Device for tracking tumor position in real time in radiotherapy

Abstract

The invention discloses a device for tracking a tumor position in real time in radiotherapy, and solves the problems of delay and inaccuracy in a traditional tumor position prediction method through a multi-model fusion method combining a deep learning model and a mathematical prediction model. Different from a traditional prediction method depending on an in-vitro marker or iconography technology, the tumor position is predicted in real time by adopting a long-short-term memory network, and the change of the tumor position along with respiratory movement is accurately captured by utilizing the strong time sequence modeling capability of the long-short-term memory network. By introducing the mathematical prediction model to compensate the future position of the tumor, the method can effectively solve the prediction error caused by system delay, and ensures that rays can accurately irradiate the tumor area in dynamic tumor tracking radiotherapy, thereby improving the accuracy and treatment effect of radiotherapy.

Description

Technical Field

[0001] The invention belongs to the field of tumor position tracking, and in particular, relates to a device for real-time tracking of tumor position during radiotherapy. Background Art

[0002] Radiotherapy (RT) plays an important role in cancer treatment. However, the treatment of thoracic and abdominal tumors faces the problem of tumor position changes caused by respiratory motion. This change causes the radiation beam to irradiate healthy tissue, increasing the uncertainty of radiotherapy. To solve this problem, this paper proposes a multi-model fusion method that combines deep learning (LSTM) with a mathematical prediction model to improve the accuracy of tumor position prediction during radiotherapy and compensate for tumor position changes caused by respiratory motion. This method predicts tumor position in real time through a deep learning model and combines a mathematical model to compensate for system delays in radiotherapy. Experimental results show that the fusion model can effectively reduce the prediction error and delay effects during radiotherapy, providing more stable and accurate technical support for precision radiotherapy.

[0003] With the continuous development of radiotherapy technology, precision radiotherapy has become the key to tumor treatment. However, the treatment of thoracic and abdominal tumors faces unique challenges, especially the change of tumor position caused by respiratory movement. Real-time positioning of tumors is crucial to the success of radiotherapy. However, the displacement of tumors caused by breathing often makes it impossible for the radiation beam to accurately irradiate the tumor area, thus affecting the treatment effect.

[0004] Traditional tumor location prediction methods rely on in vitro markers or imaging techniques, but these methods have large delays and inaccuracies. In recent years, deep learning technology, especially long short-term memory networks (LSTM), has been widely used in time series prediction and has shown great potential in the biomedical field. However, existing deep learning models often face system delay problems in practical applications, which leads to a decrease in the accuracy of tumor location prediction.

[0005] To overcome this problem, this paper proposes a multi-model fusion method that combines a deep learning model with a mathematical prediction model. Through this method, we can predict the tumor location in real time and compensate for the change of tumor location in the case of system delay, thereby improving the accuracy of radiotherapy.

[0006] In recent years, significant progress has been made in the prediction of tumor position in radiotherapy. Traditional methods mainly rely on in vitro respiratory motion data and perform position prediction through machine learning algorithms. However, these methods often have delay and accuracy issues. Deep learning methods, especially models based on recurrent neural networks (RNN) and long short-term memory networks (LSTM), have become an important means of tumor position prediction in recent years. LSTM has shown good performance in processing respiratory motion data due to its unique gating mechanism, which can effectively handle long-term dependency problems.

[0007] However, a single deep learning model still has the problem of system delay, especially in dynamic tumor tracking radiation technology, where delay can cause the radiation to fail to accurately irradiate the tumor area, affecting the treatment effect. To this end, a mathematical prediction model is proposed to compensate for this delay and reduce the error caused by delay by predicting the future position of the tumor.

[0008] In view of this, the present invention is proposed. Summary of the invention

[0009] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a device for real-time tracking of tumor positions during radiotherapy, thereby solving the problems raised in the above-mentioned background technology.

[0010] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0011] A device for real-time tracking of tumor position during radiotherapy, comprising: a data acquisition module, the data acquisition module acquires extracorporeal respiratory motion data through a sensor or imaging equipment, the data reflecting the patient's respiratory signal and its influence on tumor motion;

[0012] A deep learning module, wherein the deep learning module uses a bidirectional long short-term memory network to train the in vitro respiratory motion data, generate a time series prediction model for the tumor position, and predict the current position of the tumor in real time based on the current respiratory signal;

[0013] Mathematical prediction module: The mathematical prediction module predicts the future position of the tumor through a mathematical model based on the tumor kinematics law;

[0014] Fusion module: The fusion module fuses the deep learning prediction results with the mathematical model prediction results through weighted averaging to optimize the final prediction value of the tumor location;

[0015] The radiotherapy equipment adjustment module adjusts the radiation beam direction of the radiotherapy equipment according to the final tumor position prediction value obtained to ensure that the radiation beam accurately irradiates the tumor area.

[0016] Optionally, the extracorporeal respiratory movement data includes a respiratory signal acquired by a sensor, and the data is used to train a deep learning model after noise removal, data normalization and smoothing.

[0017] Optionally, the deep learning module uses a bidirectional long short-term memory network to train the in vitro respiratory motion data to generate a time series prediction model for the tumor location in the following steps:

[0018] S201: Use an optical tracking system and other sensor devices to synchronously collect the patient's respiratory movement data. Then, identify and delete outliers in the data using the quartile method, and calculate the quartiles (Q1, Q3) and upper and lower bounds (UB, LB) of the data. Outliers outside these ranges are marked and deleted. The expressions are: UB = Q3 + 1.0 × (Q3 - Q1), LB = Q1 - 1.0 × (Q3 - Q1);

[0019] S202: Construct a DeepBi-LSTM network with 1 input layer and 7 LSTM layers; each LSTM output is connected to the Dropout layer with the parameter set to 0.2, and the last layer uses the hyperbolic tangent function as the activation function of the fully connected layer, and upload the data processed in step S201 to the DeepBi-LSTM network.

[0020] Optionally, the following steps need to be performed after uploading the data processed in step S201 to the DeepBi-LSTM network:

[0021] S2021: Input the preprocessed respiratory motion data to an input layer, where the input of the input layer is time series data, represented as Where N is the batch size, T is the time step, and F is the number of features at each time step;

[0022] S2022: Use a 7-layer bidirectional long short-term memory network. Each Bi-LSTM layer processes the current time further data and considers the previous and next time sequence information. The forward and reverse LSTMs are used to generate hidden states respectively. and And concatenate the two to form the final hidden state h t For the next level of processing

[0023] S2023: After the Bi-LSTM output of each layer, a Dropout layer is connected to randomly discard the output of some neurons to reduce the overfitting risk of the model and enhance the generalization ability of the model;

[0024] S2024: After the output of the last layer of Bi-LSTM, connect a fully connected layer and use the hyperbolic tangent activation function to map the output of LSTM into the predicted value y of the tumor location t .

[0025] Optionally, based on the tumor kinematics, a mathematical model is used to predict the future position of the tumor:

[0026] A1: Use the speed and acceleration of the tumor in space to predict its future position. The position of the tumor y t The motion at a further time t is described by the following acceleration equation: Among them, y t is the tumor position at time t, v t is the speed at the current moment, a t is the acceleration at the current moment, Δt is the predicted time step;

[0027] A2: The motion parameters of the tumor are estimated based on the patient's breathing curve and historical tumor location data. These estimated values ​​can be calculated using the following formula: Among them, v t and v t-1 are the speeds at the current moment and the previous moment respectively, a t is the acceleration at the current moment;

[0028] A3: Using the above tumor kinematic equation, based on the current time y t , speed v t and acceleration a t , and the set time step Δt, predict the position y of the tumor at the next time t+1 t+1 .

[0029] Optionally, the steps of fusing the obtained deep learning prediction results with the mathematical model prediction results by weighted averaging to optimize the final prediction value of the tumor location are:

[0030] B1: Set a weighting coefficient α to represent the weight of the deep learning model and the mathematical model in the final prediction result, which is expressed as: in, is the final tumor location prediction value after fusion, is the real-time prediction result of the deep learning model. It is the future position prediction result of the mathematical model;

[0031] B2: Dynamically adjust the weighting coefficient α according to the prediction accuracy and real-time performance of each model. If the prediction error of the deep learning model at the current moment is small, Bi-LSTM can be given a higher weight. If the mathematical model has a small error in predicting future positions, the mathematical model can be given a higher weight.

[0032] B3: The dynamic adjustment of the weighting coefficient α can be optimized based on error analysis, and its expression is: Among them, MSE (Math) is the mean square error of the mathematical model, and λ is the regularization coefficient, which controls the range of the weighting coefficient.

[0033] Optionally, the steps of adjusting the radiation beam direction of the radiotherapy device according to the obtained final tumor position prediction value to ensure that the radiation beam accurately irradiates the tumor area are:

[0034] C1: Prediction value by final tumor location Determine the target position of the radiation beam, that is, the spatial coordinates of the tumor;

[0035] C2: Target location Calculated by the positioning system of the radiotherapy equipment, combined with the patient's CT image and real-time prediction data to ensure accurate tumor positioning, the expression is: Where f(·) represents the function for calculating the target position;

[0036] C3: According to the target position Adjust the radiation beam direction of the radiotherapy equipment to ensure that the radiation beam accurately irradiates the tumor area;

[0037] C4: Radiation beam direction of radiotherapy equipment θ t It needs to be calculated according to the spatial position of the tumor, which is expressed as: Among them, θ t is the angle or direction of the radiation beam, g(·) is the function for adjusting the direction of the radiation beam, and device parameters include various control parameters of the radiotherapy equipment.

[0038] Optionally, due to the time delay between the radiotherapy equipment and the prediction model, there is a time difference between the prediction result and the equipment response. After the time difference occurs, a real-time feedback mechanism is required to ensure that the delay does not affect the accurate positioning of the radiation beam. If there is a delay, it is adjusted according to historical data and the delay compensation model, and its expression is: in, is the delay difference between the predicted and actual tumor location.

[0039] The respiratory data collected in the experiment of the present invention comes from two places, one is the volunteer, and the other is the respiratory phantom, and the marker is placed on the chest and abdomen of the volunteer. When collecting the respiratory data of the phantom, the marker is placed on the phantom motion module.

[0040] The respiratory data used is the respiratory curve data of the CyberKnife system treatment process at Georgetown University Hospital, where the data is collected by the CyberKnife optical tracking system. This is an open source data set that contains respiratory records of 103 patients with chest and abdominal tumors, all of whom have malignant tumors in their lungs. Each patient has three continuous markers on the chest and abdomen, and the motion curves of the three parts of the markers with respiratory movement are collected separately, with a total of 306 respiratory motion trajectories, which do not contain the patient's personal information. The movement of the three markers is recorded using an optical tracking device with 26Hz sampling. During the tumor radiotherapy process affected by respiratory movement, it is necessary to use external markers placed on the chest and / or abdomen. The marker movement data is used to determine the respiratory movement during the gap between two x-ray image acquisitions, which is performed at irregular intervals. This method can be used to minimize the radiation dose of imaging. The signal of the external marker is collected at a frequency of approximately 27Hz, and the movement of the tracking marker is detected and tracked using an optical tracking system.

[0041] The markers tracked by NDI are objects. NDI can track four small balls (A, B, C, D) at one time to calculate the three-dimensional coordinates and set the origin coordinates. The default sampling frame rate of the NDI system is 60Hz, and the sampling frame rate can be set to 5Hz, 10Hz, 20Hz, and 30Hz.

[0042] Outliers in the respiratory motion sequence data will distort the experimental results. Only by eliminating outliers can better experimental results be achieved. In the process of deleting outliers, the abnormal data is judged according to the quartiles. The quartiles are to divide all the data in order of size and divide all the data into four equal parts. Each equal part contains 25% of the data. There are three quartiles, represented by Q1, Q2 and Q3 respectively. The upper and lower bounds of the box plot are represented by UB and LB, and the calculation is as follows: UB = Q3 + 1.0 * (Q3-Q1); LB = Q1-1.0 * (Q3-Q1). If the value in the respiratory trajectory is greater than UB, then the value is equal to UB. Similarly, if the value is less than LB, then the value is equal to LB.

[0043] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0044] The present invention overcomes the delay and inaccuracy problems existing in traditional tumor position prediction methods by combining a multi-model fusion method of a deep learning model with a mathematical prediction model. Unlike traditional prediction methods that rely on in vitro markers or imaging techniques, this application uses a long short-term memory network (LSTM) to predict tumor position in real time, using its powerful time series modeling capabilities to accurately capture changes in tumor position with respiratory movement. By introducing a mathematical prediction model to compensate for the future position of the tumor, this method can effectively solve the prediction error caused by system delays, ensuring that in dynamic tumor tracking radiotherapy, the rays can accurately irradiate the tumor area, thereby improving the accuracy and therapeutic effect of radiotherapy.

[0045] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0047] Figure 1 Block diagram of a device for structural real-time tracking of tumor location.

[0048] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0049] The present invention will now be described in further detail with reference to the accompanying drawings.

[0050] See also Figure 1 As shown, in this embodiment, a device for real-time tracking of tumor position during radiotherapy is provided, including a data acquisition module, which acquires extracorporeal respiratory motion data through a sensor or an imaging device, wherein the data reflects the patient's respiratory signal and its influence on tumor motion;

[0051] A deep learning module, wherein the deep learning module uses a bidirectional long short-term memory network to train the in vitro respiratory motion data, generate a time series prediction model for the tumor position, and predict the current position of the tumor in real time based on the current respiratory signal;

[0052] The mathematical prediction module predicts the future position of the tumor through a mathematical model based on the kinematic law of the tumor; the mathematical prediction model is based on the tumor kinematic model to model the movement trajectory of the tumor to predict the future position of the tumor.

[0053] A fusion module, which fuses the deep learning prediction results with the mathematical model prediction results through a weighted average method to optimize the final prediction value of the tumor location; the weighted average method determines the weight coefficients of the prediction results of different models based on the prediction error and real-time performance of the deep learning model and the mathematical prediction model.

[0054] The radiotherapy equipment adjustment module adjusts the radiation beam direction of the radiotherapy equipment according to the final tumor position prediction value obtained to ensure that the radiation beam accurately irradiates the tumor area.

[0055] The present invention overcomes the delay and inaccuracy problems existing in traditional tumor position prediction methods by combining a multi-model fusion method of a deep learning model with a mathematical prediction model. Unlike traditional prediction methods that rely on in vitro markers or imaging techniques, this application uses a long short-term memory network (LSTM) to predict tumor position in real time, using its powerful time series modeling capabilities to accurately capture changes in tumor position with respiratory movement. By introducing a mathematical prediction model to compensate for the future position of the tumor, this method can effectively solve the prediction error caused by system delays, ensuring that in dynamic tumor tracking radiotherapy, the rays can accurately irradiate the tumor area, thereby improving the accuracy and therapeutic effect of radiotherapy.

[0056] In this embodiment, the extracorporeal respiratory motion data includes a respiratory signal obtained by a sensor, and the data is used to train a deep learning model after noise removal, data normalization and smoothing.

[0057] In this embodiment, an optical tracking system (cyber knife system) and other sensor devices (chest and abdomen displacement sensor, CT scanner, infrared camera) are used to synchronously collect the patient's respiratory movement data. Then, the outliers in the data are identified and deleted by the quartile method, and the quartiles (Q1, Q3) and upper and lower bounds (UB, LB) of the data are calculated. Outliers beyond these ranges are marked and deleted, and the expressions are: UB = Q3 + 1.0 × (Q3-Q1), LB = Q1-1.0 × (Q3-Q1);

[0058] Construct a DeepBi-LSTM network with 1 input layer and 7 LSTM layers; each LSTM output is connected to the Dropout layer, the parameter (DropoutRatio) is set to 0.2, and the last layer uses the hyperbolic tangent function as the activation function of the fully connected layer, and upload the data processed in step S201 to the DeepBi-LSTM network.

[0059] In this embodiment, the following steps need to be performed after uploading the data processed in step S201 to the DeepBi-LSTM network:

[0060] The preprocessed respiratory motion data is input into an input layer, the input of which is time series data, represented as Where N is the batch size, T is the time step, and F is the number of features at each time step;

[0061] A 7-layer bidirectional long short-term memory network (Bi-LSTM) is used. Each Bi-LSTM layer processes the current time further data and considers the previous and next time sequence information. The forward and reverse LSTMs are used to generate hidden states respectively. and And concatenate the two to form the final hidden state h t For the next level of processing

[0062] After the Bi-LSTM output of each layer, a Dropout layer is connected to randomly discard the output of some neurons (the discard ratio is 0.2) to reduce the overfitting risk of the model and enhance the generalization ability of the model;

[0063] After the output of the last layer of Bi-LSTM, a fully connected layer is connected and the hyperbolic tangent (Tanh) activation function is used for mapping to convert the output of LSTM into the predicted value y of the tumor location. t Bi-LSTM can simultaneously consider the current time step and the time series data before and after it by passing information forward and backward. This enables the model to capture long-term dependencies and the relationship between the previous and next time series, and more accurately predict the position change of the tumor when processing respiratory motion data. For example, the position of the tumor depends not only on the input signal at the current moment, but may also be affected by the previous or next time steps.

[0064] Adding a Dropout layer (with a dropout rate of 0.2) after each Bi-LSTM layer can effectively reduce overfitting. Overfitting is a common problem in deep learning, especially when there is relatively little training data or the model structure is complex. By randomly discarding the output of a portion of neurons, the Dropout layer forces each neuron in the network to not be overly dependent on a specific feature, thereby enhancing the generalization ability of the model and improving its adaptability to unseen data, thereby improving the stability and reliability of tumor location prediction.

[0065] After the last layer of Bi-LSTM, a fully connected layer is connected and mapped using the hyperbolic tangent (Tanh) activation function. The Tanh function can compress the output to between [-1,1], which helps to reduce the fluctuation range of the model output and make the model more stable. At the same time, the fully connected layer is used to further convert the output of LSTM into the predicted value of the tumor location, so that the model can extract effective spatial location information from the processed complex time series data.

[0066] In this embodiment, based on the kinematics of the tumor, the future position of the tumor is predicted by a mathematical model:

[0067] A1: Use the speed and acceleration of the tumor in space to predict its future position. The position of the tumor y t The motion at a further time t is described by the following acceleration equation: Among them, y t is the tumor position at time t, v t is the speed at the current moment, a t is the acceleration at the current moment, Δ t is the time step of prediction;

[0068] A2: The motion parameters of the tumor are estimated based on the patient's breathing curve and historical tumor location data. These estimated values ​​can be calculated using the following formula: Among them, v t and v t-1 are the speeds at the current moment and the previous moment respectively, a t is the acceleration at the current moment.

[0069] A3: Using the above tumor kinematic equation, based on the current time y t , speed v t and acceleration a t , and the set time step Δt, predict the position y of the tumor at the next time t+1 t+1 . The present invention utilizes the dynamic behavior of the tumor in space (i.e., position, velocity, and acceleration), combined with the patient's breathing curve and historical tumor position data, to infer the future position of the tumor through kinematic equations. By combining the acceleration equation and the kinematic model, the position error caused by system delay can be effectively compensated, ensuring that the radiotherapy equipment can make rapid and accurate adjustments during the process of tumor position changes. This is crucial for real-time dynamic tumor tracking radiotherapy, which can provide higher accuracy, reduce radiation damage to healthy tissues, and optimize treatment effects.

[0070] In this embodiment, the steps of fusing the obtained deep learning prediction results with the mathematical model prediction results by weighted average method to optimize the final prediction value of the tumor location are as follows:

[0071] B1: Set a weighting coefficient α to represent the weight of the deep learning model and the mathematical model in the final prediction result. The weighting coefficient α is between 0 and 1, indicating the relative importance of the Bi-LSTM model and the mathematical model, which is expressed as: in, is the final tumor location prediction value after fusion, is the real-time prediction result of the deep learning model. It is the future position prediction result of the mathematical model;

[0072] B2: Dynamically adjust the weighting coefficient α according to the prediction accuracy and real-time performance of each model. If the prediction error of the deep learning model at the current moment is small, Bi-LSTM can be given a higher weight. If the mathematical model has a small error in predicting future positions, the mathematical model can be given a higher weight.

[0073] B3: The dynamic adjustment of the weighting coefficient α can be optimized based on error analysis, and its expression is: Among them, MSE (Math) is the mean square error of the mathematical model, and λ is the regularization coefficient, which controls the range of the weighting coefficient. By setting the weighting coefficient, the weights of the deep learning model (such as Bi-LSTM) and the mathematical model in the final tumor position prediction can be dynamically adjusted according to the prediction accuracy and real-time performance of different models. Such a weighting method can flexibly respond to different prediction errors, ensuring that the most appropriate model is selected for adjustment and optimization in different situations, thereby improving the accuracy of tumor position prediction. In this way, the high real-time performance of the deep learning model and the future position prediction ability of the mathematical model can be combined to make up for their respective shortcomings and achieve more accurate and stable tumor position prediction. For tumor localization in radiotherapy, dynamically adjusting the weighting coefficient can enable radiotherapy equipment to better adapt to the real-time changing tumor position and optimize the treatment effect.

[0074] For example: Assume that at the current moment, the prediction error of the Bi-LSTM model is small (for example, MSE(DL) = 0.05), while the prediction error of the mathematical model is large (for example, MSE (Math) =0.2). At this time, the weighting coefficient α will be biased towards the Bi-LSTM model to fully utilize the advantages of the deep learning model in real-time prediction.

[0075] If the error of Bi-LSTM is small, α will increase, giving Bi-LSTM a higher weight. If the error of the mathematical model is small, α will decrease, giving more weight to the mathematical model. The final tumor location prediction value is obtained through the dynamically adjusted weight coefficient α. This predicted value will be used to adjust the radiation beam direction of the radiotherapy equipment to ensure that the radiotherapy rays are accurately irradiated to the tumor area.

[0076] For example, at the current moment, the tumor location given by the deep learning model is The future position given by the mathematical model is The final position of the fused tumor is

[0077] The radiotherapy equipment will adjust the direction of the radiation beam based on this fused position to ensure that the rays are accurately irradiated to the tumor area, reduce damage to healthy tissue, and optimize the treatment effect.

[0078] In this embodiment, the steps of adjusting the radiation beam direction of the radiotherapy device according to the obtained final tumor position prediction value to ensure that the radiation beam accurately irradiates the tumor area are:

[0079] C1: Prediction value by final tumor location Determine the target position of the radiation beam, that is, the spatial coordinates of the tumor. This position will serve as the basis for adjusting the radiotherapy equipment.

[0080] C2: Target location Calculated by the positioning system of the radiotherapy equipment, combined with the patient's CT image and real-time prediction data to ensure accurate tumor positioning, the expression is: Where f(·) represents the function for calculating the target position, combining the patient position and CT images to accurately locate the tumor;

[0081] C3: According to the target position Adjust the direction of the radiation beam of the radiotherapy device to ensure that the radiation beam accurately irradiates the tumor area.

[0082] C4: Radiation beam direction of radiotherapy equipment θ t It needs to be calculated according to the spatial position of the tumor, which is expressed as: Among them, θ t is the angle or direction of the radiation beam, g(·) is the function for adjusting the direction of the radiation beam, and device parameters include various control parameters of the radiotherapy equipment.

[0083] By combining tumor position prediction with real-time data, radiotherapy equipment can dynamically adjust the direction of the radiation beam according to the actual tumor position. This method effectively solves the error accumulation and delay compensation problems that may exist in traditional radiotherapy, ensuring that the radiation beam can accurately irradiate the tumor area in real time, reducing damage to healthy tissues and improving the accuracy and effectiveness of radiotherapy treatment. By combining CT images, real-time prediction data and the positioning system of radiotherapy equipment, this method ensures that the tumor is always within the irradiation range of the radiation beam during treatment, thereby maximizing the treatment effect and reducing side effects.

[0084] Example: Suppose a patient is undergoing dynamic tumor tracking radiotherapy, which uses radiotherapy equipment (such as the CyberKnife system) to track the tumor position in real time and adjust the radiation beam direction according to the tumor movement. During the treatment, the patient's tumor position will change with physiological factors such as breathing. Real-time and accurate prediction of the tumor position and adjustment of the device's radiation beam direction are the key to ensuring treatment effectiveness.

[0085] 1. Determine the radiation beam target position based on the final tumor position prediction value

[0086] At the beginning of treatment, the deep learning model is trained on the patient's breathing signal and historical tumor location data. With this data, the model is able to predict the location of the tumor in real time. Suppose the model predicts the location of the tumor at the current time t is (for example, 3D coordinates: [x=10, y=5, z=2]), which is the final predicted position of the tumor.

[0087] Prediction value: The spatial coordinates of the tumor at the current moment

[0088] Target position: This predicted value will be used as the target position of the radiation beam, that is, the radiotherapy device will determine the irradiation direction based on this position.

[0089] 2. Combine CT images and real-time data to accurately calculate the target position

[0090] The positioning system of the radiotherapy equipment combines the patient's CT image and the real-time prediction data of the tumor (from the Bi-LSTM model). The CT image provides the three-dimensional position coordinates of the tumor in the body, while the Bi-LSTM model provides information on the dynamic changes of the tumor under the influence of factors such as respiratory movement. The combination of the two calculates the precise target position of the tumor through the following formula

[0091] Here, f(·) is a function that combines the real-time prediction data with the CT image to calculate the target position. The localization system of the radiotherapy device uses this calculation to ensure that the exact position of the tumor in the body is accurately determined, taking into account the patient's position and the geometric characteristics of the CT image.

[0092] 3. Adjust the radiation beam direction according to the target location

[0093] By accurately calculating the tumor target location The radiotherapy device will adjust the direction of its radiation beam θ t , ensuring that the radiotherapy rays can accurately irradiate the tumor area. At this time, the device will adjust the direction of the radiation beam according to the following formula:

[0094] Among them, θ t is the direction of the radiation beam, g(·) is the function that calculates the direction of the radiation beam based on the target position, and device parameters include various control parameters of the device, such as the angle, intensity, and irradiation depth of the radiation beam.

[0095] Assume that the target location of the tumor is The control parameters of the radiotherapy device specify the angle that needs to be adjusted. The device will calculate the angle of the radiation that needs to be adjusted based on the target position and the control parameters. For example, the radiation beam direction is adjusted to θ t =[15°, 30°].

[0096] 4. Equipment adjustment ensures accurate radiation beam exposure

[0097] The radiotherapy equipment adjusts the radiation beam direction θ t Adjustments are made to ensure that the radiotherapy rays can accurately irradiate the tumor area. For example, the radiotherapy equipment adjusts the radiation beam direction to θ according to the above adjustments. t = [15°, 30°], so that the radiation beam directly irradiates the spatial location of the tumor Thereby achieving precise treatment.

[0098] In this embodiment, due to the time delay between the radiotherapy equipment and the prediction model, there is a time difference between the prediction result and the equipment response. After the time difference occurs, a real-time feedback mechanism is required to ensure that the delay does not affect the accurate positioning of the radiation beam. If there is a delay, it is adjusted according to the historical data and the delay compensation model, and the expression is: in, is the delay difference between the predicted and actual tumor location.

[0099] The system delay between the radiotherapy equipment and the prediction model is compensated through a real-time feedback mechanism, thereby effectively reducing the impact of the delay on the accuracy of tumor position prediction and ensuring that the radiotherapy rays can accurately irradiate the tumor area. This mechanism can solve the problem of tumor position prediction error caused by system delay and improve the treatment effect of radiotherapy.

[0100] Specifically, the delay compensation mechanism adjusts the movement trajectory of the tumor based on historical data and the delay compensation model, compensates for the prediction error caused by the response time of the radiotherapy equipment, and ensures that the position of the tumor is consistent with the response of the equipment in time, thereby improving the accuracy and treatment effect of radiotherapy.

[0101] For example, suppose a patient's tumor has a 50ms delay in prediction due to changes in respiratory movement. If no compensation is made, the radiotherapy device will irradiate according to the wrong tumor location and may miss the tumor area. By introducing a delay compensation mechanism, the actual location of the tumor can be estimated based on historical data, and the direction of the device's rays can be adjusted so that the radiotherapy rays can compensate for the delay in real time and accurately irradiate the tumor area, avoiding incorrect irradiation and treatment failure.

[0102] Existing technology solution 1: Tumor location prediction based on traditional kinematic model

[0103] This method uses a tumor kinematic model (such as uniform linear motion or uniformly accelerated linear motion model) to predict the tumor position. By collecting respiratory motion data, a simple kinematic equation is used to estimate the tumor position. This method assumes that the movement of the tumor is regular and continuous.

[0104] Advantages: simple calculation and easy to implement.

[0105] Disadvantages: It cannot provide sufficient accuracy for complex tumor motion (such as nonlinear motion or irregular motion caused by various physiological factors), and the prediction error is large, especially when the system is delayed and the data is incomplete.

[0106] Existing technology solution 2: Tumor location prediction based on Kalman filtering

[0107] Kalman filter is a recursive algorithm widely used in state estimation of dynamic systems. This method combines the tumor kinematic model with actual observation data and can predict the movement trajectory of the tumor through recursive updating. Kalman filter is suitable for systems affected by continuous time and noise.

[0108] Advantages: Able to handle noise and uncertainty in dynamic systems and can better compensate for measurement errors.

[0109] Disadvantages: The model assumptions are relatively strict. It assumes that the system state follows a Gaussian distribution and cannot handle highly nonlinear or complex motion patterns.

[0110] Existing technology solution 3: Tumor location prediction based on deep neural network (DNN)

[0111] This method uses a deep neural network (such as a multi-layer perceptron DNN) to train historical respiratory signals and tumor location data to generate a tumor location prediction model. The model makes predictions by learning the nonlinear relationship between input features (such as respiratory signals, patient position, tumor movement patterns, etc.).

[0112] Advantages: It can capture complex nonlinear relationships and is suitable for high-dimensional data. Disadvantages: It requires a large amount of training data, takes a long time to train the model, has overfitting problems, and cannot directly perform time series prediction.

[0113] Comparative Experimental Setup

[0114] Experimental objective: To verify the advantages of the present invention (tumor location prediction method based on deep learning and mathematical prediction model) by comparing with the existing technical solutions.

[0115] Experimental conditions: Dataset: An open source dataset from Georgetown University Hospital was used, which contains respiratory records from 103 patients with thoracic and abdominal tumors. Each patient has three continuous markers that record the changes in the position of the tumor during respiratory movement, and the data sampling frequency is 26Hz.

[0116] Evaluation indicators: prediction error (such as root mean square error RMSE, mean absolute error MAE) model stability (volatility during model training and prediction) delay compensation effect (comparison of the time lag between the response time of radiotherapy equipment and the prediction of tumor location in actual application)

[0117] Comparison scheme selection: Compare the existing technical schemes 1, 2 and 3, use the same dataset, evaluation indicators and experimental settings, and verify the performance of each scheme under the same conditions.

[0118] Experimental results 1: prediction error comparison

[0119]

[0120] The present invention is superior to other existing technical solutions in terms of MAE, RMSE and nRMSE, indicating that it has significant advantages in tumor location prediction accuracy.

[0121] Experimental results 2: Delay compensation effect comparison

[0122]

[0123] Analysis: The present invention performs best in terms of system delay compensation, with high prediction accuracy and shortest equipment adjustment time, proving its real-time advantage in clinical applications.

[0124] Through comparative experiments, combined with data and charts of MAE, RMSE, nRMSE and delay compensation effects, the obvious advantages of the present invention (based on the fusion method of Bi-LSTM and mathematical model) in tumor position prediction accuracy, system delay compensation and radiotherapy equipment response time can be clearly demonstrated. In particular, in terms of delay compensation and real-time performance, the present invention has significant improvements over existing technical solutions (such as kinematic models, Kalman filtering and DNN), and can provide more accurate tumor position prediction, thereby optimizing radiotherapy treatment effects.

[0125] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A device for real-time tracking of tumor position during radiotherapy, characterized in that: include: A data acquisition module, which acquires extracorporeal respiratory motion data through sensors or imaging equipment, wherein the data reflects the patient's respiratory signal and its effect on tumor motion; A deep learning module, wherein the deep learning module uses a bidirectional long short-term memory network to train the in vitro respiratory motion data, generate a time series prediction model for the tumor position, and predict the current position of the tumor in real time based on the current respiratory signal; Mathematical prediction module: The mathematical prediction module predicts the future position of the tumor through a mathematical model based on the tumor kinematics law; Fusion module: The fusion module fuses the deep learning prediction results with the mathematical model prediction results through weighted averaging to optimize the final prediction value of the tumor location; The radiotherapy equipment adjustment module adjusts the radiation beam direction of the radiotherapy equipment according to the final tumor position prediction value obtained to ensure that the radiation beam accurately irradiates the tumor area.

2. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: The extracorporeal respiratory motion data includes a respiratory signal obtained by a sensor, and the data is used to train a deep learning model after noise removal, data normalization and smoothing.

3. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: The deep learning module uses a bidirectional long short-term memory network to train the in vitro respiratory motion data to generate a time series prediction model for the tumor location in the following steps: S201: Use an optical tracking system and other sensor devices to synchronously collect the patient's respiratory movement data. Then, identify and delete outliers in the data using the quartile method, and calculate the quartiles (Q1, Q3) and upper and lower bounds (UB, LB) of the data. Outliers outside these ranges are marked and deleted. The expressions are: UB = Q3 + 1.0 × (Q3 - Q1), LB = Q1 - 1.0 × (Q3 - Q1); S202: Construct a DeepBi-LSTM network with 1 input layer and 7 LSTM layers; each LSTM output is connected to the Dropout layer with the parameter set to 0.2, and the last layer uses the hyperbolic tangent function as the activation function of the fully connected layer, and upload the data processed in step S201 to the DeepBi-LSTM network.

4. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: The following steps need to be performed to upload the data processed in step S201 to the DeepBi-LSTM network: S2021: Input the preprocessed respiratory motion data to an input layer, where the input of the input layer is time series data, represented as Where N is the batch size, T is the time step, and F is the number of features at each time step; S2022: Use a 7-layer bidirectional long short-term memory network. Each Bi-LSTM layer processes the current time further data and considers the previous and next time sequence information. The forward and reverse LSTMs are used to generate hidden states respectively. and And concatenate the two to form the final hidden state h t For the next level of processing S2023: After the Bi-LSTM output of each layer, a Dropout layer is connected to randomly discard the output of some neurons to reduce the overfitting risk of the model and enhance the generalization ability of the model; S2024: After the output of the last layer of Bi-LSTM, connect a fully connected layer and use the hyperbolic tangent activation function to map the output of LSTM into the predicted value y of the tumor location t .

5. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: Based on the kinematics of tumors, mathematical models are used to predict the future position of tumors: A1: Use the speed and acceleration of the tumor in space to predict its future position. The position of the tumor y t The motion at a further time t is described by the following acceleration equation: Among them, y t is the tumor position at time t, v t is the speed at the current moment, a t is the acceleration at the current moment, Δt is the predicted time step; A2: The motion parameters of the tumor are estimated based on the patient's breathing curve and historical tumor location data. These estimated values ​​can be calculated using the following formula: Among them, v t and v t-1 are the speeds at the current moment and the previous moment respectively, a t is the acceleration at the current moment; A3: Using the above tumor kinematic equation, based on the current time y t , speed v t and acceleration a t , and the set time step Δt, predict the position y of the tumor at the next time t+1 t+1 .

6. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: The steps of fusing the deep learning prediction results with the mathematical model prediction results by weighted averaging to optimize the final prediction value of the tumor location are as follows: B1: Set a weighting coefficient α to represent the weight of the deep learning model and the mathematical model in the final prediction result, which is expressed as: in, is the final tumor location prediction value after fusion, is the real-time prediction result of the deep learning model. It is the future position prediction result of the mathematical model; B2: Dynamically adjust the weighting coefficient α according to the prediction accuracy and real-time performance of each model. If the prediction error of the deep learning model at the current moment is small, Bi-LSTM can be given a higher weight. If the mathematical model has a small error in predicting future positions, the mathematical model can be given a higher weight. B3: The dynamic adjustment of the weighting coefficient α can be optimized based on error analysis, and its expression is: Among them, MSE (Math) is the mean square error of the mathematical model, and λ is the regularization coefficient, which controls the range of the weighting coefficient.

7. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: According to the final tumor position prediction value, the steps to adjust the radiation beam direction of the radiotherapy equipment to ensure that the radiation beam accurately irradiates the tumor area are as follows: C1: Prediction value by final tumor location Determine the target position of the radiation beam, that is, the spatial coordinates of the tumor; C2: Target location Calculated by the positioning system of the radiotherapy equipment, combined with the patient's CT image and real-time prediction data to ensure accurate tumor positioning, the expression is: Where f(·) represents the function for calculating the target position; C3: According to the target position Adjust the radiation beam direction of the radiotherapy equipment to ensure that the radiation beam accurately irradiates the tumor area; C4: Radiation beam direction of radiotherapy equipment θ t It needs to be calculated according to the spatial position of the tumor, which is expressed as: Among them, θ t is the angle or direction of the radiation beam, g(·) is the function for adjusting the direction of the radiation beam, and device parameters include various control parameters of the radiotherapy equipment.

8. The device for real-time tracking of tumor position during radiotherapy according to claim 1, characterized in that: Due to the time delay between the radiotherapy equipment and the prediction model, there is a time difference between the prediction result and the equipment response. After the time difference occurs, a real-time feedback mechanism is needed to ensure that the delay does not affect the accurate positioning of the radiation beam. If there is a delay, it is adjusted according to the historical data and the delay compensation model. The expression is: in, is the delay difference between the predicted and actual tumor location.

Images