Multi-organ motion coupling compensation system based on spatiotemporal prediction model and execution method thereof

Abstract

The application belongs to the technical field of intelligent medical treatment assistance, and provides a multi-organ motion coupling compensation system based on a space-time prediction model and an execution method thereof, which comprises a multi-source heterogeneous data acquisition and preprocessing module, a double-dimension gold standard construction module, a space-time fusion prediction model construction module, a training and verification module, a time sequence trajectory prediction module and a motion compensation instruction generation module. Through time sequence synchronization and space registration of multi-modal data, the application uses a space-time Transformer architecture to construct a prediction model, accurately quantifies multi-organ motion coupling coefficients, and predicts an output motion trajectory that is physically reasonable and meets accuracy standards. Meanwhile, relying on double gold standards and an optimized training strategy, combined with a parameter efficient fine-tuning strategy, the application realizes personalized adaptation for specific patients, solves defects such as low multi-source data synchronization accuracy and insufficient coupling relationship modeling in existing radiotherapy motion compensation technologies, and promotes the clinical landing and upgrading of precise radiotherapy technologies.

Description

Technical Field

[0001] This application belongs to the field of intelligent medical treatment assistance technology, specifically, it relates to a multi-organ motion coupling compensation system based on a spatiotemporal prediction model and its execution method. Background Technology

[0002] Radiotherapy is one of the core treatment methods for thoracic and abdominal tumors. Its core goal is to deliver the treatment dose precisely to the tumor target area while protecting the surrounding normal tissues and organs to the greatest extent and reducing the risk of radiotherapy complications. Therefore, the accuracy of target area localization and motion compensation directly determines the treatment effect.

[0003] During radiotherapy, patients with thoracic and abdominal tumors are affected by physiological activities such as breathing and heartbeat. The tumor and adjacent organs, such as the diaphragm, liver, and gastrointestinal tract, will undergo dynamic displacement and deformation. Moreover, this movement is not isolated. Due to the anatomical structure and physiological driving force, there is a significant motion coupling relationship between organs, which further increases the difficulty of target tracking and compensation. If only the tumor is predicted to move, the prediction deviation is likely to be too large, causing dose distribution deviation, resulting in damage to normal organs or insufficient local dose to the tumor.

[0004] Existing tumor motion compensation techniques have many limitations. Traditional respiratory gating techniques rely on a single respiratory phase, sacrificing treatment efficiency and failing to adapt to irregular breathing. Real-time tracking techniques mostly rely on single-modal data, lack the ability to collaboratively fuse multi-source data, and do not fully consider the inherent spatial topological coupling relationship between organs. They also lack the quantification and physical constraints of coupling coefficients, which easily outputs physically unreasonable motion trajectories, thus hindering the clinical implementation of precision radiotherapy. Summary of the Invention

[0005] To address the shortcomings of existing radiotherapy motion compensation techniques, such as low synchronization accuracy of multi-source data, insufficient modeling of coupling relationships, and poor generalization, this application provides a multi-organ motion coupling compensation system and its execution method based on a spatiotemporal prediction model.

[0006] In one scheme, a multi-organ motion coupling compensation system based on a spatiotemporal prediction model includes a multi-source heterogeneous data acquisition and preprocessing module, a two-dimensional gold standard construction module, a spatiotemporal fusion prediction model construction module, a training and verification module, a time-series trajectory prediction module, and a motion compensation instruction generation module.

[0007] The multi-source heterogeneous data acquisition and preprocessing module includes a data acquisition unit and a data preprocessing unit. The data acquisition unit collects multi-organ dynamic image data and multi-dimensional motion monitoring data as sample data according to the required sample size. The data preprocessing unit performs feature extraction, normalization and coupling region segmentation on the sample data, and outputs a standardized multimodal fusion feature vector.

[0008] The dual-dimensional gold standard construction module includes a motion trajectory gold standard construction unit and a coupling characteristic gold standard construction unit. The motion trajectory gold standard construction unit is based on the three-dimensional coordinates of the organ centroid annotated by multi-organ dynamic image data, and combines the body surface time series data to correct the breakpoints. After calibration by electrocardiogram signal, a continuous trajectory sequence is generated. The coupling characteristic gold standard construction unit is based on the boundary annotation and geometric calculation of multi-organ dynamic image data to determine the reference interval of coupling coefficient, coupling direction and spatiotemporal significant interval.

[0009] The spatiotemporal fusion prediction model construction module constructs a spatiotemporal Transformer fusion prediction model including a spatial coupling branch, a temporal prediction branch, and an adaptive optimization unit. The spatial coupling branch uses an 8-head self-attention mechanism to calculate the attention weights between nodes and generate coupling coefficients. The temporal prediction branch focuses on key respiratory nodes based on the temporal attention layer to predict the three-dimensional motion trajectories of multiple organs. The adaptive optimization unit dynamically adjusts the organ feature weights based on the coupling coefficient threshold.

[0010] The training and validation module includes a model training unit and a validation and evaluation unit. The model training unit adopts a transfer learning strategy and uses the gold standard of motion trajectory as a label and the gold standard of coupling characteristics as a constraint. The model parameters are iteratively adjusted through a dual-objective loss function. The validation and evaluation unit evaluates the model prediction error based on the indicators of trajectory prediction accuracy and coupling coefficient accuracy.

[0011] The temporal trajectory prediction module receives the processed multimodal fusion feature input from the patient, calls the trained prediction model, calculates the real-time coupling coefficient through the spatial coupling branch, and outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a future preset time period through the temporal prediction branch.

[0012] The motion compensation command generation module converts the predicted three-dimensional motion trajectory into compensation commands containing displacement, motion direction, velocity, and acceleration parameters according to the control protocol of the linear accelerator action unit.

[0013] In one approach, the gold standard annotation process for motion trajectories includes: Doctors use medical image annotation software to delineate the centroid and contour of the tumor and coupled organs for each respiratory phase of the fused image. The system automatically extracts the coordinate values ​​of the centroid of each organ in the three-dimensional coordinate system and sorts them according to the respiratory phase to generate a sequence of three-dimensional motion trajectories of the organs. The trajectory sequence is smoothed to generate a gold standard dataset of motion trajectories containing the temporal characteristics of three-dimensional displacement, velocity, and acceleration of each organ.

[0014] In one approach, the gold standard annotation process for coupling characteristics includes: Based on the gold standard of motion trajectory, doctors mark the inter-organ coupling characteristics, including the significant coupling interval, the direction of coupled motion, and the coupling strength level. Based on the marked coupling characteristics, reference values ​​of coupling coefficients are generated through geometric calculations, which serve as the basis for the gold standard of coupling coefficients. Statistical analysis is performed on all marked reference values ​​of coupling coefficients to generate reference intervals of coupling coefficients for different tumor types.

[0015] Furthermore, the spatial coupling branch includes the GAT topology modeling unit, the multi-head self-attention calculation unit, and the coupling coefficient generation unit.

[0016] The GAT topology modeling unit uses tumors and associated coupled organs as nodes and Euclidean distances and anatomical adjacency relationships between organs as edge weights to construct an organ topology graph. The multi-head self-attention calculation unit uses an 8-head self-attention mechanism to calculate the attention weights between nodes. The coupling coefficient generation unit obtains the basic coupling coefficient based on displacement correlation, time synchronization, and spatial proximity weights, and then combines the multi-head attention weights to generate the final coupling coefficient. The coupling coefficient generation unit introduces a reference interval for the gold standard of coupling coefficients to constrain the final coupling coefficients.

[0017] Furthermore, the temporal prediction branch includes a temporal feature encoding unit and a trajectory decoding unit.

[0018] The temporal feature encoding unit expands the spatial coupling feature matrix along the time axis, focuses on key nodes at the end of inspiration and expiration to strengthen attention weights, and learns the temporal coordination rules of multi-organ movement; the trajectory decoding unit uses a single-layer decoder to output the three-dimensional coordinate sequence of the tumor and associated coupled organs with the same frequency as the body surface data in the future preset time period.

[0019] Furthermore, the basic coupling coefficient is calculated by weighted summation based on three geometric characteristics: displacement correlation, time synchronization, and spatial proximity.

[0020] Among them, displacement correlation is the Pearson correlation coefficient of the three-dimensional displacement sequence of the associated organ and the tumor, with a weight of 0.5; temporal synchronicity is the deviation rate of the occurrence time of the motion peak and motion trough of the associated organ and the tumor, with a weight of 0.3; spatial proximity is the normalized value of the extreme Euclidean distance between the associated organ and the tumor, with a weight of 0.2.

[0021] Furthermore, the coupling coefficient threshold adopts a two-layer setting logic combining a global reference threshold and a patient-specific threshold. The global reference threshold is based on the statistical distribution of the gold standard data for each coupling characteristic, and the median of the coupling coefficient for different tumor types is calculated as the initial judgment threshold. The patient-specific threshold is calculated by taking the basic coupling coefficient of each associated coupled organ of the patient, and using the elbow method, sorting each basic coupling coefficient according to the parameter value, and taking the elbow point where the rate of decrease of the coupling coefficient parameter value increases sharply as the personalized threshold.

[0022] In one approach, the time synchronization between multi-organ dynamic imaging data and multi-dimensional motion monitoring data adopts the master clock of the PTP precision clock protocol, using the treatment trigger signal of the linear accelerator as a reference to unify the system clock of each acquisition device.

[0023] The spatial dimension uses the treatment plan CT as the spatial reference. First, rigid registration is performed using bony landmarks, and then soft tissue deformation is corrected by elastic registration using Deformable UNet. After registration, the spatial coordinates of the dynamic imaging data of various organs and the motion monitoring data of various dimensions are unified into the DICOM coordinate system.

[0024] Furthermore, to achieve the above objectives, this application also provides a multi-organ motion coupling compensation execution method based on a spatiotemporal prediction model, used to implement the aforementioned multi-organ motion coupling compensation system based on a spatiotemporal prediction model. The method includes the following steps: S1. Collect multi-organ dynamic imaging data and multi-dimensional motion monitoring data of the target patient for at least one complete respiratory cycle, call the system's multi-source heterogeneous data acquisition and preprocessing module, perform spatiotemporal registration, feature extraction and fusion, obtain multimodal fusion features, and simultaneously extract the patient's personalized anatomical parameters and respiratory pattern features; S2. Input the multimodal fusion features obtained in S1 into the spatiotemporal fusion prediction model trained by the system, and use a parameter efficient fine-tuning strategy to perform personalized fine-tuning of the spatiotemporal fusion prediction model to adapt to the personalized organ movement patterns of the target patient. S3. The system receives the patient's multimodal fusion feature data in real time and calculates the real-time coupling coefficient between each organ and the tumor through the spatial coupling branch. When the real-time coupling coefficient exceeds the personalized threshold, the adaptive optimization unit dynamically adjusts the feature weight of the corresponding organ. The temporal prediction branch outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a future preset time period based on the coupling features and temporal patterns. S4. The system's motion compensation command generation module is based on the coupling coefficient priority, prioritizing the adaptation of organ motion patterns with higher parameter values ​​to correct command parameters. S5. The radiotherapy equipment receives instructions and drives the motion unit to adjust the direction and position of the radiation.

[0025] Furthermore, in step S2, the execution logic of the parameter efficient fine-tuning strategy is to fix the backbone parameters of the model and only fine-tune the adapter parameters of the model's attention layer and fully connected layer; the fine-tuning process uses the AdamW optimizer with a learning rate of 1.2 times the original learning rate of the model, and completes the adaptation in 1 to 2 iterations.

[0026] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application utilizes a spatiotemporal Transformer architecture to construct a predictive model by performing temporal synchronization and spatial registration of multimodal data. It accurately quantifies the motion coupling coefficients of multiple organs, predicting and outputting physically reasonable and accurate motion trajectories. Simultaneously, relying on dual-gold standards and optimized training strategies, combined with efficient parameter fine-tuning strategies, it achieves personalized adaptation for specific patients. This approach demonstrates strong generalization ability and good clinical compatibility, effectively improving target dose delivery accuracy, reducing the risk of damage to normal organs, and addressing the shortcomings of existing radiotherapy motion compensation techniques, such as low multi-source data synchronization accuracy and insufficient coupling relationship modeling. This promotes the clinical application and upgrading of precision radiotherapy technology. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the system modules of a multi-organ motion coupling compensation system based on a spatiotemporal prediction model in one embodiment of this application; Figure 2 This is a schematic diagram of the method flow of a multi-organ motion coupling compensation execution method based on a spatiotemporal prediction model in one embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0030] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0031] To address the shortcomings of existing radiotherapy motion compensation techniques, such as low synchronization accuracy of multi-source data, insufficient modeling of coupling relationships, and poor generalization, this application provides a multi-organ motion coupling compensation system based on a spatiotemporal prediction model. Please refer to [link to relevant documentation]. Figure 1 The system includes a multi-source heterogeneous data acquisition and preprocessing module, a two-dimensional gold standard construction module, a spatiotemporal fusion prediction model construction module, a training and validation module, a time-series trajectory prediction module, and a motion compensation instruction generation module.

[0032] In this embodiment, the multi-source heterogeneous data acquisition and preprocessing module includes a data acquisition unit and a data preprocessing unit. The data acquisition unit acquires multi-organ dynamic imaging data and multi-dimensional motion monitoring data as sample data according to the required sample size; the data preprocessing unit performs feature extraction, normalization, and coupled region segmentation on the sample data, and outputs a standardized multimodal fusion feature vector.

[0033] Specifically, multi-organ dynamic imaging data includes at least 4D-CT images, dynamic MRI images, and PET-CT images. 4D-CT images are acquired using a 16-slice or higher 4D-CT scanner, dividing the respiratory cycle into 10 to 12 phases, acquiring images covering 3 to 5 complete respiratory cycles, with a slice thickness not exceeding 1 mm and a pixel pitch of 0.5 × 0.5 mm. This provides a basis for the three-dimensional structure and motion trajectory of multiple organs, serving as a spatial registration reference. Dynamic MRI images are acquired using a 3.0T dynamic MRI scanner, acquiring fast gradient echo sequences at a sampling frequency of 0.5 fps, without contrast enhancement, with a sampling duration of at least 30 seconds, also covering 3 to 5 respiratory cycles. This captures details of soft tissue deformation and assists in identifying coupling region boundaries. PET-CT images are acquired using a PET-CT scanner, performing a single scan at the corresponding median respiratory phase, acquiring SUV values ​​normalized, with slice thickness consistent with 4D-CT, used to distinguish the metabolic boundaries between tumors and normal organs, avoiding misjudgment of coupling regions.

[0034] Multidimensional motion monitoring data includes at least surface optical tracking data, abdominal pressure sensor data, and electrocardiogram (ECG) signals, all acquired synchronously with multi-organ dynamic imaging data acquisition. Surface optical tracking data is acquired using an infrared surface optical tracking system with a sampling frequency of at least 10 fps, simultaneously marking three tracking points on the chest wall or abdomen to aid in constructing a correlation model between surface and internal organ motion, providing real-time temporal signals. Abdominal pressure sensor data is acquired using a flexible pressure sensor attached to the abdomen in the tumor projection area, with a sampling frequency of at least 10 fps, which helps characterize respiratory intensity and enhances the reliability of the surface-internal motion correlation. ECG signals are acquired using an ECG monitor, extracting QRS complexes at a sampling frequency of 1000 fps, serving as a temporal alignment benchmark to distinguish the coupled motion of respiration and the heart.

[0035] For example, the sample size requirement is at least 500 cases, which can be allocated according to tumor type: 200 cases of lung cancer, 200 cases of liver cancer, and 100 cases of pancreatic cancer, and divided into training set (70%), validation set (20%), and test set (10%).

[0036] Furthermore, the sample data preprocessing includes extracting the three-dimensional displacement, velocity, and acceleration time-series features of each organ, calculating the relative motion coefficient between organs; removing outliers caused by coughing and changes in body position, and using linear interpolation to fill in missing data in the acquisition gaps; normalizing continuous features such as motion amplitude and pressure values ​​to the 0-1 interval; and automatically segmenting the multi-organ coupling boundaries based on U-Net++ to form a standardized multimodal fusion feature vector.

[0037] Furthermore, the time synchronization between multi-organ dynamic imaging data and multi-dimensional motion monitoring data adopts the master clock of the PTP (Precision Time Protocol). Based on the treatment trigger signal of the linear accelerator, the acquisition of multi-modal data is synchronously started to avoid timing deviation caused by asynchronous acquisition. Among them, the PTP Precision Time Protocol is IEEE 1588v2, which is a common clock synchronization solution in the fields of medical equipment and industrial control. It is usually deployed in radiotherapy rooms. Its master clock is the supporting clock of the linear accelerator, and mainstream devices are equipped with it and can be directly connected to achieve microsecond-level synchronization of the system clocks of various imaging devices, motion monitoring devices, and electrocardiogram monitors. Exemplarily, the software layer can select the sampling rate of body surface optical data as the time reference, and use cubic spline interpolation to complete low-sampling-rate data such as dynamic MRI and 4D-CT to 10fps, retaining the non-linear characteristics of organ movement and avoiding motion trajectory distortion caused by interpolation; for high-sampling-rate data such as electrocardiogram and pressure sensors, downsampling and key feature extraction are adopted, that is, extracting the QRS complex of the electrocardiogram and the respiratory peak of the pressure sensor, and only retaining the key timing points related to organ movement to reduce data redundancy. For lost frames and noisy frames during the acquisition process, forward prediction and neighborhood mean correction are used to ensure the continuity of timing data. The timing deviation of all data after time synchronization is controlled within 10ms, and it can be verified by the correspondence between the QRS complex of the electrocardiogram and the respiratory phase of 4D-CT.

[0038] In the spatial dimension, the treatment plan CT is used as the spatial reference, and bony landmarks such as vertebrae and ribs, and vascular landmarks such as the aortic arch are used as anatomical marker points. Based on each anatomical marker point, rigid body transformations such as translation, rotation, and scaling are performed on MRI and PET-CT to achieve general spatial alignment, with an accuracy requirement of less than or equal to 0.8mm; then, a deep learning registration network of Deformable UNet is used to perform elastic deformation registration on the soft tissue areas of tumors and coupled organs to capture the spatial deviation caused by organ deformation, with an accuracy of less than or equal to 0.5mm; at the same time, based on the body surface optical-internal organ motion association model trained with data, the three-dimensional displacement of the body surface marker points can be mapped to the motion trend of the internal organs, and then the spatial alignment of the body surface data and the imaging data can be achieved. After registration, the spatial coordinates of each multi-organ dynamic imaging data and each multi-dimensional motion monitoring data are unified into the DICOM coordinate system to generate a multi-modal fusion three-dimensional spatio-temporal data volume as the input of the spatio-temporal prediction model.

[0039] In this embodiment, the gold standard is a crucial basis for verifying the accuracy of tumor motion prediction and the quantification accuracy of coupling coefficients. The dual-dimensional gold standard construction module includes a motion trajectory gold standard construction unit and a coupling characteristic gold standard construction unit. The motion trajectory gold standard construction unit is based on the three-dimensional coordinates of the organ centroid annotated by multi-organ dynamic imaging data, combined with body surface time-series data to correct breakpoints, and generates a continuous trajectory sequence after electrocardiogram signal calibration, which serves as the motion trajectory gold standard for verifying the model's prediction accuracy. The coupling characteristic gold standard construction unit, based on the boundary annotation and geometric calculation of multi-organ dynamic imaging data, determines the coupling coefficient reference interval, coupling direction, and spatiotemporal significant interval, forming the coupling characteristic gold standard for verifying the quantification accuracy of the coupling coefficient.

[0040] The coupling coefficient is essentially the motion correlation between organs based on anatomical and physiological driving relationships, specifically the organ-tumor coupling coefficient. It is defined as the contribution of the organ's three-dimensional motion to the tumor's three-dimensional displacement under physiological drives such as respiration and cardiac movement, and its spatiotemporal correlation. The value ranges from [0,1]. A larger value indicates a stronger effect of organ motion on tumor displacement; 0 represents no coupling; 1 represents complete coupling, meaning the tumor moves synchronously with the organ.

[0041] Specifically, the annotation process involves at least two professional radiation oncologists as primary annotators and one chief physician as arbitrator. The annotation results from the two physicians are validated for consistency using preset indicators. If the validation results fail to meet the preset indicators, the chief physician in charge of arbitration will re-annotate, and the arbitration result will be used as the gold standard. The annotation tool can utilize 3D Slicer to develop custom multi-organ motion annotation plugins, satisfying the requirements of 3D contour drawing of multiple organs, organ position marking at each phase of the respiratory cycle, quantitative calculation of relative displacement between organs, and automatic export of annotation results. The annotation data is based on fused images of 4D-CT and dynamic MRI, covering the complete respiratory cycle, including key motion nodes such as end-inspiratory and end-expiratory phases. For example, for lung cancer patients, the core organ group for annotation includes the tumor, diaphragm, heart, and ipsilateral lung lobe; for liver cancer or pancreatic cancer patients, the core organ group for annotation includes the tumor, diaphragm, stomach, liver, or adjacent intestinal segments of the pancreas.

[0042] Furthermore, the gold standard annotation process for motion trajectories is as follows: Doctors use medical image annotation software such as 3D Slicer to delineate the centroid and contour of the tumor and coupled organs for each respiratory phase of the fused image; then the system automatically extracts the coordinate values ​​of the centroid of each organ in the three-dimensional coordinate system and sorts them according to the respiratory phase to generate a sequence of three-dimensional motion trajectories of the organs; finally, the trajectory sequence is smoothed to remove abnormal points caused by coughing, changes in body position, etc., to generate a gold standard dataset of motion trajectories containing the temporal features of three-dimensional displacement, velocity, and acceleration of each organ.

[0043] Furthermore, the annotation process for the gold standard of coupling characteristics includes physicians marking inter-organ coupling features, such as significant coupling intervals, coupling direction, and coupling strength level, based on the gold standard of motion trajectory. For example, a significant coupling interval might be the interval where liver displacement occurs due to diaphragmatic elevation at the end of inspiration; the coupling direction can be distinguished by positive or negative coupling, such as positive coupling (diaphragmatic elevation causing upward liver displacement) and negative coupling (gastrointestinal peristalsis causing leftward tumor displacement); the coupling strength level can be graded according to subsequent quantitative reference values. Based on the marked coupling features, and combined with the contribution of associated organ motion to tumor displacement, geometric calculations are performed to generate coupling coefficient reference values, serving as the basis for the coupling coefficient gold standard. Finally, statistical analysis is performed on all marked coupling coefficient reference values ​​to generate coupling coefficient reference intervals for different tumor types.

[0044] Furthermore, during the model training phase, the gold standard for motion trajectory is used as the label for the model output, and the gold standard for coupling characteristics is used as the constraint condition for quantifying the coupling coefficient, thus avoiding unconstrained fitting of the model. During the model validation and testing phase, the tumor motion trajectory predicted by the model is compared with the gold standard, and indicators such as root mean square error and mean absolute error are calculated to verify the prediction accuracy. The coupling coefficients output by the model are compared with the gold standard for coupling characteristics, and cosine similarity is calculated to verify the quantization accuracy.

[0045] In this embodiment, the spatiotemporal fusion prediction model construction module constructs a spatiotemporal Transformer fusion prediction model including a spatial coupling branch, a temporal prediction branch, and an adaptive optimization unit. The model's input layer transforms preprocessed multimodal sample data into model-recognizable feature vectors, which are then fused into three categories: image features, motion features, and auxiliary features. Specifically, image features can be encoded using ResNet50 to capture organ contours, deformations, and coupling boundary features from 4D-CT / MRI / PET-CT, outputting a 256-dimensional feature vector; motion features can be encoded using 1D CNN to capture temporal motion features from surface optics and pressure sensors, outputting a 128-dimensional feature vector; and auxiliary features, such as periodic features extracted from ECG signals and patient anatomical parameters, output a 64-dimensional feature vector. These three types of features are concatenated into a 448-dimensional fusion feature vector, which is then input into the dual-branch structure of spatial coupling and temporal prediction.

[0046] Furthermore, the spatial coupling branch includes a GAT topology modeling unit, a multi-head self-attention calculation unit, and a coupling coefficient generation unit. The GAT topology modeling unit uses the tumor and related coupled organs such as the diaphragm and gastrointestinal tract as nodes, and the Euclidean distance and anatomical adjacency between organs as edge weights, replacing the fully connected spatial graph of a conventional Transformer to construct an organ topology graph. For example, the liver and stomach are close and anatomically adjacent, resulting in high edge weights; the liver and distal intestine are far apart, resulting in low edge weights. Through GAT's node-level attention mechanism, the coupling weights of neighboring organs are explicitly strengthened, outputting a topology-aware spatial feature matrix, avoiding the problem of indiscriminate weighting of global attention. The multi-head self-attention calculation unit uses an 8-head self-attention mechanism to calculate the attention weights between nodes. In the spatiotemporal Transformer, 8-head self-attention can be extended to spatiotemporal multi-head, with some heads focusing on the spatial dimension of lesion region association and others focusing on the temporal dimension of tumor movement such as respiration and heartbeat. By fusing multi-view and multi-node association information, a 512-dimensional multi-organ coupling feature matrix is ​​generated and output, quantifying the image weight of each organ for tumor movement. The coupling coefficient generation unit obtains the basic coupling coefficient based on displacement correlation, time synchronization and spatial proximity weights, and then generates the final coupling coefficient by combining multi-head attention weights. Among them, the coupling coefficient generation unit introduces the reference interval of the coupling coefficient gold standard to constrain the final coupling coefficient.

[0047] Specifically, the coupling coefficient adopts a quantization logic of "physics first, learning later." First, a basic coupling coefficient is calculated using geometric features, giving it biophysical meaning and constraining the model's fitting direction. Then, the final coupling coefficient is learned and corrected using the self-attention mechanism of the spatiotemporal Transformer, and used to capture complex nonlinear coupling relationships. The basic coupling coefficient is calculated based on three geometric features: displacement correlation, temporal synchronicity, and spatial proximity, through weighted summation. Specifically, displacement correlation r1 is the Pearson correlation coefficient of the three-dimensional displacement sequences of the associated organ and tumor, with a weight of 0.5, reflecting the consistency of the movement trends of the organ and tumor; temporal synchronicity r2 is the deviation rate of the occurrence time of the movement peaks and troughs of the associated organ and tumor, with a weight of 0.3, reflecting the temporal coordination of the movement of the organ and tumor; spatial proximity r3 is the normalized value of the limiting Euclidean distance between the associated organ and tumor, with a weight of 0.2, reflecting the anatomical association between the organ and tumor. The formula for calculating the basic coupling coefficient is as follows: C0=0.5×r1+0.3×r2+0.2×r3 The value ∈ [0,1]. Furthermore, using the basic coupling coefficient as a priori constraint, the nonlinear coupling relationship between organs is learned through a multi-head self-attention mechanism to correct the basic coupling coefficient. Therefore, the final coupling coefficient is calculated using the following formula: C X→T =C0×Attention(X,T) in, Attention(X,T) It is the organ of spacetime Transformer computation. X With tumors T The self-attention weights, with values ​​∈ [0.8, 1.2], only slightly modify the basic coupling coefficient. Specifically, the final coupling coefficient is designed to be directionally specific for clinical radiotherapy, and is thus a three-dimensional directional coupling coefficient. C X→T-x , C X→T-y , C X→T-z Quantify organs separately X Tumor in the x / y / z three-dimensional direction T The coupling contribution.

[0048] Furthermore, the temporal prediction branch includes a temporal feature encoding unit and a trajectory decoding unit. Based on the coupling feature matrix output by the spatial coupling branch, it captures the temporal patterns of multi-organ movement and predicts the movement trajectory for a future preset time period. The temporal feature encoding unit expands the spatial coupling feature matrix along the time axis, with 10 to 12 phases per respiratory cycle, focusing on key nodes such as the end of inspiration and end of expiration, strengthening attention weights, and learning the temporal coordination patterns of multi-organ movement. The trajectory decoding unit uses a single-layer decoder. Based on the encoded spatiotemporal features, it outputs the three-dimensional coordinate sequence of the tumor and associated coupled organs with frequencies consistent with the body surface data within the future preset time period. Simultaneously, it aligns the predicted trajectory with the gold standard of the movement trajectory, providing a foundation for subsequent loss calculation.

[0049] Furthermore, the adaptive optimization unit determines a global reference threshold based on the coupling coefficients output by the spatial coupling branch and the gold standard for coupling characteristics, dynamically adjusting the feature weights of each organ. The adjustment logic is as follows: when the coupling coefficient between an organ and the tumor is greater than the global reference threshold, it is identified as a major coupled organ, and its feature weight is automatically increased. After each frame of prediction, the threshold and weight coefficients are dynamically fine-tuned by comparing the prediction error with the gold standard for coupling characteristics, strengthening the contribution of major coupled organs to the prediction results and suppressing interference from non-major organs. The global reference threshold is determined based on the statistical distribution of the gold standard data for each coupling characteristic, using the median coupling coefficient of different tumor types as the initial judgment threshold.

[0050] Furthermore, the coupling coefficient threshold employs a two-tiered setting logic combining a global reference threshold and a patient-specific threshold. The patient-specific threshold is calculated by assigning the baseline coupling coefficients of each associated coupled organ to an elbow point, sorting the baseline coupling coefficients by their parameter values, and selecting the elbow point where the rate of decrease in the coupling coefficient parameter value suddenly increases as the personalized threshold. For example, the baseline coupling coefficients of a liver cancer patient are sorted as follows: diaphragm 0.85, stomach 0.62, intestines 0.31, kidneys 0.15. The elbow point is the intestines, where the rate of decrease in parameter value suddenly increases (0.31). Therefore, the patient's personalized threshold is 0.31, and only the diaphragm and stomach are included as the main coupled organs. During treatment, if the model detects changes in organ movement patterns, such as gastrointestinal distension leading to an increase in the coupling coefficient between the stomach and the tumor, the personalized threshold can be dynamically adjusted via online feedback to ensure accurate identification of the main coupled organs.

[0051] In this embodiment, the training and validation module includes a model training unit and a validation and evaluation unit. The model training unit employs a transfer learning strategy, using the gold standard for motion trajectory as a label and the gold standard for coupling characteristics as a constraint, and iteratively adjusts the model parameters through a bi-objective loss function. The validation and evaluation unit evaluates the model's prediction error based on metrics such as trajectory prediction accuracy and coupling coefficient accuracy.

[0052] Specifically, the model training unit divides the preprocessed sample data into training, validation, and test sets in a 7:2:1 ratio. Augmentation operations are performed on the training set samples, including random phase shifting, motion amplitude scaling, coupling coefficient perturbation, and random masking of single-modal features. After augmentation, the training set data size can be doubled. Then, a transfer learning strategy is employed, using public multi-organ motion datasets such as the LIDC-IDRI extended dataset to pre-train the multi-head self-attention layer of the spatial coupling branch and initialize parameters. The temporal prediction branch and adaptive optimization unit can use Xavier initialization to ensure reasonable parameter distribution and accelerate training convergence. For example, hyperparameters can be preset to a batch size of 8, an initial learning rate of 0.0001, 100 training epochs, a weight decay of 0.00001, 8 multi-head self-attention heads, and a prediction window of 3 seconds. The trajectory prediction loss uses mean squared error to calculate the 3D coordinate deviation between the model's predicted trajectory and the gold standard motion trajectory; the coupling coefficient loss uses cosine similarity loss to calculate the deviation between the model's output coupling coefficient and the gold standard coupling characteristics.

[0053] During iterative training, training set samples are input into the model. The coupling coefficient and spatial feature matrix are calculated through the spatial coupling branch, the predicted trajectory is output through the temporal prediction branch, and the adaptive optimization unit adjusts the feature weights, completing one forward propagation. Then, based on the bi-objective loss function, the loss value between the model output and the gold standard is calculated, and the gradient of each layer's parameters is obtained through backpropagation. Parameter updates can use the AdamW optimizer, updating model parameters in conjunction with gradient values, while applying weight decay to suppress overfitting. For the multi-head self-attention calculation unit and the coupling coefficient generation unit, a larger base learning rate, such as 1.2 times, is used for priority optimization. The learning rate adjustment uses a cosine annealing learning rate strategy, decaying to 0.5 of the original value every 20 rounds. When the validation set loss does not decrease for 10 consecutive rounds, an early stopping mechanism is triggered to avoid ineffective training.

[0054] The model fine-tuning logic can evaluate model performance based on the validation set. If the trajectory prediction error is too large, the temporal attention weights of the temporal prediction branch are fine-tuned; if the coupling coefficient deviation is large, the number of multi-head self-attention heads and the weights of the basic coupling coefficient in the spatial coupling branch are adjusted; if the accuracy of a certain type of tumor sample is too low, the training weights of that type of sample are increased. The conditions for judging model training convergence include: the validation set loss fluctuation is less than or equal to 0.00001 for 10 consecutive rounds; the validation set tumor prediction error is less than or equal to 1.2 mm, and the coupling coefficient calculation error is less than or equal to 0.05; the difference between the training set loss and the validation set loss is less than or equal to 0.01, with no obvious overfitting. When all the judgment conditions are met, the model training is judged to have converged.

[0055] Furthermore, the trajectory prediction accuracy evaluation indicators for the validation and evaluation unit include root mean square error (RMSE) and mean absolute error (MAE). RMSE is the square root mean of the deviations between the model's predicted trajectory and the gold standard motion trajectory in three-dimensional coordinates, with a standard value set at 1.2 mm. This is used to evaluate the accuracy of tumor motion prediction and directly affects the compensation effect. MAE is the mean absolute deviation of the model's predicted trajectory from the gold standard motion trajectory in three-dimensional coordinates, with a standard value set at 0.8 mm. This helps to assess the stability of the prediction error and avoid extreme deviations. The evaluation indicators for the accuracy of the coupling coefficient include cosine similarity and coupling coefficient deviation. Cosine similarity is the cosine similarity between the model's output coupling coefficient and the gold standard coupling characteristics, with a standard value set at 0.95. This assesses the rationality of the coupling coefficient calculation to ensure it aligns with clinical patterns. Coupling coefficient deviation is the mean absolute deviation of the model's output coupling coefficient from the gold standard coupling characteristics, with a standard value set at 0.05. This quantifies the calculation error of the coupling coefficient.

[0056] In this embodiment, the temporal trajectory prediction module receives the processed multimodal fusion feature input of the patient, calls the trained prediction model, calculates the real-time coupling coefficient through the spatial coupling branch, and outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a future preset time period through the temporal prediction branch.

[0057] In this embodiment, the motion compensation command generation module converts the predicted three-dimensional motion trajectory into compensation commands containing parameters such as displacement, direction of motion, velocity, and acceleration, according to the control protocol of the linear accelerator motion unit, such as a robotic arm. Simultaneously, it ensures that the commands conform to the physical constraints of the equipment. The compensation commands can be sent in real-time to the linear accelerator's motion control system, such as the robotic arm control system, via the equipment's local area network. The output delay is required to be no more than 2ms to ensure that the motion unit can synchronously respond to tumor movement and complete accurate compensation.

[0058] Furthermore, after each frame of compensation command is output, the system collects the actual compensation action data of the robotic arm and other motion units in real time and the CBCT images during treatment. It compares the actual tumor location with the model's predicted trajectory, calculates the error, dynamically fine-tunes the coupling coefficient threshold and feature fusion weight, and optimizes the generation of the next round of prediction and compensation commands, forming a real-time closed loop of "prediction-compensation-feedback-optimization" to continuously improve compensation accuracy.

[0059] As one implementation, this application also provides a multi-organ motion coupling compensation execution method based on a spatiotemporal prediction model, used to implement a multi-organ motion coupling compensation system based on a spatiotemporal prediction model in the aforementioned embodiments. Please refer to [link to relevant documentation]. Figure 2 The method includes the following steps: Step S1: Collection and processing of small sample data from target patients: The system collects dynamic imaging data of multiple organs and multidimensional motion monitoring data of the target patient for at least one complete respiratory cycle. It calls the system's multi-source heterogeneous data acquisition and preprocessing module to perform spatiotemporal registration, feature extraction and fusion to obtain multimodal fusion features. At the same time, it extracts the patient's personalized anatomical parameters and respiratory pattern features.

[0060] Step S2, Model Personalization and Fine-tuning: The multimodal fusion features obtained in step S1 are input into the spatiotemporal fusion prediction model trained by the system. The spatiotemporal fusion prediction model is then individually fine-tuned using an efficient parameter fine-tuning strategy to adapt to the personalized organ movement patterns of the target patient.

[0061] Step S3: Real-time Coupled Prediction and Trajectory Generation: The system receives multimodal fusion feature data from patients in real time and calculates the real-time coupling coefficient between each organ and the tumor through the spatial coupling branch. When the real-time coupling coefficient exceeds the personalized threshold, the adaptive optimization unit dynamically adjusts the feature weights of the corresponding organs. The temporal prediction branch outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a preset time period based on coupling features and temporal patterns.

[0062] Step S4: Optimization and Conversion of Compensation Instructions: The system's motion compensation command generation module prioritizes coupling coefficients, adapting to the motion patterns of organs with higher coupling coefficients, correcting command parameters, and ensuring accurate matching between commands and tumor motion trajectories, while also adapting to the physical motion constraints of the device's action units.

[0063] Step S5: Dynamic compensation execution: The radiotherapy equipment receives instructions and drives the motion unit to adjust the direction and position of the radiation rays, thereby achieving dynamic compensation for tumor movement.

[0064] For example, during the pre-treatment localization phase for the target patient, no more than 5 frames of multimodal fusion data combining 4D-CT and body surface optics are collected, covering a complete respiratory cycle. This means that only routine data from the localization phase is used for small-sample acquisition. The small-sample data from the target patient is input into a pre-trained spatiotemporal prediction model. An efficient parameter fine-tuning strategy is employed, adjusting only the adapter parameters of the model's attention and fully connected layers while fixing the core parameters. This allows for rapid adaptation to the target patient's personalized features, such as anatomical structure, breathing pattern, and coupling strength. The model, adapted through small-sample learning, is applied to the target patient's first two treatments. Actual organ motion data during treatment is collected using CBCT images and compared with the model's predicted trajectory. Errors are calculated, and the adapter parameters are rapidly fine-tuned online using gradient descent. The fine-tuning process employs the AdamW optimizer with a learning rate 1.2 times the original learning rate. Adaptation is completed in 1 to 2 iterations. Subsequent treatments utilize a personalized model for real-time prediction and compensation, eliminating the need for further adjustments.

[0065] In summary, this application employs a millisecond-level spatiotemporal registration strategy combining PTP clock hard synchronization and Deformable UNet elastic registration. This strategy achieves a multimodal data temporal deviation of ≤10ms and a spatial registration accuracy of ≤0.5mm, effectively solving the synchronization and adaptation challenges of multimodal data, meeting the millisecond-level response requirements of radiotherapy, and providing a high-quality data foundation for subsequent coupled modeling and prediction. Furthermore, the standardized data preprocessing process ensures the consistency and usability of the sample data. The spatiotemporal prediction model adopts a fusion architecture of spatiotemporal Transformer and GAT. GAT explicitly constructs an organ topology map, strengthening the capture of spatial proximity coupling relationships. Combined with a multi-head self-attention mechanism, the coupling coefficient is quantified, avoiding the shortcomings of global attention ignoring inherent topological relationships and preventing physically unreasonable motion trajectory outputs. The modeling results possess both mathematical rationality and clinical interpretability. Simultaneously, the dual-dimensional gold standard provides accurate labels and constraints for the model, the dual-objective loss function effectively suppresses overfitting, transfer learning pre-training and adaptive fine-tuning strategies shorten the training cycle, and the combined efficient parameter fine-tuning strategy can quickly adapt to the personalized characteristics of new patients, comprehensively enhancing the model's training and generalization capabilities. The system has a clear collaborative architecture among its modules and is compatible with a variety of commonly used clinical equipment such as 4D-CT, dynamic MRI, and surface optics. It does not require modification of existing radiotherapy hardware and continuously improves compensation accuracy through personalized threshold correction and closed-loop optimization mechanisms. It is suitable for various thoracic and abdominal tumor radiotherapy scenarios, such as lung cancer, liver cancer, and pancreatic cancer.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-organ motion coupling compensation system based on a spatiotemporal prediction model, characterized in that, It includes a multi-source heterogeneous data acquisition and preprocessing module, a two-dimensional gold standard construction module, a spatiotemporal fusion prediction model construction module, a training and validation module, a time-series trajectory prediction module, and a motion compensation instruction generation module; The multi-source heterogeneous data acquisition and preprocessing module includes a data acquisition unit and a data preprocessing unit; the data acquisition unit acquires multi-organ dynamic image data and multi-dimensional motion monitoring data as sample data according to the sample requirements; the data preprocessing unit performs feature extraction, normalization and coupling region segmentation on the sample data, and outputs a standardized multimodal fusion feature vector; The dual-dimensional gold standard construction module includes a motion trajectory gold standard construction unit and a coupling characteristic gold standard construction unit. The motion trajectory gold standard construction unit is based on the three-dimensional coordinates of the organ centroid annotated by the multi-organ dynamic image data, and combines the body surface time series data to correct breakpoints. After calibration by electrocardiogram signal, a continuous trajectory sequence is generated. The coupling characteristic gold standard construction unit determines the coupling coefficient reference interval, coupling direction, and spatiotemporal significant interval based on the boundary annotation and geometric calculation of the multi-organ dynamic image data. The spatiotemporal fusion prediction model construction module constructs a spatiotemporal Transformer fusion prediction model including a spatial coupling branch, a temporal prediction branch, and an adaptive optimization unit. The spatial coupling branch uses an 8-head self-attention mechanism to calculate the attention weights between nodes and generate coupling coefficients; the temporal prediction branch focuses on key respiratory nodes based on the temporal attention layer to predict the three-dimensional motion trajectories of multiple organs; the adaptive optimization unit dynamically adjusts the organ feature weights based on the coupling coefficient threshold. The training and validation module includes a model training unit and a validation and evaluation unit. The model training unit adopts a transfer learning strategy and uses the motion trajectory gold standard as a label and the coupling characteristic gold standard as a constraint to iteratively adjust the model parameters through a dual-objective loss function. The validation and evaluation unit evaluates the model prediction error based on the trajectory prediction accuracy and the coupling coefficient accuracy. The temporal trajectory prediction module receives the processed multimodal fusion feature input of the patient, calls the trained prediction model, calculates the real-time coupling coefficient through the spatial coupling branch, and outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a future preset time period through the temporal prediction branch. The motion compensation command generation module converts the predicted three-dimensional motion trajectory into compensation commands containing displacement, motion direction, velocity, and acceleration parameters according to the control protocol of the linear accelerator action unit.

2. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 1, characterized in that, The gold standard annotation process for motion trajectories includes: Doctors use medical image annotation software to delineate the centroid and contour of the tumor and coupled organs for each respiratory phase of the fused image. The system automatically extracts the coordinate values ​​of the centroid of each organ in the three-dimensional coordinate system and sorts them according to the respiratory phase to generate a sequence of three-dimensional motion trajectories of the organs. The trajectory sequence is smoothed to generate a gold standard dataset of motion trajectories containing the temporal characteristics of three-dimensional displacement, velocity, and acceleration of each organ.

3. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 2, characterized in that, The gold standard labeling process for coupling characteristics includes: Based on the gold standard of motion trajectory, the doctor marks the inter-organ coupling characteristics, including the significant coupling interval, the direction of coupling motion, and the coupling strength level. Based on the marked coupling characteristics, a coupling coefficient reference value is generated through geometric calculation, which serves as the basis for the coupling coefficient gold standard. Statistical analysis is performed on all marked coupling coefficient reference values ​​to generate coupling coefficient reference intervals for different tumor types.

4. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 3, characterized in that, The spatial coupling branch includes a GAT topology modeling unit, a multi-head self-attention calculation unit, and a coupling coefficient generation unit; The GAT topology modeling unit uses tumors and associated coupled organs as nodes, and Euclidean distances and anatomical adjacency relationships between organs as edge weights to construct an organ topology graph; the multi-head self-attention calculation unit uses an 8-head self-attention mechanism to calculate the attention weights between the nodes. The coupling coefficient generation unit obtains the basic coupling coefficient based on displacement correlation, time synchronization and spatial proximity weights, and then combines the multi-head attention weights to generate the final coupling coefficient. The coupling coefficient generation unit introduces a reference range for the gold standard of the coupling coefficient to constrain the final coupling coefficient.

5. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 4, characterized in that, The temporal prediction branch includes a temporal feature encoding unit and a trajectory decoding unit; The temporal feature encoding unit expands the spatial coupling feature matrix along the time axis, focuses on key nodes at the end of inspiration and expiration to strengthen attention weights, and learns the temporal coordination rules of multi-organ movement; the trajectory decoding unit adopts a single-layer decoder to output the three-dimensional coordinate sequence of tumors and associated coupled organs with the same frequency as the body surface data within a future preset time period.

6. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 4, characterized in that, The basic coupling coefficient is calculated by weighted summation based on three geometric features: displacement correlation, time synchronization, and spatial proximity. Wherein, the displacement correlation is the Pearson correlation coefficient of the three-dimensional displacement sequence of the associated organ and the tumor, with a weight of 0.5; the temporal synchronization is the deviation rate of the occurrence time of the motion peak and motion trough of the associated organ and the tumor, with a weight of 0.3; and the spatial proximity is the normalized value of the limiting Euclidean distance between the associated organ and the tumor, with a weight of 0.

2.

7. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to claim 6, characterized in that, The coupling coefficient threshold adopts a two-layer setting logic that combines a global reference threshold and a patient-specific threshold; The global reference threshold is based on the statistical distribution of the gold standard data of each coupling characteristic, and the median of the coupling coefficient of different tumor types is calculated as the initial judgment threshold. The patient-specific threshold is calculated by taking the basic coupling coefficients of each associated coupled organ of the patient, using the elbow point method to sort the basic coupling coefficients according to the parameter value, and taking the elbow point where the rate of decrease of the coupling coefficient parameter value increases sharply as the personalized threshold.

8. The multi-organ motion coupling compensation system based on a spatiotemporal prediction model according to any one of claims 1 to 7, characterized in that, The time synchronization between the multi-organ dynamic imaging data and the multi-dimensional motion monitoring data adopts the master clock of the PTP precision clock protocol, and uses the treatment trigger signal of the linear accelerator as the reference to unify the system clock of each acquisition device. The spatial dimension uses the treatment plan CT as the spatial reference. First, rigid registration is performed using bony landmarks, and then soft tissue deformation is corrected by elastic registration using DeformableUNet. After registration, the spatial coordinates of the dynamic imaging data of each organ and the multidimensional motion monitoring data are unified to the DICOM coordinate system.

9. A multi-organ motion coupling compensation execution method based on a spatiotemporal prediction model, characterized in that, For implementing the multi-organ motion coupling compensation system based on a spatiotemporal prediction model as described in claim 8, the method includes the following steps: S1. Collect multi-organ dynamic imaging data and multi-dimensional motion monitoring data of the target patient for at least one complete respiratory cycle, call the multi-source heterogeneous data acquisition and preprocessing module of the system, perform spatiotemporal registration, feature extraction and fusion to obtain multimodal fusion features, and simultaneously extract the patient's personalized anatomical parameters and respiratory pattern features; S2. Input the multimodal fusion features obtained in S1 into the spatiotemporal fusion prediction model trained by the system, and use a parameter efficient fine-tuning strategy to perform personalized fine-tuning of the spatiotemporal fusion prediction model to adapt to the personalized organ movement patterns of the target patient. S3. The system receives the patient's multimodal fusion feature data in real time and calculates the real-time coupling coefficient between each organ and the tumor through spatial coupling branches. When the real-time coupling coefficient exceeds the personalized threshold, the adaptive optimization unit dynamically adjusts the feature weights of the corresponding organs. The temporal prediction branch, based on coupling characteristics and temporal patterns, outputs the three-dimensional motion trajectory of the tumor and associated coupled organs within a future preset time period. S4. The system's motion compensation command generation module is based on the coupling coefficient priority, prioritizing the adaptation of organ motion patterns with higher parameter values ​​to correct command parameters. S5. The radiotherapy equipment receives instructions and drives the motion unit to adjust the direction and position of the radiation.

10. The multi-organ motion coupling compensation execution method based on a spatiotemporal prediction model according to claim 9, characterized in that, In step S2, the execution logic of the parameter efficient fine-tuning strategy is to fix the backbone parameters of the model and only fine-tune the adapter parameters of the model's attention layer and fully connected layer. The fine-tuning process uses the AdamW optimizer with a learning rate of 1.2 times the original learning rate of the model, and completes the adaptation in 1 to 2 iterations.

Images