A precise positioning guide system for hyperthermia care

Abstract

The application provides a kind of accurate positioning guide system for hyperthermia nursing, method is first through deep learning algorithm to CT, MRI, PET-CT multimodal image is registered fusion and intelligent synchronous segmentation, accurately identify tumor and surrounding key structure;Further construct the model of fusion multi-objective optimization and reinforcement learning, generate personalized treatment path;In treatment execution, real-time compensation physiological movement is realized accurate tracking by using hybrid prediction model;And continuously optimize system through real-time feedback data;Corresponding system includes image fusion, intelligent segmentation, path planning, real-time guidance and monitoring feedback unit, is related to medical navigation and intelligent control technical field, wherein, it realizes the change from experience dependence to intelligent self-adaptation in treatment positioning, significantly improves positioning accuracy, treatment safety and efficiency, is applicable to precise medical scenarios such as tumor hyperthermia.

Description

Technical Field

[0001] This invention relates to the field of medical navigation and intelligent control technology, and in particular to an accurate positioning and guidance system for thermotherapy nursing. Background Technology

[0002] As an important treatment for tumors, the efficacy of hyperthermia for tumors highly depends on the precise positioning of the treatment equipment. However, current clinical positioning technologies face many challenges: Relying on operator experience and marking the body surface using two-dimensional medical images has problems such as large subjective errors, low positioning accuracy (often exceeding 10mm), and poor repeatability, making it difficult to meet the requirements of modern precision medicine.

[0003] While using rigid structures to fix patients and treatment equipment can improve stability, it lacks flexibility, cannot adapt to individual anatomical differences, and cannot compensate for physiological movements (such as breathing and heartbeat) during treatment, resulting in poor patient comfort.

[0004] Relying solely on single imaging data such as CT or MRI provides limited information dimensions and lacks a comprehensive assessment of tumor metabolic activity and surrounding dangerous organs, making it difficult to achieve individualized optimal pathway planning.

[0005] Spatial positioning using surface markers is susceptible to marker displacement, line-of-sight obstruction, or electromagnetic interference, and only provides spatial location information, lacking intelligent decision-making and adaptive adjustment capabilities.

[0006] In recent years, while multimodal image fusion technology and artificial intelligence algorithms have made progress in the field of medical diagnostics, solutions for their deep integration with treatment execution devices to form a real-time closed-loop intelligent guidance system remain lacking. In particular, how to effectively fuse the high-resolution anatomical information of CT, the superior soft tissue contrast of MRI, and the functional metabolic information of PET, and use deep learning for precise analysis to guide the real-time dynamic adjustment of treatment equipment, is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a positioning guidance method and system with high positioning accuracy, high intelligence, and real-time adaptability to physiological movements. Specific objectives include: achieving automatic registration and deep fusion of multimodal image data; simultaneously and accurately segmenting multiple key anatomical structures using an improved deep learning model; generating personalized optimal treatment paths based on multi-objective optimization and reinforcement learning; achieving sub-millimeter-level stable tracking by real-time compensation of target area movement through advanced predictive models; and constructing a complete intelligent closed loop of perception-decision-execution-feedback to improve the safety and effectiveness of treatment.

[0008] The technical solution adopted by the embodiments of this application to solve its technical problem is: A precise positioning guidance system for thermotherapy care includes the following steps: S1. Multimodal image acquisition and fusion: Simultaneously acquire CT, MRI and PET-CT image data of the same patient, and perform three-dimensional spatial registration and fusion through a deep learning-based multi-scale feature matching algorithm to generate fused image volume data; S2. Intelligent synchronous segmentation of anatomical structures: The improved 3DU-Net++ neural network, which integrates image input, integrates attention gating mechanism and dense connection structure. In one forward propagation, it synchronously outputs three-dimensional segmentation masks of tumor area, dangerous organ area, and vascular and nerve structure, and generates segmentation uncertainty heat map based on Monte Carlo Dropout. S3. Personalized treatment path planning: Based on the segmentation mask, a multi-objective optimization model is established. The model takes maximizing the tumor area coverage as the first objective, minimizing the radiation dose to dangerous organs as the second objective, and minimizing the motion complexity of the treatment path as the third objective. A parameter adaptive adjustment mechanism based on reinforcement learning is introduced to output the optimal treatment path scheme including spatial coordinates, incident angle, and motion trajectory. S4. Real-time dynamic positioning guidance and compensation: Based on the optimal treatment path plan, the actuator is driven to perform positioning. At the same time, a hybrid prediction model combining Kalman filter and long short-term memory network is used to predict and compensate for target area movement caused by breathing and heartbeat in real time, forming a closed-loop control system. S5. Treatment process monitoring and feedback optimization: During the treatment process, temperature distribution in the treatment area and actual location of the equipment are collected in real time through temperature and position sensors. The data is fed back to the multi-objective optimization model to dynamically adjust the treatment parameters and iteratively optimize the reinforcement learning model based on historical treatment data.

[0009] Preferably, in step S1, the multi-scale feature matching algorithm specifically includes: Extract bony structural features from CT images, soft tissue boundary features from MRI images, and metabolically active region features from PET-CT images. Non-rigid alignment of multiple sets of feature points is performed using a deformable registration network; An adaptive weighted fusion module is used to dynamically allocate fusion weights based on the signal-to-noise ratio and contrast of each modality image in different anatomical regions.

[0010] Preferably, in step S2, the 3DU-Net++ neural network is further connected to an uncertainty evaluation unit, which is used to output the segmentation confidence of each voxel; When the confidence level of a certain region in the segmentation uncertainty heatmap is lower than a preset threshold, a manual review prompt mechanism is triggered, and the region to be reviewed is highlighted on the interactive interface.

[0011] Preferably, in step S3, the mathematical expression of the multi-objective optimization model is: in, For tumor coverage, For radiation dose to dangerous organs, For path motion complexity, The weight coefficients are dynamically adjusted through reinforcement learning, and satisfy the following conditions: + + =1.

[0012] Preferably, in step S3, the parameter adaptive adjustment mechanism based on reinforcement learning specifically includes: Construct a Markov decision process model with treatment pathway parameters as actions and real-time treatment feedback and long-term efficacy evaluation as rewards; Using the deep deterministic policy gradient algorithm to adjust the weight coefficients Conduct online learning and optimization.

[0013] Preferably, in step S4, the method for constructing the hybrid prediction model includes: An LSTM network was trained using preprocessed temporal image data to learn respiratory movement cycle patterns. By using the output of the LSTM network as the observation input of the Kalman filter and combining it with the kinematic model of the actuator, the position of the target area within the next 0.5-1.0 seconds can be predicted.

[0014] Preferably, step S4 further includes a dynamic constraint step for safety boundaries: Based on the real-time distance between the vascular and neural structures in the segmentation mask and the target area, the safety protection boundary is dynamically calculated and updated; When the actuator trajectory is predicted or detected to potentially intrude into the safety protection boundary, the control module automatically triggers a deceleration or stop command and issues visual and auditory alarms on the human-machine interface.

[0015] Preferably, in step S5, the feedback optimization process specifically includes: Establish treatment efficacy evaluation indicators, which comprehensively consider real-time temperature target achievement rate, positioning trajectory tracking error, and post-treatment imaging changes; Based on this evaluation metric, the constraints in the multi-objective optimization model and the reward function of the reinforcement learning model are fine-tuned in real time using an online learning algorithm.

[0016] It includes a multimodal image acquisition and fusion unit, an intelligent segmentation and analysis unit, a treatment pathway planning unit, a real-time guidance and control unit, a monitoring feedback and optimization unit, and a system bus and communication interface unit; The multimodal image acquisition and fusion unit is used to perform step S1; The intelligent segmentation and analysis unit has the improved 3DU-Net++ neural network built in, which is used to execute step S2; The treatment path planning unit has a built-in multi-objective optimization model and reinforcement learning engine, which is used to execute step S3; The real-time guidance and control unit includes an actuator, sensors, and a hybrid prediction model, and is used to execute step S4; The monitoring feedback and optimization unit is used to execute step S5; The system bus and communication interface unit is used to connect and coordinate the data and instruction transmission between the various units.

[0017] Preferably, the real-time guidance and control unit further includes an augmented reality human-computer interaction unit, used for: The optimal treatment path plan, real-time segmentation results, predicted target area location, and safety boundary are overlaid and displayed in three dimensions on a real-scene video stream on the patient's body surface. It provides gesture and voice interaction interfaces, allowing operators to confirm and fine-tune key points during the guidance process.

[0018] The advantages of the embodiments of this application are: By using multimodal information fusion and deep learning for precise segmentation, the accuracy of the positioning benchmark is guaranteed from the source. Combined with sub-second motion prediction and compensation, sub-millimeter-level stable tracking accuracy is achieved throughout the treatment process, which is far superior to existing technologies. Based on each patient's unique anatomical and functional images, treatment plans are automatically generated and continuously optimized through multi-objective optimization and reinforcement learning, which greatly reduces the reliance on the operator's personal experience.

[0019] The hybrid prediction model effectively overcomes the long-standing challenge of physiological movement in precision treatment, enabling the device to actively adapt to the target area rather than passively waiting or ignoring it; the uncertainty assessment mechanism provides quality monitoring for key segmentation; and the dynamic safety boundary and real-time monitoring form a double insurance to minimize the risk of injury to dangerous organs.

[0020] The system has the ability to continuously learn from clinical feedback. As treatment cases accumulate, its planning and control will become more and more precise and efficient, and its value will increase over time, making high-precision thermotherapy technology easier to promote and apply in clinical practice, benefiting more patients. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the accurate positioning guidance method for thermotherapy nursing of the present invention; Figure 2 This is a schematic diagram of the image fusion and segmentation process of the present invention; Figure 3 This is a flowchart illustrating the accurate positioning guidance system for thermotherapy nursing of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. In addition, for the sake of convenience, the terms "upper," "lower," "left," and "right" are equivalent to the upper, lower, left, and right directions of the accompanying drawings themselves, and the terms "first," "second," etc., are used for descriptive purposes and have no other special meaning.

[0023] This application provides an accurate positioning guidance system for thermotherapy care, solving the problems in the prior art. By using multimodal information fusion and deep learning for precise segmentation, the accuracy of the positioning benchmark is guaranteed from the source. Combined with sub-second motion prediction and compensation, sub-millimeter-level stable tracking accuracy is achieved throughout the treatment process, which is far superior to the prior art. Based on the unique anatomical and functional images of each patient, the treatment plan is automatically generated and continuously optimized through multi-objective optimization and reinforcement learning, greatly reducing the dependence on the operator's personal experience.

[0024] The hybrid prediction model effectively overcomes the long-standing challenge of physiological movement in precision treatment, enabling the device to actively adapt to the target area rather than passively waiting or ignoring it; the uncertainty assessment mechanism provides quality monitoring for key segmentation; and the dynamic safety boundary and real-time monitoring form a double insurance to minimize the risk of injury to dangerous organs.

[0025] The system has the ability to continuously learn from clinical feedback. As treatment cases accumulate, its planning and control will become more and more precise and efficient, and its value will increase over time, making high-precision thermotherapy technology easier to promote and apply in clinical practice, benefiting more patients.

[0026] The technical solution in this application is to solve the above problems, and the overall approach is as follows: Example This embodiment provides an accurate positioning guidance system for thermotherapy care, such as... Figure 1-3 As shown, it includes the following steps: S1. Multimodal image acquisition and fusion: Simultaneously acquire CT, MRI and PET-CT image data of the same patient, and perform three-dimensional spatial registration and fusion through a deep learning-based multi-scale feature matching algorithm to generate fused image volume data; The process involved simultaneously acquiring CT, MRI, and PET-CT images of the patient. A deep learning-based multi-scale feature matching algorithm was used for registration: bony features from CT, soft tissue boundary features from MRI, and metabolically active area features from PET-CT were extracted, and precise spatial alignment was achieved through a deformable registration network. Finally, an adaptive weighted fusion module was used to dynamically synthesize fused image volume data based on the quality of each modality in different anatomical regions, providing a comprehensive data foundation for subsequent analysis.

[0027] S2. Intelligent synchronous segmentation of anatomical structures: The improved 3DU-Net++ neural network, which integrates image input, integrates attention gating mechanism and dense connection structure. In one forward propagation, it synchronously outputs three-dimensional segmentation masks of tumor area, dangerous organ area, and vascular and nerve structure, and generates segmentation uncertainty heat map based on Monte Carlo Dropout. The fused image data is input into a specially designed improved 3DU-Net++ neural network. This network integrates an attention gating mechanism, enabling it to focus on key regions of the tumor boundary, and employs dense connections to enhance feature reuse. In a single forward pass, the network can simultaneously output a high-precision 3D segmentation mask for the tumor region, dangerous organs, and important vascular and neural structures. The network is connected to an uncertainty assessment unit based on Monte Carlo Dropout technology, which generates a segmentation confidence heatmap, providing clear guidance for manual review and ensuring the reliability of key region segmentation.

[0028] S3. Personalized treatment path planning: Based on the segmentation mask, a multi-objective optimization model is established. The model takes maximizing the tumor area coverage as the first objective, minimizing the radiation dose to dangerous organs as the second objective, and minimizing the motion complexity of the treatment path as the third objective. A parameter adaptive adjustment mechanism based on reinforcement learning is introduced to output the optimal treatment path scheme including spatial coordinates, incident angle, and motion trajectory. Based on the segmentation results described above, a multi-objective optimization mathematical model is constructed. The objective functions are: maximizing tumor coverage, minimizing radiation dose to at-risk organs, and minimizing the complexity of the treatment device's movement path. The weight coefficients of the three objectives are not fixed but dynamically and adaptively adjusted through a built-in reinforcement learning engine. This engine constructs a Markov decision process model and continuously learns and optimizes using a deep deterministic policy gradient algorithm, thereby generating a unique optimal spatial coordinates, incident angle, and movement trajectory scheme for each patient.

[0029] S4. Real-time dynamic positioning guidance and compensation: Based on the optimal treatment path plan, the actuator is driven to perform positioning. At the same time, a hybrid prediction model combining Kalman filter and long short-term memory network is used to predict and compensate for target area movement caused by breathing and heartbeat in real time, forming a closed-loop control system. The control actuator positions itself according to the planned path. To overcome target area displacement caused by physiological movements such as breathing and heartbeat, a hybrid prediction model is employed: a Long Short-Term Memory (LSTM) network is used to learn the patient's individualized respiratory movement patterns, and its predicted output is used as the observation value of a Kalman filter. Combined with the device's kinematic model, this achieves accurate prediction of the target area's position 0.5-1.0 seconds in the future, with the feedforward control actuator providing synchronous compensation to form a closed-loop control. The system also features a dynamic safety boundary protection function, monitoring the distance to hazardous structures in real time to ensure treatment safety.

[0030] S5. Treatment process monitoring and feedback optimization: During the treatment process, temperature distribution in the treatment area and actual location of the equipment are collected in real time through temperature and position sensors. The data is fed back to the multi-objective optimization model to dynamically adjust the treatment parameters and iteratively optimize the reinforcement learning model based on historical treatment data.

[0031] During treatment, integrated temperature and position sensors monitor the temperature distribution and actual device orientation in the treatment area in real time. This data is fed back to the path planning module for dynamic fine-tuning of treatment parameters. Simultaneously, based on accumulated treatment data, the reinforcement learning model and evaluation metrics are continuously iterated and optimized, enabling the system to self-improve.

[0032] In step S1, the multi-scale feature matching algorithm specifically includes: Extract bony structural features from CT images, soft tissue boundary features from MRI images, and metabolically active region features from PET-CT images. Non-rigid alignment of multiple sets of feature points is performed using a deformable registration network; An adaptive weighted fusion module is used to dynamically allocate fusion weights based on the signal-to-noise ratio and contrast of each modality image in different anatomical regions.

[0033] In step S2, the 3DU-Net++ neural network is also connected to an uncertainty evaluation unit, which is used to output the segmentation confidence of each voxel; When the confidence level of a certain region in the segmentation uncertainty heatmap is lower than a preset threshold, a manual review prompt mechanism is triggered, and the region to be reviewed is highlighted on the interactive interface.

[0034] In step S3, the mathematical expression of the multi-objective optimization model is: in, For tumor coverage, For radiation dose to dangerous organs, For path motion complexity, The weight coefficients are dynamically adjusted through reinforcement learning, and satisfy the following conditions: + + =1.

[0035] In step S3, the parameter adaptive adjustment mechanism based on reinforcement learning specifically includes: Construct a Markov decision process model with treatment pathway parameters as actions and real-time treatment feedback and long-term efficacy evaluation as rewards; Using the deep deterministic policy gradient algorithm to adjust the weight coefficients Conduct online learning and optimization.

[0036] In step S4, the method for constructing the hybrid prediction model includes: An LSTM network was trained using preprocessed temporal image data to learn respiratory movement cycle patterns. By using the output of the LSTM network as the observation input of the Kalman filter and combining it with the kinematic model of the actuator, the position of the target area within the next 0.5-1.0 seconds can be predicted.

[0037] Step S4 also includes a dynamic constraint step for the safety boundary: Based on the real-time distance between the vascular and neural structures in the segmentation mask and the target area, the safety protection boundary is dynamically calculated and updated; When the actuator trajectory is predicted or detected to potentially intrude into the safety protection boundary, the control module automatically triggers a deceleration or stop command and issues visual and auditory alarms on the human-machine interface.

[0038] In step S5, the feedback optimization process specifically includes: Establish treatment efficacy evaluation indicators, which comprehensively consider real-time temperature target achievement rate, positioning trajectory tracking error, and post-treatment imaging changes; Based on this evaluation metric, the constraints in the multi-objective optimization model and the reward function of the reinforcement learning model are fine-tuned in real time using an online learning algorithm.

[0039] It includes a multimodal image acquisition and fusion unit, an intelligent segmentation and analysis unit, a treatment pathway planning unit, a real-time guidance and control unit, a monitoring feedback and optimization unit, and a system bus and communication interface unit; The multimodal image acquisition and fusion unit is used to perform step S1. The patient is fixed on the treatment bed, and the tumor area image data is acquired sequentially by CT, MRI, and PET-CT equipment. During the acquisition process, respiratory gating technology is used to reduce motion artifacts. The three modal image data are transmitted to a GPU-accelerated server, and feature points are extracted and non-rigid alignment is completed through a multi-scale feature matching algorithm. An adaptive weighted fusion module dynamically allocates weights to generate 1mm thick three-dimensional fused image volume data in DICOM format, which is then transmitted to the intelligent segmentation and analysis unit. The intelligent segmentation and analysis unit, with a built-in improved 3DU-Net++ neural network, is used to execute step S2. The intelligent segmentation and analysis unit receives fused image volume data and inputs it into the improved 3DU-Net++ neural network. After one forward propagation, the neural network outputs a three-dimensional segmentation mask of the tumor region, dangerous organs, and vascular and neural structures. At the same time, the uncertainty assessment unit generates a segmentation uncertainty heatmap. If the confidence score of a certain region in the heatmap is lower than 0.85, the system highlights the region on the interactive interface and triggers a manual review prompt. After the operator confirms or corrects the segmentation, the segmentation result is transmitted to the treatment path planning unit. The treatment path planning unit, with its built-in multi-objective optimization model and reinforcement learning engine, is used to execute step S3. The unit receives segmentation mask data, loads the multi-objective optimization model, and initializes the weight coefficients α=0.6, β=0.3, γ=0.1 (satisfying α+β+γ=1). Based on a reinforcement learning-based parameter adaptive adjustment mechanism, a Markov decision process model is constructed, using treatment path parameters (spatial coordinates, incident angle, and trajectory) as actions and real-time treatment feedback and long-term efficacy evaluation as rewards. The weight coefficients are optimized online through a deep deterministic policy gradient algorithm, outputting the optimal treatment path scheme, which is then transmitted to the real-time guidance and control unit. The real-time guidance and control unit, including actuators, sensors, and a hybrid prediction model, is used to execute step S4. The real-time guidance and control unit receives the optimal treatment path plan and drives the six-axis robotic arm to move to the initial positioning position. The hybrid prediction model learns the patient's respiratory motion cycle pattern through an LSTM network, using the output as the observation input for a Kalman filter. Combined with the robotic arm's kinematic model, it predicts the target area's position in real time within 0.5-1.0 seconds, compensating for motion errors caused by breathing and heartbeat. The augmented reality human-computer interaction unit displays the optimal treatment path, real-time segmentation results, predicted target area position, and safety boundaries in a 3D overlay on a video stream of the patient's real-world scene. The operator confirms key points via gestures or voice. The system calculates the distance between the vascular and neural structures and the target area in real time, dynamically updates the safety protection boundary, and immediately triggers deceleration or stop commands and issues visual and auditory alarms if the robotic arm's trajectory is detected to potentially intrude into the safety boundary. The monitoring, feedback, and optimization unit executes step S5. During the treatment process, temperature and position sensors collect real-time data on the temperature distribution of the treatment area and the actual position of the robotic arm, with sampling frequencies of 100Hz and 200Hz, respectively. The data is transmitted to the monitoring, feedback, and optimization unit. Based on the treatment efficacy evaluation indicators, the monitoring, feedback, and optimization unit comprehensively considers the real-time temperature target achievement rate, positioning trajectory tracking error, and post-treatment image changes. It then uses an online learning algorithm to fine-tune the constraints of the multi-objective optimization model and the reward function of the reinforcement learning model in real time. The fine-tuned parameters are fed back to the treatment path planning unit to dynamically adjust the treatment parameters, thereby achieving closed-loop optimization of the treatment process. The system bus and communication interface unit is used to connect and coordinate the data and command transmission between various units.

[0040] The real-time guidance and control unit also includes an augmented reality human-computer interaction unit for: The optimal treatment path plan, real-time segmentation results, predicted target area location, and safety boundary are overlaid and displayed in three dimensions on a real-scene video stream on the patient's body surface. It provides gesture and voice interaction interfaces, allowing operators to confirm and fine-tune key points during the guidance process.

[0041] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for accurate positioning guidance in thermotherapy nursing, characterized in that, Includes the following steps: S1. Multimodal image acquisition and fusion: Simultaneously acquire CT, MRI and PET-CT image data of the same patient, and perform three-dimensional spatial registration and fusion through a deep learning-based multi-scale feature matching algorithm to generate fused image volume data; S2. Intelligent synchronous segmentation of anatomical structures: The improved 3DU-Net++ neural network, which integrates image input, integrates attention gating mechanism and dense connection structure. In one forward propagation, it synchronously outputs three-dimensional segmentation masks of tumor area, dangerous organ area, and vascular and nerve structure, and generates segmentation uncertainty heat map based on Monte Carlo Dropout. S3. Personalized treatment path planning: Based on the segmentation mask, a multi-objective optimization model is established. The model takes maximizing the tumor area coverage as the first objective, minimizing the radiation dose to dangerous organs as the second objective, and minimizing the motion complexity of the treatment path as the third objective. A parameter adaptive adjustment mechanism based on reinforcement learning is introduced to output the optimal treatment path scheme including spatial coordinates, incident angle, and motion trajectory. S4. Real-time dynamic positioning guidance and compensation: Based on the optimal treatment path plan, the actuator is driven to perform positioning. At the same time, a hybrid prediction model combining Kalman filter and long short-term memory network is used to predict and compensate for target area movement caused by breathing and heartbeat in real time, forming a closed-loop control system. S5. Treatment process monitoring and feedback optimization: During the treatment process, temperature distribution in the treatment area and actual location of the equipment are collected in real time through temperature and position sensors. The data is fed back to the multi-objective optimization model to dynamically adjust the treatment parameters and iteratively optimize the reinforcement learning model based on historical treatment data.

2. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S1, the multi-scale feature matching algorithm specifically includes: Extract bony structural features from CT images, soft tissue boundary features from MRI images, and metabolically active region features from PET-CT images. Non-rigid alignment of multiple sets of feature points is performed using a deformable registration network; An adaptive weighted fusion module is used to dynamically allocate fusion weights based on the signal-to-noise ratio and contrast of each modality image in different anatomical regions.

3. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S2, the 3DU-Net++ neural network is also connected to an uncertainty evaluation unit, which is used to output the segmentation confidence of each voxel; When the confidence level of a certain region in the segmentation uncertainty heatmap is lower than a preset threshold, a manual review prompt mechanism is triggered, and the region to be reviewed is highlighted on the interactive interface.

4. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S3, the mathematical expression of the multi-objective optimization model is: in, For tumor coverage, For radiation dose to dangerous organs, For path motion complexity, The weight coefficients are dynamically adjusted through reinforcement learning, and satisfy the following conditions: + + =1.

5. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S3, the parameter adaptive adjustment mechanism based on reinforcement learning specifically includes: Construct a Markov decision process model with treatment pathway parameters as actions and real-time treatment feedback and long-term efficacy evaluation as rewards; Using the deep deterministic policy gradient algorithm to adjust the weight coefficients Conduct online learning and optimization.

6. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S4, the method for constructing the hybrid prediction model includes: An LSTM network was trained using preprocessed temporal image data to learn respiratory movement cycle patterns. By using the output of the LSTM network as the observation input of the Kalman filter and combining it with the kinematic model of the actuator, the position of the target area within the next 0.5-1.0 seconds can be predicted.

7. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, Step S4 further includes a dynamic constraint step for the safety boundary: Based on the real-time distance between the vascular and neural structures in the segmentation mask and the target area, the safety protection boundary is dynamically calculated and updated; When the actuator trajectory is predicted or detected to potentially intrude into the safety protection boundary, the control module automatically triggers a deceleration or stop command and issues visual and auditory alarms on the human-machine interface.

8. The accurate positioning guidance method for thermotherapy nursing according to claim 1, characterized in that, In step S5, the feedback optimization process specifically includes: Establish treatment efficacy evaluation indicators, which comprehensively consider real-time temperature target achievement rate, positioning trajectory tracking error, and post-treatment imaging changes; Based on this evaluation metric, the constraints in the multi-objective optimization model and the reward function of the reinforcement learning model are fine-tuned in real time using an online learning algorithm.

9. A precise positioning guidance system for implementing the method according to any one of claims 1-8, characterized in that, It includes a multimodal image acquisition and fusion unit, an intelligent segmentation and analysis unit, a treatment pathway planning unit, a real-time guidance and control unit, a monitoring feedback and optimization unit, and a system bus and communication interface unit; The multimodal image acquisition and fusion unit is used to perform step S1; The intelligent segmentation and analysis unit has the improved 3DU-Net++ neural network built in, which is used to execute step S2; The treatment path planning unit has a built-in multi-objective optimization model and reinforcement learning engine, which is used to execute step S3; The real-time guidance and control unit includes an actuator, sensors, and a hybrid prediction model, and is used to execute step S4; The monitoring feedback and optimization unit is used to execute step S5; The system bus and communication interface unit is used to connect and coordinate the data and instruction transmission between the various units.

10. A precise positioning and guidance system for thermotherapy nursing according to claim 9, characterized in that, The real-time guidance and control unit also includes an augmented reality human-computer interaction unit for: The optimal treatment path plan, real-time segmentation results, predicted target area location, and safety boundary are overlaid and displayed in three dimensions on a real-scene video stream on the patient's body surface. It provides gesture and voice interaction interfaces, allowing operators to confirm and fine-tune key points during the guidance process.

Images