Personalized tumor motion evolution prediction method, system, medium, and program product

Abstract

This invention belongs to the field of medical image processing and radiotherapy, and relates to a personalized method, system, medium, and program product for predicting tumor motion evolution. The method includes multi-source data acquisition and anatomical feature extraction, construction of a basic model family incorporating the Cardiac-induced Motion Coupling Factor (CMCF), individualized transfer learning, and specialized model generation. This invention aims to construct a basic model family through anatomical partitioning and introduce the CMCF as a spatial constraint. By combining transfer learning with an evolutionary strategy based on clustering experience playback, it achieves high-precision, self-optimizing personalized tumor motion prediction, solving the model failure problem under complex respiratory patterns.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and radiotherapy, and relates to a personalized method and system for predicting tumor motion evolution. More specifically, it relates to a method for predicting the real-time three-dimensional motion trajectory of lung tumors using body surface point clouds, anatomical prior knowledge, and transfer learning algorithms. Background Technology

[0002] In radiotherapy for lung tumors, tumor displacement caused by both respiratory and cardiac movements is a key factor affecting irradiation accuracy. Current tumor tracking methods have the following problems: (1) Lack of anatomical specificity: Most existing models are general mappings and do not consider the significant differences in biomechanics and respiratory deformation patterns of different lung lobes (upper left, lower left, upper right, middle right, and lower right).

[0003] (2) Ignoring cardiac motion: Tumors near the heart or aortic wall will generate high-frequency micro-motions with the heartbeat. Traditional respiratory prediction models cannot effectively decouple the combined motion of breathing and heartbeat.

[0004] (3) Model staticization and cold start: Once the model is trained, it is difficult to absorb new patient data for model optimization, and the speed of personalized adaptation to new patients is slow, making it difficult to balance group generalization and individual differences.

[0005] Therefore, there is an urgent need for new tumor motion prediction methods that take into account cardiogenic motion, reveal specificity, and provide personalized adaptation for patients. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a personalized tumor motion evolution prediction method, system, medium and program product, the purpose of which is to take into account cardiogenic motion, reveal specificity and personalized adaptation to patients.

[0007] To achieve the above objectives, according to one aspect of the present invention, a personalized tumor motion evolution prediction method is provided, comprising: Multi-source data acquisition and anatomical feature extraction: Respiratory features extracted from body surface point clouds, lung lobe information extracted from 4DCT, and tumor target region locations are transformed to the same spatial reference. The body surface point cloud is divided into multiple sub-blocks, and the motion correlation between each sub-block and the five lung lobe regions is calculated. Different anatomical weights are assigned to each sub-block, resulting in a feature matrix containing information from the multiple sub-blocks with the highest motion significance and their respective lung lobe labels. ; Introducing the basic model family for CMCF: Defining the cardiogenic motion coupling factor for tumors located near the heart. Based on the location of the tumor in the lung lobe, the historical database was divided into five subsets corresponding to the five lung lobes, and feature matrices were used simultaneously. and cardiogenic motion coupling factor Train the base models corresponding to the 5 subsets; Personalized transfer learning and dedicated model generation: Based on the lung lobe to which the new patient's tumor belongs, the corresponding weights of the basic model are called; during the treatment plan preparation stage, the patient's synchronous data is obtained, the shallow feature layer of the basic model is frozen, and only the parameters of the deep mapping layer are updated to generate a personalized dedicated model for the patient.

[0008] Furthermore, it also includes anomaly detection and triggering logic in real-time prediction: real-time monitoring of the input sequence of body surface point cloud and calculation of its reconstruction residual R; if R < threshold T, maintain the prediction of the dedicated model; if R ≥ threshold T, it is judged as an anomaly, and the model is switched to the basic model to perform downgraded prediction until the respiratory signal returns to stability.

[0009] Furthermore, reconstruct the residuals ,in, It is a predicted feature based on historical stationary series.

[0010] Furthermore, it also includes the evolution of basic models based on clustering experience playback: after the data of new patients are validated, clustering is performed according to respiratory frequency, amplitude and phase characteristics, and representative respiratory samples are selected and stored in the experience pool; old samples in the experience pool are extracted and mixed with the data of newly enrolled patients to incrementally train the five basic models.

[0011] Furthermore, the feature vector In the form of: , in, It is a small piece on the body surface With lung lobes The correlation between them It is the first The mean z-axis coordinates of each sub-block at time t, i = 1~K, j = 1~5, where K is the number of sub-blocks with the highest motion significance selected.

[0012] Furthermore, the correlation is the Pearson correlation coefficient, in which case, The calculation formula is as follows: , in, =1~10 represents the sequence number of the 10 phases into which respiration is divided. Indicates small patches on the body surface In the Mean Z-axis coordinates of respiratory phase. Indicates small patches on the body surface Z-axis mean across all phases Indicates lung lobe In the Z-axis centroid coordinates of the respiratory phase Indicates lung lobe Mean Z-axis centroid coordinates across all breathing phases.

[0013] Furthermore, cardiogenic motion coupling factor ,in, This represents the minimum Euclidean distance from the tumor center to the heart. The attenuation coefficient is... It is the natural logarithm.

[0014] According to another aspect of the present invention, a personalized tumor motion evolution prediction system is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement a personalized tumor motion evolution prediction method as described in any of the preceding claims.

[0015] According to another aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements a personalized tumor motion evolution prediction method as described in any of the preceding claims.

[0016] According to another aspect of the invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements a personalized tumor motion evolution prediction method as described in any of the preceding claims.

[0017] In summary, the technical solutions conceived in this invention, compared with the prior art, can achieve the following beneficial effects: 1. This invention aims to construct a basic model family through anatomical partitioning and introduce the Cardiac-induced Motion Coupling Factor (CMCF) as a spatial constraint. By combining transfer learning and an evolutionary strategy based on clustering experience playback, it achieves high-precision, self-optimizing personalized tumor motion prediction and solves the model failure problem under complex respiratory patterns.

[0018] 2. High-precision adaptation: Fine-grained modeling is performed for different lung lobes and heart distances, further reducing prediction errors.

[0019] 3. Cold start and rapid response: Transfer learning significantly reduces the need for training data and modeling time for new patients.

[0020] 4. System stability: The anomaly detection triggering logic effectively handles non-stable respiratory signals, ensuring clinical irradiation safety.

[0021] 5. Lifelong learning: The model continues to evolve as more cases are enrolled, exhibiting strong generalization ability and a long lifespan. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall system workflow of a preferred embodiment of the present invention (including classification training, transfer learning and real-time prediction modules).

[0023] Figure 2 This is a preferred embodiment of the architecture diagram for calculating the cardiogenic motion coupling factor (CMCF) and fusing features.

[0024] Figure 3 This is a flowchart of the anomaly detection module and model switching logic in a preferred embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of the basic model evolution mechanism based on clustering experience replay in a preferred embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0027] A preferred embodiment of the present invention provides a personalized tumor motion evolution prediction method comprising the following steps: 1. Multi-source data acquisition and anatomical feature extraction 1.1. Coordinate Unification: Coordinate transformation was performed on the 3D point cloud data of the body surface acquired by the structured light camera and the lung tumor target area data acquired by 4DCT to ensure that the body surface point cloud and the in vivo tumor target area are on the same spatial reference. Respiratory features were extracted from the body surface point cloud, and lung lobe information and tumor target area location were extracted from the 4DCT.

[0028] 1.2. Feature Extraction of Body Surface Point Cloud Based on Lung Lobe Correlation: The body surface point cloud is divided into M×N sub-blocks, where M and N are empirical values. Based on prior anatomical knowledge, the motion correlation (e.g., Pearson correlation coefficient) between each sub-block and the five lung lobe regions is calculated, and different anatomical weights are assigned. Feature vectors are selected. It contains information on the K sub-blocks with the highest motion saliency and their respective lung lobe labels. Feature vector In the form of: , in, It is a small piece on the body surface With lung lobes The Pearson correlation coefficient between them It is the first The mean z-axis coordinates of each sub-block at time t, i = 1~K, j = 1~5.

[0029] Pearson correlation coefficient value The calculation formula is as follows: , in, It is a small piece on the body surface With lung lobes The Pearson correlation coefficient between them =1~10 represents the sequence number of the 10 phases into which respiration is divided. Indicates small patches on the body surface In the Mean Z-axis coordinates of respiratory phase. Indicates small patches on the body surface Z-axis mean across all phases Indicates lung lobe In the Z-axis centroid coordinates of the respiratory phase Indicates lung lobe Mean Z-axis centroid coordinates across all breathing phases.

[0030] 2. Introduce the basic model family construction of CMCF 2.1. CMCF Condition Input: For tumors located near the heart, define the cardiogenic motion coupling factor. : , in, This represents the minimum Euclidean distance from the tumor center to the heart. Let be the attenuation coefficient, and e be the natural logarithm. As a conditional input to a temporal neural network, it can guide the model to adaptively adjust its focus on high-frequency motion features. The temporal neural network can be a conventional network, such as LSTM or TCN.

[0031] 2.2. Classification Training: Based on the location of the tumor in the lung lobe, the historical database was divided into five subsets corresponding to the five lung lobes, and feature vectors were used simultaneously. and cardiogenic motion coupling factor The base models for five subsets were trained, with each subset trained independently. The feature vectors of the point clouds from each patient's body surface were then processed. The time series data of tumor target centroid motion and tumor target area centroid motion are aligned on the timestamp, and a coupling factor is added to train a body surface-tumor motion association model. The input is... and The output is the three-dimensional coordinates of the centroid of the tumor target area.

[0032] 3. Personalized transfer learning and specialized model generation 3.1. Model Loading: Based on the lung lobe to which the new patient's tumor belongs, the corresponding base model weights are called. Since the patient needs to undergo 4D CT to delineate the tumor target area before radiotherapy, the lung lobe to which the new patient's tumor belongs can be identified based on the CT information, which is known information.

[0033] 3.2. Fine-tuning: During the treatment plan preparation phase, acquire the patient's synchronous data, freeze the shallow feature layers of the neural network association model selected in step 2, and update the parameters only for the deep mapping layers to quickly generate a personalized model. The duration of the patient's synchronous data is preferably 1-3 minutes, which covers approximately several dozen respiratory cycles, sufficient for model fine-tuning. Since the network layers near the input end of the neural network association model are mainly responsible for extracting general temporal motion features from body surface motion, exhibiting strong commonality among different patients; while the network layers near the output end are mainly responsible for mapping temporal features to tumor motion coordinates for specific patients, showing significant individual differences, the shallow network parameters of the base model are frozen, and only the deep network parameters are updated. The parameter update uses new patient synchronous data as a supervision signal, employs the same loss function as the base model training, and iteratively optimizes the parameters of the deep network through backpropagation and gradient descent algorithms.

[0034] 4. Anomaly detection and triggering logic in real-time prediction 4.1. Anomaly detection module: Real-time monitoring of the input sequence of point cloud on the body surface and calculation of its reconstruction residual R.

[0035] , in, It is a predicted feature based on historical stationary series.

[0036] 4.2. Triggering Logic: If R < threshold T, maintain the prediction of the dedicated model.

[0037] If R ≥ threshold T, it is determined to be an unexpected movement such as coughing or body slippage. The system automatically triggers an early warning and switches to a more robust baseline model to perform downgraded prediction until the respiratory signal returns to stability. Stopping prediction would cause the radiation beam to fail to track the tumor target area, mistakenly irradiating normal tissue. The more robust downgraded prediction ensures the safe continuity of radiation therapy.

[0038] 5. Evolution of the basic model based on clustering experience replay 5.1. Sample Clustering: For data validated from new patients, K-means clustering is performed based on respiratory rate, amplitude, and phase characteristics to select representative respiratory samples for storage in the experience pool. After K-means clustering, each cluster represents a typical respiratory pattern, and the cluster center or the sample closest to the cluster center is the representative sample of that pattern.

[0039] 5.2. Anti-forgetting Training: Old samples from the experience pool were periodically mixed with data from newly enrolled patients to incrementally train the five base models. Regularization constraints were introduced to ensure that the models did not lose their predictive ability for existing respiratory patterns while learning new features. The incremental training method involved mixing old samples from the experience pool with data from new patients and continuing supervised training on the corresponding lung lobe base models to update network parameters. Regularization constraints could employ conventional methods, such as using elastic weights to solidify EWC or L2 regularization. Taking EWC as an example, EWC applies a secondary penalty to important weights related to existing respiratory patterns, limiting the magnitude of model parameter updates and ensuring that the model does not destroy learned respiratory pattern features when learning features from newly enrolled patients.

[0040] The advantages of this invention are as follows: Anatomy-guided modeling: For the first time, the lung lobe to which the tumor belongs was used as the basis for classification training, and the distance to the heart was quantified as a dynamic constraint, which improved the model's explanatory power for complex motions.

[0041] Two-layer prediction mechanism: Combining the accuracy of the dedicated model with the robustness of the base model, dynamic switching is achieved through anomaly detection.

[0042] Dynamic evolution strategy: By adopting cluster-based experience playback, the catastrophic forgetting problem in incremental learning of deep learning is solved, thereby optimizing and upgrading the basic model to better serve subsequent patients.

[0043] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A personalized method for predicting tumor movement and evolution, characterized in that, include: Multi-source data acquisition and anatomical feature extraction: Respiratory features extracted from body surface point clouds, lung lobe information extracted from 4DCT, and tumor target region locations are transformed to the same spatial reference. The body surface point cloud is divided into multiple sub-blocks, and the motion correlation between each sub-block and the five lung lobe regions is calculated. Different anatomical weights are assigned to each sub-block, resulting in a feature matrix containing information from the multiple sub-blocks with the highest motion significance and their respective lung lobe labels. ; Construction of a basic family of models for CMCF (cardiac motion coupling factor): Defining cardiac motion coupling factors for tumors located near the heart. Cardiac motion coupling factor ,in, This represents the minimum Euclidean distance from the tumor center to the heart. The attenuation coefficient is... It is the natural logarithm; Based on the location of the tumor in the lung lobe, the historical database was divided into five subsets corresponding to the five lung lobes, and feature matrices were used. and cardiogenic motion coupling factor Train the base model corresponding to 5 subsets, and output the three-dimensional coordinates of the centroid of the tumor target area; Personalized transfer learning and specialized model generation: Based on the lung lobe to which the new patient's tumor belongs, the corresponding base model weights are invoked; During the treatment plan preparation phase, the patient's synchronous data is acquired, the shallow feature layer of the base model is frozen, and only the parameters of the deep mapping layer are updated to generate a personalized model for the patient.

2. The personalized tumor motion evolution prediction method as described in claim 1, characterized in that, It also includes anomaly detection and triggering logic in real-time prediction: real-time monitoring of the input sequence of body surface point cloud and calculation of its reconstruction residual R; if R < threshold T, maintain the prediction of the dedicated model; if R ≥ threshold T, it is judged as an anomaly, and the model is switched to the basic model to perform downgraded prediction until the respiratory signal returns to stability.

3. The personalized tumor motion evolution prediction method as described in claim 2, characterized in that, Reconstructing residuals ,in, It is a predicted feature based on historical stationary series.

4. The personalized tumor motion evolution prediction method as described in claim 1, characterized in that, It also includes the evolution of basic models based on clustering experience playback: after the data of new patients are validated, clustering is performed according to respiratory frequency, amplitude and phase characteristics, and representative respiratory samples are selected and stored in the experience pool; old samples in the experience pool are extracted and mixed with the data of newly enrolled patients to incrementally train the five basic models.

5. The personalized tumor motion evolution prediction method as described in claim 1, characterized in that, Feature vector In the form of: ， in, It is a small piece on the body surface With lung lobes The correlation between them It is the first The mean z-axis coordinates of each sub-block at time t, i = 1~K, j = 1~5, where K is the number of sub-blocks with the highest motion significance selected.

6. The personalized tumor motion evolution prediction method as described in claim 5, characterized in that, The correlation mentioned is the Pearson correlation coefficient. The calculation formula is as follows: ， in, =1~10 represents the sequence number of the 10 phases into which respiration is divided. Indicates small patches on the body surface In the Mean Z-axis coordinates of respiratory phase. Indicates small patches on the body surface Z-axis mean across all phases Indicates lung lobe In the Z-axis centroid coordinates of the respiratory phase Indicates lung lobe Mean Z-axis centroid coordinates across all breathing phases.

7. A personalized tumor motion evolution prediction system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement a personalized tumor motion evolution prediction method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a personalized tumor motion evolution prediction method as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements a personalized tumor motion evolution prediction method as described in any one of claims 1 to 6.

Images