triggers an alarm signal and switches to manual control mode, while retaining the current positioning data to support clinical emergency response and ensure treatment safety.

[0090] In summary, the REAC-Net model algorithm subsystem, through a five-layer hierarchical architecture design and combined with innovative technologies such as dynamic multimodal fusion, causal perception, online learning, and fault-tolerant control, breaks through the limitations of traditional algorithms, achieving efficient processing, accurate decision-making, and dynamic optimization of multimodal data. It provides core algorithmic support for sub-millimeter-level body position adjustment and dynamic error compensation of exoskeleton radiotherapy fixation devices, highlighting the innovation and clinical applicability of this invention, and effectively solving the core defects of existing technologies such as poor algorithm adaptability, insufficient prediction accuracy, and weak anti-interference ability.

[0091] 2) Innovative Advantages of the REAC-Net Model The REAC-Net model addresses the core shortcomings of existing radiotherapy positioning control algorithms, such as weak anti-interference ability, poor adaptability, insufficient interpretability, and lack of safety. It combines these shortcomings with the needs of clinical radiotherapy scenarios to achieve multi-dimensional technological innovation, effectively making up for the deficiencies of traditional algorithms. Specific innovative advantages are as follows: 1. Possesses accurate dynamic causal reasoning capabilities, improving the reliability of error judgment. This invention, through dynamic causal structure learning and causal masking mechanisms, can accurately identify the causal sources of positional errors (such as the patient's actual positional displacement and respiratory movements) and non-causal interferences (such as sensor noise and electromagnetic interference). It effectively avoids misjudging non-causal correlations as causal relationships, solving the problem of error calculation distortion caused by the confusion between causal and non-causal factors in traditional algorithms. Simultaneously, this capability ensures that the model maintains stable performance output even when the data distribution in radiotherapy scenarios changes dynamically (such as patient position switching and sensor state fluctuations), guaranteeing the consistency of positional control accuracy and fully adapting to the dynamic needs of complex radiotherapy scenarios.

[0092] 2. Achieve multimodal adaptive fusion to adapt to the dynamic needs of different radiotherapy scenarios. Breaking through the technical limitations of traditional multimodal fusion's "fixed weights," the REAC-Net model can dynamically adjust the fusion weights of each sensor modality according to real-time changes in the radiotherapy scenario, achieving optimal utilization and efficient fusion of multimodal data. Specifically, when the patient is stationary, the model prioritizes relying on ToF sensors (spatial position data) and absolute magnetic encoders (joint position data) to ensure high accuracy in body positioning. When the patient exhibits physiological micro-movements such as respiratory movements, the model automatically switches to relying on a 9-axis IMU (temporal attitude data) and a thin-film pressure sensor (contact state data) to achieve precise capture and dynamic compensation of physiological movements. When a sensor malfunctions or its data is abnormal, the model automatically reduces the fusion weight of that modality and uses other high-confidence sensor data as a substitute, ensuring the accuracy of the fused features and the continuity of the system, thus overcoming the shortcomings of traditional fusion algorithms, such as poor adaptability and weak anti-interference capabilities.

[0093] 3. Construct a closed-loop adaptive learning mechanism to improve individualized adaptation capabilities and clinical efficiency. The REAC-Net model integrates a closed-loop adaptive learning mechanism encompassing data acquisition, error analysis, parameter optimization, and model updates. It can automatically analyze historical treatment error data and reverse-optimize initial positioning parameters, reducing the accumulation of initial positioning errors. Simultaneously, the model can adaptively adjust postural control strategies based on each patient's individual anatomical characteristics and physiological movement patterns, achieving individualized adaptation. Furthermore, the model can learn postural adaptation patterns from new patients online, continuously iterating and optimizing error calculation and motion control accuracy as the number of treatments increases. This significantly shortens the postural adjustment time for subsequent treatments, improves clinical treatment efficiency, and addresses the problem of traditional algorithms being "highly generalizable but weakly individualized," lowering the operational threshold for clinical medical staff.

[0094] 4. Enhance interpretability and safety to meet the requirements of clinical application of medical devices. To address the "black box" problem of existing deep learning models, the REAC-Net model employs a fully connected signal flow (SiFu) mechanism to ensure that each decision-making step is traceable to specific raw sensor data, achieving interpretability of the decision-making process. Simultaneously, the model automatically generates detailed error analysis reports, clearly presenting the source of errors, the calculation process, and the basis for adjustments, facilitating review and traceability by clinical medical staff and meeting the traceability requirements of medical equipment. Furthermore, this mechanism is deeply integrated with the REAC-Net model, automatically intercepting unreasonable postural adjustment commands, mitigating safety risks such as excessive postural adjustment and pressure-induced injuries at the algorithmic level, and comprehensively ensuring patient treatment safety.

[0095] 3) Implementation details and parameter settings To ensure the engineering implementation and clinical applicability of the REAC-Net model, a scientific training deployment strategy and safety assurance mechanism were formulated, taking into account the hardware conditions of radiotherapy equipment and clinical treatment needs. The specific implementation details and parameter settings are as follows: 1. Training and Deployment Strategies A three-tiered implementation strategy of "offline pre-training + personalized fine-tuning + edge deployment" is adopted to comprehensively consider model accuracy, individualized adaptability, and real-time response performance. The specific parameters and implementation process are as follows: Pre-training: Offline training is conducted based on a large amount of historical radiotherapy data accumulated in clinical practice (including multi-position sensor data, position error data, patient anatomical feature data, etc.) to optimize the initial parameter configuration of the model and ensure that the model has basic position error calculation, multimodal data fusion and dynamic compensation capabilities, laying the foundation for subsequent personalized fine-tuning; Personalized fine-tuning: For each patient, initial positioning data, body surface pressure distribution data, and physiological movement characteristic time series data are collected. The pre-trained model is then individually fine-tuned, and the model parameters are adaptively adjusted to suit the individual characteristics and physiological movement patterns of the patient, ensuring that the model achieves sub-millimeter-level accuracy in controlling the patient's posture. Edge deployment: The trained model is quantized into INT8 format and deployed on the edge computing unit of the radiotherapy equipment. It does not rely on cloud computing resources, effectively reducing data transmission latency and ensuring the model's real-time response capability. Real-time performance: Clinically tested and verified, the model's single inference time is <50ms, which can meet the real-time control requirements for body position monitoring, error calculation and dynamic compensation during radiotherapy, ensuring precise synchronization between body position adjustment and radiotherapy beam output.

[0096] 2. Security Guarantee Mechanism To eliminate security risks caused by abnormal input data or parameter deviations, the REAC-Net model integrates a dual security protection mechanism to achieve full-process review and control of the decision-making process, as detailed below: Pre-approval: After receiving multimodal sensing data, the model first conducts a comprehensive review of the rationality of the input data, accurately filtering out abnormal data that exceeds the normal range, has obvious noise, or has data conflicts; then, it performs noise reduction and calibration on the abnormal data that can be corrected; if the abnormal data cannot be corrected, it automatically triggers an alarm and activates backup data to prevent abnormal data from causing the model to make incorrect decisions from the source. Post-processing review: After the model generates the body position adjustment instructions, the safety of the output instructions is immediately verified to check whether the instructions exceed the radiotherapy safety thresholds (such as the body position adjustment range, body surface contact pressure threshold, etc.). If the instructions have safety risks, the instructions are automatically intercepted and correction suggestions are generated to ensure the safety and rationality of the output instructions and protect the patient's treatment safety.

[0097] In summary, the advantages of the embodiments of the present invention are as follows: 1. Whole-body biomimetic exoskeleton topology for multi-position radiotherapy: Core Features: A pioneering modular full-body exoskeleton mechanical topology suitable for radiotherapy scenarios. The biomimetic design of shoulder, elbow, hip, and knee joints mimics the human kinetic chain, providing reliable support for body weight while actively guiding and rigidly fixing patients in various treatment positions, including supine, prone, lateral, sitting, and even semi-standing. Leveraging its modular design, the system can quickly reconstruct the fixation range for different treatment target areas such as the head and neck, chest, abdomen, and pelvis, flexibly adapting to the needs of radiotherapy in multiple locations.

[0098] Key protection features: The modular full-body exoskeleton's topological structure design, bionic joint layout, detachable connection of each module, and multi-position adaptation method effectively solve the inherent technical defects of traditional radiotherapy fixation devices, such as single position and poor adaptability.

[0099] 2. A multimodal fusion localization method combining optical global measurement, embedded local sensing, and deep learning. Core: Abandoning the existing approach of relying solely on optical or mechanical sensors for positioning, this innovative approach deeply integrates the global coordinate system constructed by the ceiling ToF sensor, the embedded IMU attitude perception data, and the joint absolute encoder position data. By establishing a unified coordinate transformation model and error compensation model, it achieves seamless millimeter-level precision mapping from the global space of the treatment room to the patient's body surface, providing a high-precision and highly reliable input benchmark for active position compensation.

[0100] Key protection areas include: the combined deployment of multimodal sensors, the construction of coordinate transformation and error compensation models, and the fusion logic of multi-source sensor data, to solve the technical problems of insufficient positioning accuracy, weak anti-interference ability, and poor data correlation in traditional positioning.

[0101] 3. Force / position hybrid intelligent closed-loop control algorithm based on model predictive control and feedforward compensation Core: Breaking through the limitations of traditional single PID feedback control, a hierarchical intelligent control architecture of "model predictive control + respiratory motion feedforward compensation" is constructed. The bottom layer ensures stable body surface contact pressure through a force / position hybrid control strategy, while the upper layer uses feedforward compensation to predictively offset systematic and physiological errors (such as respiratory motion disturbances). While achieving "rigid fixation" of body position and ensuring positioning accuracy, it also takes into account "flexible contact" of the human body, realizing a dynamic balance between body position stability and patient treatment comfort.

[0102] Key protection areas include: force / position hybrid control logic, respiratory motion feedforward compensation mechanism, and implementation of hierarchical control architecture, addressing the technical pain points of traditional control algorithms that cannot simultaneously ensure both stability and comfort, and that error compensation is lagging.

[0103] 4. The REAC-Net model, an adaptive control network for radiotherapy exoskeletons, overcomes the limitations of traditional artificial intelligence networks. Core Features: The REAC-Net model is a pioneering adaptive control network for radiotherapy exoskeleton designed specifically for high-precision positioning in radiotherapy. Unlike traditional general-purpose AI networks, it possesses four core innovative features, adapting to the non-stationary data characteristics of radiotherapy scenarios and the clinical requirements of precision radiotherapy: ① Dynamic Causal Structure Learning Capability: By decoupling causal and non-causal factors in positioning errors, it accurately identifies the true causes of errors, avoiding decision-making biases caused by non-causal interference; ② Multimodal Adaptive Fusion Mechanism: It innovatively combines the Transformer architecture with graph neural networks to achieve dynamic optimal fusion of multi-source sensor data, improving data utilization and anti-interference capabilities; ③ Closed-Loop Online Learning Capability: Based on the characteristics of non-stationary medical data, a dynamic parameter update mechanism is designed, enabling the model to continuously adapt to individual patient anatomical features and physiological movement differences; ④ Interpretability Enhancement Design: Employing a fully connected signal flow (SiFu) mechanism, each decision step of the model is traceable to specific sensor data, making the decision-making process transparent and traceable, meeting the clinical requirements of precision radiotherapy.

[0104] Key protection areas: The overall architecture of the REAC-Net model and the implementation methods of its four core innovative features (dynamic causal structure learning, multimodal fusion mechanism, online learning mechanism, and interpretable design), which address the technical shortcomings of traditional artificial intelligence networks, such as poor adaptability, weak anti-interference, lack of interpretability, and inability to adapt to non-stationary medical data.

[0105] 5. A comprehensive radiation compatibility engineering system that integrates design and verification. Core: Breaking through the limitations of existing technologies that only focus on the selection of a single material, we innovatively establish a comprehensive radiation compatibility engineering system that spans the entire process from "material selection - structural optimization - dose simulation - field verification - clinical quality control". In terms of materials, we select low-radiation-attenuation materials such as carbon fiber reinforced PEEK composite materials and zirconia ceramic bearings. In terms of structure, we optimize the design according to the beam eye angle (BEV) avoidance principle. Through dual verification by Monte Carlo dose simulation and actual phantom measurement, combined with clinical quality control procedures, we ensure that the exoskeleton device provides strong mechanical support while minimizing interference with therapeutic radiation (6MV photon beam dose disturbance <1%), meeting the core physical requirements of fixation devices for precise radiotherapy.

[0106] Key protection areas include: the construction process of a comprehensive radiation compatibility engineering system, material selection standards, structural optimization principles, and dose simulation and experimental verification methods, to solve the technical problems of poor radiation compatibility, excessive dose disturbance, and inability to adapt to precise radiotherapy in traditional fixation devices.

[0107] 6. An integrated intelligent body positioning device encompassing initialization, perception, decision-making, execution, and closed-loop optimization. Core Concept: Breaking away from the limitations of traditional radiotherapy exoskeletons as merely "passive fixation devices," this device redefines itself as an "intelligent active fixation carrier," establishing an integrated intelligent positioning fixation architecture encompassing the entire process of "initialization-sensing-decision-execution-closed-loop optimization." Utilizing multimodal sensors to perceive patient position, respiratory movements, and even micro-tremors in real time, the device autonomously performs situational assessment, decision analysis, and positional fine-tuning via the REAC-Net model algorithm. Simultaneously, an online learning mechanism is introduced to continuously iterate and optimize model parameters, achieving long-term improvement in positioning accuracy and personalized patient adaptation, constructing a closed-loop positioning maintenance system with self-learning and self-correction capabilities. This device can continuously suppress positioning errors during tens of minutes of radiotherapy, making it particularly suitable for ultra-high-precision radiotherapy clinical applications such as SRS, SBRT, and FLASH-RT.

[0108] Key protection points: The overall process architecture of the integrated intelligent body positioning fixation device, the "initialization-sensing-decision-execution-closed-loop optimization" operation logic, and the collaborative linkage mechanism between the device and the REAC-Net model, effectively solve the technical defects of traditional fixation devices that can only passively fix the body, cannot dynamically compensate for positional errors, and have poor individual patient adaptability.

[0109] The following example illustrates a simulation scenario: AI-driven exoskeleton-assisted multi-position radiotherapy fixation. 1. Example Background and Scene Setting Simulation subjects: Adult patients with lung tumors who cannot tolerate conventional supine / prone positions and require stereotactic body radiotherapy (SBRT) in a seated position.

[0110] Treatment site: Lung tumor target area.

[0111] Simulation environment: Virtual radiotherapy linear accelerator treatment room, equipped with an AI-driven exoskeleton-assisted multi-position radiotherapy fixation device.

[0112] Simulation duration: 5 minutes (300 seconds) for a single treatment session.

[0113] Set a challenge: a. Position Adaptation Challenge: Patients with lung tumors have limited positioning and need to be seated to complete radiotherapy, which tests the device's ability to adapt to special positions. b. Positioning accuracy challenge: Stereotactic body radiotherapy (SBRT) is a high-dose, high-precision radiotherapy technique, which tests the device's ability to achieve sub-millimeter-level positional repeatability accuracy throughout the entire single treatment process and in fractionated treatments; c. Challenges in handling sudden disturbances: During radiotherapy (at 150 seconds), a patient is simulated to cough violently due to lung tumor stimulation, resulting in a sudden body position shift disturbance. This tests the device's REAC Net model's ability to quickly identify sudden body position shifts and its fully automatic closed-loop adaptive correction capability. d. Interference resistance and fault tolerance challenge: During radiotherapy (at 200 seconds), a single set of ceiling ToF sensors is simulated to be affected by electromagnetic interference in the machine room, resulting in abnormal data noise, which tests the device's anti-interference mechanism and fault tolerance capability; f. Comfort and balance control challenges: Uneven pressure distribution on the patient's body surface during prolonged seated radiotherapy tests the device's REAC Net model's ability to adaptively control comfort and maintain balance.

[0114] 2. System initialization and hardware deployment (simulation configuration) Initialization is completed according to the technical solution of this invention: 2.1 Deployment of the bionic exoskeleton module: It adopts a biomimetic topology structure that allows for detachable and combinable head, neck, torso, and limbs, which can be quickly assembled and adapted according to the body parameters of the simulated patient; the system has preset sitting posture locking parameters; the main structure is made of carbon fiber reinforced PEEK composite material, and the joints are equipped with low electron density zirconia ceramic bearings; the overall structure strictly follows the beam eye angle (BEV) avoidance design to accurately avoid radiation shielding.

[0115] 2.2 Deployment of Multimodal Sensing and Positioning Module: Six sets of ToF sensors were deployed in the ceiling array of the virtual radiotherapy treatment room to construct a global three-dimensional coordinate system for the treatment room. The sensor data acquisition frame rate was set to 30fps. An attitude sensing unit with an integrated 9-axis IMU and an absolute magnetic encoder was embedded in each joint of the exoskeleton. A thin-film pressure sensor matrix was deployed across the entire contact layer between the exoskeleton and the human body surface to collect two-dimensional body surface pressure distribution data in real time. All sensing devices established a real-time data link with the REAC-Net module to complete timestamp alignment, data noise reduction, and initial coordinate calibration.

[0116] 2.3 Six Degrees of Freedom Driven and AI System Startup: The linear piezoelectric actuator and rotary torque motor have completed the zero-point calibration of all joints and have the precision adjustment capability of six degrees of freedom of three-dimensional translation and three-dimensional rotation. Load the REAC Net deep learning model, initialize the preset configuration of coordinate transformation matrix, PID control basic parameters, and sensor confidence threshold; at the same time, set the human comfort pressure range of 8-12 kPa and the postural error warning threshold of 2 mm; the system is linked to the radiotherapy treatment planning system (TPS) to import the patient's lung target area sitting reference postural data.

[0117] 3. Simulation Execution Flow (by Timeline) Phase 1: System Positioning Initialization and Baseline Establishment (t=0s) Operation: The patient positions themselves in the bionic exoskeleton device in a preset treatment sitting posture. The system calls the preset target area positioning reference parameters in TPS to complete the initial coarse positioning.

[0118] Actions of this invention: 1. Six sets of ToF sensors simultaneously collect 3D point cloud data of the patient's seated body surface and exoskeleton frame, which are then converted into voxel networks by an algorithm to establish a unified global spatial reference coordinate system; 2. The joint end 9-axis IMU and absolute magnetic encoder acquire the initial joint rotation angle, spatial attitude and position information in real time; 3. The thin-film pressure sensor collects the initial seated body surface pressure distribution, and the REAC-Net model completes the initial adaptive matching of PID control parameters; 4. The system runs a coordinate transformation mathematical model to construct a three-level coordinate mapping relationship of "treatment equipment-exoskeleton base-patient" and synchronously archives the coordinate parameters of sitting posture treatment.

[0119] Compared with existing technologies: Traditional radiotherapy fixation devices (such as thermoplastic films and vacuum negative pressure pads) only support standard positions such as supine and prone, and cannot achieve effective fixation for seated radiotherapy; moreover, they lack a global digital coordinate reference and rely on manual positioning, resulting in large positioning errors (usually greater than 2mm) and poor position repeatability for fractional treatment, making it difficult to meet the high-precision radiotherapy requirements of SBRT.

[0120] Phase Two: Dynamic Perception and Intelligent Compensation of Respiratory Movements (t=0s-150s) Scenario: The patient maintains a stable seated posture and breathes spontaneously, with the chest rising and falling periodically and the respiratory rate at 0.25 Hz, completing the exoskeleton-assisted positional fixation.

[0121] Actions of this invention: 1. The REAC-Net model feature extraction layer performs step-by-step operations, successively completing the voxelization processing of ToF point cloud data, the extraction of respiratory time-series features of 9-axis IMU by LSTM network, the capture of micro-displacement of magnetic encoder joints by embedding encoding, and the extraction of local features of body surface pressure distribution by CNN convolutional network, comprehensively capturing multi-dimensional signal changes caused by respiratory motion. 2. The dynamic fusion layer integrates Transformer multi-head attention and graph neural network architecture, dynamically allocates the confidence weights of each sensor in real time, introduces causal masking to accurately distinguish between physiological respiratory movements and abnormal body position shifts, and effectively filters environmental electromagnetic and structural noise. 3. The decision output layer relies on the physiological motion prediction algorithm to predict the timing pattern of respiratory fluctuations in real time and generate feedforward compensation control commands. 4. The six-degree-of-freedom drive module links the linear piezoelectric actuator and the rotary torque motor to achieve sub-millimeter-level dynamic fine-tuning; 5. The pressure sensor dynamically monitors the pressure on the patient's body surface during the entire sitting posture, combined with REAC. The Net model provides real-time adjustment to keep the surface contact pressure stable within the ideal comfort range of 8–12 kPa, thus avoiding localized pressure.

[0122] Verification results: The REAC-Net network autonomously identifies and filters physiological respiratory movements without the need for additional respiratory gating devices. The positional error is stably controlled within the sub-millimeter level throughout the process, with no invalid false alarms.

[0123] Phase 3: Sudden cough triggers postural shift and closed-loop adaptive correction (t=150s) Scenario: 150 seconds into radiotherapy, a simulated patient experiences a sudden, severe cough due to lung discomfort, causing body swaying and significant target area positional shift. Real-time data acquisition by a multimodal sensor reveals a patient positional shift of -5mm along the X-axis, 3mm along the Z-axis, and 3° around the X-axis, exceeding the error warning threshold.

[0124] Actions of this invention: 1. Multimodal sensor networks capture real-time features of point cloud topological abrupt changes, joint posture anomalies, and abnormal surface pressure distribution. 2. The REAC-Net error calculation module decouples causal and non-causal factors, accurately quantifying the six-degree-of-freedom positional offset; 3. The motion planning module uses an improved bidirectional RRT algorithm to generate an optimal fine-tuned trajectory that balances accuracy, comfort, and structural stability; 4. The six-degree-of-freedom drive module has a fast response, completing sub-millimeter-level attitude closed-loop correction within 3 seconds and returning to the reference position; 5. The system synchronously optimizes PID control parameters online, fine-tunes the support points of the exoskeleton for sitting posture, and suppresses subsequent secondary body slippage; 6. The system is linked with the radiotherapy equipment in real time, and intelligently starts and stops the radiation in a timely manner according to the body position deviation. After completing the sub-millimeter level closed-loop compensation correction, it continuously locks the sitting posture and maintains high-precision body position stability throughout the process.

[0125] Verification Results: The system can accurately identify sudden postural shifts induced by coughing within seconds. Relying on the REAC-Net artificial intelligence algorithm, it achieves fully automatic closed-loop adaptive correction and synchronously links with radiotherapy equipment to intelligently control the start and stop of radiation. No manual intervention or secondary positioning is required, thus avoiding the clinical risks of target deflection and excessive radiation to normal tissues from the source.

[0126] Phase 4: Sensor Data Interference Resistance and Fault Tolerance Simulation (t=200s) Scenario: When radiotherapy is in progress for 200 seconds, the simulated single-group ceiling ToF sensor is affected by electromagnetic interference in the machine room, resulting in abnormal data noise and an occasional single-point failure of the simulation equipment.

[0127] Actions of this invention: 1. The confidence detection module of the REAC-Net model monitors the data quality of each sensor in real time. If the confidence of an abnormal ToF sensor drops significantly (below the preset threshold), the primary circuit breaker fault tolerance mechanism is immediately and automatically triggered. 2. The system dynamically reduces the fusion weight of abnormal sensors and automatically switches to use the high-confidence data from the remaining 5 groups of ToF sensors, 9-axis IMU and absolute magnetic encoder as the main fusion source; 3. The adaptive update layer iterates the model fusion parameters online, intelligently shields non-causal noise interference, and ensures that the body position positioning accuracy is not reduced and the system operation is uninterrupted throughout the process.

[0128] Verification results: It has the ability to self-identify single-point sensor faults, self-weight adaptation, and autonomous fault tolerance. Single-point anomalies do not affect the overall positioning accuracy of the machine or the clinical radiotherapy process.

[0129] Phase 5: Simulation of Comfortable Posture Pressure Control Throughout the Entire Process (300 seconds) Scenario: Patients undergoing radiotherapy in a fixed sitting position for extended periods are prone to local pressure, soreness, and discomfort, which can then induce involuntary body movements.

[0130] Actions of this invention: 1. A thin-film pressure sensor matrix continuously collects the two-dimensional body surface pressure distribution during a seated posture. 2. The PID optimization module of the REAC-Net model dynamically and adaptively adjusts the exoskeleton's support force and contact posture; 3. When the local pressure approaches the upper limit of the threshold, the system automatically fine-tunes the support structure to avoid soft tissue compression damage and always maintains a stable 8-12 kPa comfort range for the human body.

[0131] Verification results: It perfectly achieves a two-way balance between rigid fixation during radiotherapy and human sitting comfort, significantly reducing unconscious postural shifts induced by physical discomfort, and improving patient compliance and postural stability throughout the treatment.

[0132] 4. Simulation Result Data Comparison Table Table 2 compares the performance of the simulation examples of this invention with existing traditional radiotherapy fixation devices.

[0133] 5. Conclusion This simulation example closely addresses real-world clinical challenges, specifically targeting the unique scenario of lung cancer patients who cannot lie supine and must undergo SBRT radiotherapy in a seated position. It fully simulates complex conditions such as physiological respiratory disturbances, sudden coughing and positional shifts, sensor electromagnetic interference, and continuous seated pressure comfort control.

[0134] Simulation results fully verify that this invention, relying on modular bionic exoskeleton, multimodal perception and positioning, six-degree-of-freedom precision actuation, and the REAC-Net deep learning algorithm, effectively solves the core shortcomings of traditional fixation devices, such as limited body position adaptability to sitting posture, low positioning accuracy, lack of dynamic compensation, poor comfort, and lack of fault tolerance. It can achieve sub-millimeter-level precise fixation in special body positions, intelligent physiological motion compensation, automatic correction of sudden deviations, and adaptive pressure balance control, fully adapting to the clinical needs of high-precision radiotherapy such as SBRT, and possesses outstanding technological innovation and strong clinical application and promotion value.

[0135] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0136] 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0137] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0138] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device, comprising a bionic exoskeleton module, a multimodal sensing and positioning module, a six-degree-of-freedom drive and execution module, and a REAC-Net artificial intelligence control module; characterized in that: The bionic exoskeleton module adopts a bionic joint structure with detachable and combinable head, neck, torso, and limbs, which can support and lock the patient in various radiotherapy positions such as supine, prone, lateral, sitting, and semi-standing. It is also made of low-radiation-disturbing material to reduce radiotherapy radiation dose interference. The multimodal sensing and positioning module is equipped with spatial positioning sensors, posture sensing units and body surface pressure sensing units, which are used to collect multimodal data of treatment room spatial position, exoskeleton joint posture, patient body surface contact pressure and physiological movement in real time, and transmit them to the REAC-Net artificial intelligence control module. The six-degree-of-freedom drive execution module is driven and cooperates with the bionic exoskeleton module, and is controlled and connected to the REAC-Net artificial intelligence control module. It can receive control commands to drive the bionic exoskeleton module to complete three-dimensional translation and three-dimensional rotation six-degree-of-freedom posture adjustment, and realize sub-millimeter-level positioning and dynamic compensation of physiological movement in radiotherapy. The REAC-Net AI control module performs feature fusion, positional error calculation, physiological motion prediction, and adaptive optimization on the received multimodal sensor data, generates motion control commands, and sends them to the six-degree-of-freedom drive execution module. This constructs an intelligent control closed loop of initialization-perception-decision-execution-closed-loop optimization, enabling precise fixation of radiotherapy position, dynamic error compensation, and adaptive control of contact pressure.

2. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 1, characterized in that: The spatial positioning sensor includes a ToF array sensor on the ceiling of the treatment room. The array of ToF sensors consists of 6 groups, which are installed on the ceiling of the treatment room and connected to the REAC-Net artificial intelligence control module via a wired link. The array of ToF sensors is used to construct a global treatment coordinate system, collect the three-dimensional spatial position information of the patient's body surface and the exoskeleton frame in real time, and convert the collected 3D point cloud data into a voxel network. This provides a unified spatial reference for the entire body positioning process and can monitor the relative displacement of the patient's body surface and the exoskeleton frame in real time, capturing body position offset signals in a timely manner.

3. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 2, characterized in that: The attitude sensing unit includes an integrated 9-axis inertial measurement unit and an absolute magnetic encoder, which are embedded in each joint of the exoskeleton and establish a real-time data transmission connection with the REAC-Net artificial intelligence control module. It is used to collect the three-dimensional rotation angle, angular velocity and spatial attitude data of each joint of the exoskeleton in real time, and synchronously feedback the motion state of the exoskeleton, providing a key basis for judging the degree of matching between the exoskeleton attitude and the treatment reference position.

4. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 1, characterized in that: The REAC-Net artificial intelligence control module includes a PID parameter optimization module based on the deep learning REAC-Net model. This module performs real-time analysis and regulation based on measured pressure data to stably control the body surface contact pressure within the ideal comfortable range of 8-12 kPa. When the pressure exceeds this threshold, the system automatically triggers an alarm signal and drives the support structure to perform adaptive fine-tuning.

5. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 4, characterized in that: The processing steps of the REAC-Net artificial intelligence control module include: Coordinate Transformation Mathematical Model: A dedicated coordinate transformation link is constructed. The transformation matrix parameters are iteratively optimized through the deep learning REAC-Net model to effectively compensate for the effects of equipment installation errors and environmental thermal drift, reduce cumulative system errors, and ensure the accuracy of coordinate system mapping at each level. The specific transformation model is as follows: in, This is the transformation matrix from the treatment device coordinate system to the exoskeleton base. Here is the attitude rotation matrix. For thermal drift compensation, Let these be the patient's coordinates in the coordinate system of the treatment device. This model provides the coordinates of the exoskeleton joints in its own coordinate system, enabling precise calculation of the patient's coordinates and providing a coordinate basis for postural adjustment.

6. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 5, characterized in that: The processing steps of the REAC-Net artificial intelligence control module also include a deep learning REAC-Net model algorithm subsystem. The system includes a five-layer hierarchical architecture, including a multimodal input layer, a feature extraction layer, a dynamic fusion layer, a decision output layer, an adaptive update layer, and a closed-loop link that interacts sequentially, with the closed-loop link returning to the feature extraction layer. The multimodal input layer receives raw data collected by various sensors; the feature extraction layer performs preprocessing and feature extraction of each modality; the dynamic fusion layer realizes intelligent fusion of multimodal features and noise filtering; the decision output layer generates positional error calculation results, motion control commands, and parameter optimization schemes; the adaptive update layer realizes online iteration and fault-tolerant adjustment of model parameters; and the model performance is continuously optimized through a closed-loop link to ensure adaptation to the complex physiological movements and environmental changes required during radiotherapy.

7. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 6, characterized in that: The feature extraction layer completes the preprocessing and feature extraction of data from each modality, including targeted preprocessing and feature extraction of raw data from ToF sensors, 9-axis IMUs, absolute magnetic encoders, and thin-film pressure sensors, converting heterogeneous multimodal data into unified and fusionable feature vectors. The processing methods and mathematical expressions for each modality of data are as follows: ToF sensor data: The acquired 3D point cloud data is converted into a voxel network to extract spatial location features and capture the relative positional relationship between the patient's body surface and the exoskeleton. The expression is: in, This represents the number of point clouds; The point cloud is represented by its 3D spatial coordinates; the point cloud is converted into a voxel mesh through voxelization. Its three-dimensional parameters are defined as follows: height H corresponds to the number of voxels in the cranio-coccygeal direction of the human body, width W corresponds to the number of voxels in the left and right directions of the human body, and depth D corresponds to the number of voxels in the ventral and dorsal directions of the human body. 9-axis IMU data: Temporal features are extracted using a Long Short-Term Memory network to adapt to the temporal characteristics of IMU data and capture the dynamic changes in exoskeleton posture. The expression is as follows: in, Let be the hidden state of the LSTM at time t; Let t be the input feature vector; LSTM is a short-term memory network; the 9-dimensional column vector corresponds to the output data of the 9-axis IMU, and T represents the transpose. For acceleration, Angular velocity, The magnetic field strength; Magnetic encoder data: Converted into learnable embedding vectors through position encoding, this captures subtle changes in joint position, providing accurate feature support for pose tracking and positioning error prediction. The expression is: in, The output position embedding vector serves as the input feature for subsequent pose tracking and positioning error prediction models, capturing subtle changes in joint position. The embedding function is a feature transformation module in deep learning, used to convert raw location data into high-dimensional feature vectors. The output vector of the absolute magnetic encoder is used to collect raw data on the angle and position of the exoskeleton joints. Pressure sensor data: Local pressure distribution features are extracted using a convolutional neural network to reflect the contact state between the patient's body surface and the exoskeleton, providing support for pressure regulation and postural stability assessment. The expression is as follows: in, The output pressure feature vector is used for postural stability assessment and positioning deviation early warning; The two-dimensional pressure distribution image corresponds to the collected pressure distribution data on the patient's body surface, which is used to reflect the contact state between the patient's body and the fixation device. It is a real matrix with M rows and N columns, corresponding to the pixel dimension of the pressure distribution image.

8. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 7, characterized in that: The dynamic fusion layer combines Transformer and graph neural network, and its specific implementation process is as follows: Dynamic confidence-aware weight allocation: fusion weights It is over time The dynamically changing matrix is ​​set by the sensor to a real-time confidence vector. The decision was made to ensure the dominant role of high-confidence sensor data. Sensor confidence calculation: For the A sensor, its confidence level at time t Taking into account two core factors: firstly, its measured value Compared with historical stable values The degree of deviation, and the health status of the sensor itself. The expression is: in: For sensors The historical average measurement value reflects the stable operating state of the sensor; For sensors The noise standard deviation is calibrated by the sensor's factory parameters and clinical measurement data. The sensor's working status is fed back in real time to provide feedback on the hardware health reported by the device driver layer; Dynamic fusion weight generation: The confidence scores are converted into normalized fusion weights using the Softmax function. At the same time, temperature parameters are introduced. To control the smoothness of the weight distribution, the expression is: Where N is the total number of sensors, the normalization process ensures that the sum of the weights of all sensors is 1, thus achieving a reasonable allocation of weights; Dynamic weighted fusion output: set up For the first The feature vectors extracted from each modality are then fused into a single feature vector. for: in It is a mode-specific projection matrix; Multi-head attention mechanism for causal perception: Introducing causal masks into attention mechanisms This enables effective differentiation between causal and non-causal factors, improving the reliability of fusion features; Query, key, and value generation: For the input sequence , For sequence length, Generate a query matrix based on the feature dimensions. Key matrix and value matrix This enables feature mapping and association capture, expressed as: Among them, W Q W K W V The learnable projection weight matrix is ​​a parameter optimized during model training, used to map input features to three different feature spaces: query Q, key K, and value V. Causal similarity calculation: Add causal masking to standard dot product attention. This mask is generated in real time by the dynamic causality discovery module and is used to filter out non-causally correlated time steps or sensor nodes. The expression is: Learning about dynamic graph structures: The connection weights between nodes (i.e., sensors) in a graph neural network are not fixed, but rather the adjacency matrix is ​​dynamically constructed based on the correlation of real-time data. This enables dynamic correlation and capture of sensor features; Dynamic adjacency matrix construction: The real-time correlation between sensors is measured using the inner product of node features. The correlation result is compressed to the (0,1) interval using the Sigmoid function to form dynamic edge weights, expressed as: in, , Let be the real-time feature vectors of sensor nodes i and j at time t, corresponding to the sensor monitoring data at that time; σ is the Sigmoid activation function, which compresses the result of the inner product of features to the (0,1) interval to ensure the rationality of the edge weights. The larger the value, the stronger the real-time correlation between the two sensor nodes. Dynamic graph convolution: Combining dynamic adjacency matrix Sum-degree matrix A variant of GCN is used for graph convolution operations, emphasizing neighboring nodes. For the central node The influence of sensor features is considered to achieve structured fusion, expressed as: in: For nodes The set of neighbors of node i, that is, the sensor nodes that are strongly correlated with node i; For the first The learnable weight matrix of the layer; The activation function is used to introduce non-linear features and enhance the model's expressive power; the degree matrix... Used for normalization processing to avoid feature distortion caused by differences in node degree and ensure the rationality of graph convolution results; Dynamic weight allocation: The fusion weights of graph nodes are automatically adjusted based on sensor confidence levels to further enhance the role of high-confidence sensors. The expression is as follows: in, Let be the real-time confidence level of sensor i. This is the weight adjustment coefficient, used to control the degree of influence of confidence level on node weights; the confidence level is calculated by combining historical error data and real-time sensor operating status to ensure the accuracy of weight allocation; Synthetic dynamic fusion function: Integrating the aforementioned dynamic confidence perception weight allocation, causal perception multi-head attention mechanism, and dynamic graph structure learning process, the final output of the dynamic fusion layer is... Represented as a composite function, this function not only achieves efficient fusion of multimodal data, but also improves the accuracy and robustness of the fused features by filtering noise and non-causal interference through causal filtering and graph structure optimization. The expression is: in, , , These are feature vectors extracted from the ToF sensor, 9-axis IMU, and absolute magnetic encoder, respectively.

9. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 8, characterized in that: The decision output layer is used to calculate the six-degree-of-freedom body position error, generate motion planning instructions, and optimize PID parameters online based on the feature vector output by the dynamic fusion layer. This provides control instructions to the exoskeleton drive unit, enabling sub-millimeter-level body position adjustment and pressure stability control. Including the following: Error Calculation Module: Through decoupled causal factor analysis, it distinguishes between causal and non-causal components in postural errors, improving the accuracy of error calculation. The expression is as follows: in, This is a causal relationship function used to extract the causal components from the error; It is a non-causal function used to extract the non-causal components in the error; the causal and non-causal factors are determined through dynamic causal structure learning to ensure that the error calculation results can truly reflect the actual deviation of the patient's body position. Motion planning module: Employing an improved bidirectional RRT algorithm, which balances the accuracy of postural adjustment, patient comfort, and fixation stability, it generates the optimal motion trajectory, expressed as: in, The movement posture of exoskeleton joints; The distance error between the current posture and the target posture is used to ensure the accuracy of body position adjustment; The patient comfort index is fed back from pressure sensor data; As an indicator of body position stability, it is comprehensively evaluated based on multimodal sensor data; The adaptive weights are dynamically adjusted based on real-time patient feedback to achieve a balance between accuracy, comfort, and stability. PID parameter optimization module: This module uses deep reinforcement learning to adjust PID control parameters online, ensuring the exoskeleton contact pressure remains stable within the ideal range of 8-12 kPa, balancing fixation stability and patient comfort. The expression is: in, , , The initial values ​​for the PID parameters are calibrated based on clinical experience and model training. , , The amount of parameter adjustment for the output of the deep learning model.

10. The artificial intelligence-driven exoskeleton-assisted multi-position radiotherapy fixation device according to claim 9, characterized in that: The adaptive update layer is used to enable the REAC-Net model to learn autonomously online and update its parameters iteratively. It also constructs a three-level fault tolerance mechanism, as detailed below: Dynamic causal structure learning: Employing a dynamic causal variational autoencoder, this method achieves accurate separation and modeling of causal and non-causal factors, providing support for error calculation and feature fusion. The loss function expression is as follows: in, Divided into causal factors Dynamic non-causal factors and static non-causal factors ; For approximate posterior distribution, To generate a distribution, It is the prior distribution; To reconstruct the loss, ensure that the model can accurately reconstruct the input data; KL divergence is used to constrain the distribution of latent variables. These are constraint coefficients; Online learning mechanism: A sliding window-based data update strategy is employed to achieve online iterative optimization of model parameters, adapting to individual differences among patients and dynamic changes during treatment. The expression is: in, Let be the model parameters at time t+1. The model parameters are at time t; The adaptive learning rate is dynamically adjusted based on the model training error to ensure the stability and convergence of parameter updates. Let be the gradient of the loss function at time t. This includes multimodal sensing data and error feedback data within the current sliding window; Fault tolerance mechanism: A three-tiered circuit breaker system is constructed. When a sensor malfunctions or data anomalies, it automatically switches to the backup plan to ensure stable system operation and avoid treatment interruptions or positioning errors caused by equipment failure. The specific tiers are as follows: Primary circuit breaker: If a single sensor's data is abnormal, the system automatically reduces the confidence weight of that sensor and replaces it with data from other high-confidence sensors to ensure the accuracy of the fused features; Intermediate circuit breaker: When there is a data conflict between multiple sensors, the system initiates a data calibration process, combines historical data with causal analysis, and filters out reliable data to ensure the accuracy of decision output; Advanced circuit breaker: System-level fault detection, the system automatically triggers an alarm signal and switches to manual control mode, while retaining the current body position data to support clinical emergency treatment and ensure treatment safety.

Images