A real-time target area tracking and dynamic regulation method and system thereof

Abstract

The present invention relates to a real-time target area tracking and dynamic regulation method and system thereof. The method includes obtaining a phase prediction model based on 4DCT reconstruction data and three-dimensional point cloud body surface contour data; obtaining a two-stage model according to the result output by the phase prediction model and the 4DCT reconstruction data; superimposing the phase prediction model and the two-stage model, and fusing to obtain a target area prediction model; acquiring real-time body surface contour data and respiratory signals; inputting the body surface contour data and the respiratory signals into the target area prediction model, and the target area prediction model predicts the current position of the in-vivo target area; calculating the target pose of the accelerator treatment couch according to the current position of the in-vivo target area and causing the accelerator treatment couch to be converted to the target pose. Through the target area prediction model, real-time input data can be used to obtain the real-time position of the in-vivo target area, which can respond to the dynamic changes of the target area in real time and achieve dynamic regulation. The target area prediction model is modeled based on 4DCT reconstruction data, and the prediction result has high accuracy.

Description

Technical Field

[0001] The present invention relates to the field of data prediction and tracking, and particularly relates to a real-time target area tracking and dynamic regulation method. Background Art

[0002] In radiotherapy, the accurate positioning and irradiation of the target area are crucial. The accurate positioning of the target area is to find the location of the target area during the treatment process, and the irradiation of the target area is the dynamic regulation of the radiation beam, specifically to adjust the radiation beam according to the obtained target area position.

[0003] There are three existing target area tracking methods, namely IGRT (Image Guided Radiotherapy), SGRT (Surface Guided Radiotherapy), and CBCT (Cone Beam Computed Tomography). Among them, IGRT is a method that uses imaging technology to monitor and adjust the target area position in real time during the treatment process. Its advantage is that it can provide high-resolution images for more accurate target area positioning. The disadvantage is that it may cause additional radiation exposure to patients due to the use of X-ray imaging, and the imaging and analysis process may prolong the treatment time and affect the treatment efficiency; SGRT is a non-invasive target area tracking method that monitors the changes in the patient's body surface by analyzing the optical markers on the patient's body surface or using a 3D camera. Its advantages are that it does not use radiation imaging, has less radiation exposure to patients, and can monitor the changes in the body surface in real time to provide immediate feedback for target area positioning. The disadvantages are that the image resolution and depth information are poor and it mainly focuses on the changes in the body surface, and it may not accurately reflect the dynamics of the target area in the body. CBCT is an imaging technology used before treatment that can provide relatively high image resolution and depth information. Its disadvantage is that it cannot track the changes in the target area position in real time. The above three methods can be summarized as follows: IGRT provides high-resolution images but is accompanied by radiation exposure; SGRT has no radiation but the image quality and depth information are limited; CBCT can provide detailed image information but cannot achieve real-time tracking.

[0004] Existing dynamic regulation methods include dynamic grating technology and gated radiotherapy. Dynamic grating technology uses a grating system with a dynamically adjustable shape to adjust the shape and size of the irradiation field according to the real-time position of the target area. However, it has the following disadvantages: (1) The regulation method is limited to specific equipment, which requires the equipment to have corresponding dynamic adjustment capabilities; (2) The adjustment of the grating shape may not fully adapt to the complex movement of the target area; (3) Frequent adjustment of the grating shape may affect the patient's treatment experience. Gated radiotherapy combines respiratory gating technology and only turns on the ray irradiation when the target area is in an ideal position, avoiding irradiation when the target area moves to a non-ideal position and reducing damage to normal tissues. But it also has the following disadvantages: (1) The gating technology may require the patient to remain stationary at a specific respiratory stage, which may prolong the time of a single treatment. (2) Not all patients are suitable for gated radiotherapy, especially those with severe respiratory problems or unable to control their breathing. (3) Patient comfort: Requiring the patient to maintain a consistent breathing pattern throughout the treatment process may be a challenge for some patients, especially for situations that require maintaining the same breathing state for a long time. Summary of the Invention

[0005] One of the objectives of the present invention is to solve the problems in the above technologies that accurate information cannot be obtained and real-time tracking and radiation-free damage can be achieved, and to provide a real-time target area tracking and dynamic regulation method and system, which can provide accurate real-time target area position information and keep the target area position within the irradiation range of the rays all the time.

[0006] In order to achieve the above objectives, the technical solutions adopted by the present invention are as follows:

[0007] A real-time target area tracking and dynamic regulation method includes the following steps:

[0008] S1: Obtain a phase prediction model based on 4DCT body surface reconstruction data and three-dimensional point cloud body surface contour data;

[0009] S2: Obtain a two-stage model according to the result output by the phase prediction model and 4DCT in-vivo target area reconstruction data;

[0010] S3: Superimpose the phase prediction model and the two-stage model, and fuse them to obtain a target area prediction model;

[0011] S4: Obtain real-time body surface contour data and respiratory signals;

[0012] S5: Input the body surface contour data and respiratory signals into the target area prediction model, and the target area prediction model predicts the current position of the in-vivo target area;

[0013] S6: Calculate the target pose of the accelerator treatment couch according to the current position of the in-vivo target area, obtain the target position of the actuator of the accelerator treatment couch based on the target pose of the accelerator treatment couch, and drive the actuator to move to the target position so as to adjust the accelerator treatment couch to the said target pose.

[0014] In the above technical solution, the 4DCT body surface reconstruction data includes a respiratory signal and a 4DCT three-dimensional point cloud contour corresponding to the respiratory signal, and the 4DCT in-vivo target area reconstruction data includes target area position information corresponding to the respiratory signal.

[0015] Since in the dataset for training the phase prediction model, the three-dimensional point cloud body surface contour is used as the input of the model, and the in-vivo target area point cloud data and the corresponding phase of the 4DCT body surface reconstruction data are used as the output. Therefore, after obtaining the phase prediction model, through this model, the phase associated with the respiratory signal corresponding to the body surface data and the in-vivo target area point cloud data corresponding to this phase can be predicted based on the body surface data.

[0016] Similarly, in the dataset for training the two-stage model, the 4DCT body surface reconstruction data corresponding to the predicted phase and the three-dimensional point cloud body surface contour data are used as the input of the model, and the target area position information in the 4DCT reconstruction data is used as a reference value. After the 4DCT body surface reconstruction data and the three-dimensional point cloud body surface contour data are associated, the three-dimensional point cloud body surface contour data can be associated with the target area position information in the 4DCT reconstruction data, that is, the corresponding target area position information can be obtained from the three-dimensional point cloud body surface contour data. After obtaining and training the two-stage model, the target area position information can be predicted based on the predicted phase and the three-dimensional point cloud body surface contour data.

[0017] After superimposing the phase prediction model and the two-stage model, a target area prediction model is obtained. When it is necessary to track the target area, only the real-time body surface data and the respiratory signal need to be input into the target area prediction model. The input data is calculated by the phase prediction model and the two-stage model, and finally the current position of the predicted in-vivo target area is output. After obtaining the current position of the in-vivo target area, the corresponding target pose of the accelerator treatment couch is obtained through conversion. The accelerator treatment couch enables the actuator of the accelerator treatment couch to reach the target position according to the target pose, thereby realizing the transformation of the pose. After the accelerator treatment couch is converted to the target pose, the ray irradiates the current position of the in-vivo target area. The in-vivo target area changes with respiration, and the real-time body surface data and the respiratory signal also change accordingly, and the position of the in-vivo target area predicted by the target area prediction model also changes accordingly. Finally, the accelerator treatment couch can follow the position of the target area in real time, so that the ray can always irradiate only at the target area.

[0018] Preferably, in step S1, the specific process is as follows:

[0019] S1.1: Obtain the respiratory signal and corresponding image information within at least one respiratory cycle through 4DCT. The image information includes the 4DCT three-dimensional point cloud contour and the target area position information. Divide the respiratory phase into N phases on average according to the respiratory signal, and associate the corresponding 4DCT three-dimensional point cloud contour and target area position information for each phase.

[0020] S1.2: Obtain the three-dimensional point cloud body surface contour of at least one respiratory cycle.

[0021] S1.3: Align the three-dimensional point cloud body surface contour with the 4DCT three-dimensional point cloud contour to obtain the phase corresponding to the three-dimensional point cloud body surface contour and perform annotation.

[0022] S1.4: Establish a data set based on the data and phase annotation obtained in S1.1 - S1.3.

[0023] S1.5: Input the data set into the first neural network model for training to obtain the phase prediction model, and the phase prediction model outputs the predicted phase.

[0024] Preferably, in step S1.3, calculate the average vector distance between the point cloud data in the three-dimensional point cloud body surface contour and the point cloud data in the N 4DCT three-dimensional point cloud contours, and use the phase corresponding to the 4DCT three-dimensional point cloud contour with the minimum average vector distance as the phase of the three-dimensional point cloud body surface contour.

[0025] Preferably, the first neural network includes a PointNet layer, a self-attention mechanism layer, and a fully connected layer, which are used to extract the features of the point cloud and perform classification.

[0026] Preferably, in step S2, use the point cloud data of the 4DCT three-dimensional point cloud contour corresponding to the predicted phase and the point cloud data of the three-dimensional point cloud body surface contour obtained by the phase prediction model as input values and input them into the second neural network model for training to obtain the two-stage model, and the two-stage model outputs the predicted target area position information. The 4DCT in-vivo target area reconstruction data includes the target area position information, which is used as the output value during the training of the two-stage model.

[0027] Preferably, the second neural network model includes a DGCNN layer, a Transformer layer, and a Pointer layer.

[0028] Preferably, the loss function of the two-stage model is defined as the difference between the predicted target area position information and the target area position information in S1.1; the loss function has an L2 regularization term. The specific loss function is:

[0029]

[0030] In the formula, represents the second loss function; represents the transpose matrix of the rotation matrix for transforming the three-dimensional point cloud body surface contour number of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the rotation matrix for transforming the three-dimensional point cloud body surface contour of each breathing phase to the point cloud data corresponding to the corresponding true phase; represents the identity matrix; represents the translation matrix for transforming the three-dimensional point cloud body surface contour of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the translation matrix for transforming the body surface point cloud contour of each breathing phase to the point cloud data corresponding to the corresponding true phase, represents the regularization coefficient, represents the training parameter.

[0031] Preferably, before step S4, the target area prediction model is optimized. The specific steps are as follows:

[0032] The data of the in-vivo target area is obtained through CBCT scanning and used as training data to be input into the phase prediction model for training, and the parameters of the phase prediction model are adjusted. Specifically, the data of the in-vivo target area obtained through CBCT scanning includes the shape, size, position of the target area, and the correlation with the body surface and in-vivo movements; the correlation of the body surface and in-vivo movements is extracted through convolution operations of the convolutional neural network; the data of the in-vivo target area is input into the phase prediction model for training through the transfer learning method.

[0033] Preferably, the body surface contour data and breathing signals of the patient are monitored in real time through the body surface optical guidance system.

[0034] Preferably, in step S6, the specific steps include:

[0035] S6.1: Calculate the conversion deviation corresponding to the target pose of the accelerator treatment couch when the current position of the in-vivo target area is converted;

[0036] S6.2: Obtain the current pose of the accelerator treatment couch, and the current pose includes the current position and the current angle;

[0037] S6.3: Based on the current pose and the coordinate system of the accelerator treatment couch, convert the conversion deviation into the pose relationship of the upper platform of the accelerator treatment couch with respect to the lower platform;

[0038] S6.4: Convert the pose relationship of the upper platform with respect to the lower platform into the position matrix corresponding to the target position of each of the actuators. Through the position matrix, the execution amount of the actuator is obtained. After the actuator moves the corresponding execution amount, the upper platform reaches the target pose.

[0039] Preferably, in step S6.1, the conversion deviation is ( , , , , , ); in step S6.3, the conversion of the conversion deviation into the pose relationship of the upper platform of the accelerator treatment couch relative to the lower platform includes:

[0040] Spatial translation transformation:

[0041] X-axis rotation transformation:

[0042] Y-axis rotation transformation:

[0043] Z-axis rotation transformation:

[0044] The matrix corresponding to the conversion deviation is:

[0045]

[0046] In the formula, represents the transformation of the tumor target area at the isocenter, represents the pose relationship of the lower platform coordinate system B-XYZ under the isocenter coordinate system O-XYZ. The accelerator treatment couch has only 4 degrees of freedom, including translation in the X, Y, and Z directions and rotation about the Z axis. Assuming its initial position is Lat, Lng, Vrt and the initial angle is Rtn, the following representation can be obtained:

[0047] = *

[0048] represents the initial pose transformation relationship of the upper platform P-XYZ coordinate system under the lower platform coordinate system B-XYZ. Assuming its initial position is , , , and the initial attitude is , , , then it can be represented as follows:

[0049]

[0050] Based on the above known conditions, the pose relationship after correcting the target area can be obtained:

[0051]

[0052]

[0053] It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system O-XYZ after correcting the target area; It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system B-XYZ after correcting the target area.

[0054] Preferably, the specific steps of step S6.4 are as follows:

[0055] S6.4.1: Establish the coordinates of the lower hinge point of the actuator relative to the lower platform coordinate system and the coordinates of the upper hinge point of the actuator relative to the upper platform coordinate system, respectively, as:

[0056]

[0057]

[0058] In the formula, It represents the coordinates of the lower hinge point of the actuator relative to the lower platform coordinate system; It represents The X-axis coordinate of the point in the lower platform coordinate system, n represents the number of actuators, Same as Similarly; It represents the starting point of the point on the lower platform;

[0059]

[0060]

[0061] In the formula, It represents the coordinates of the upper hinge point of the actuator relative to the upper platform coordinate system; It represents The X-axis coordinate of the point in the upper platform coordinate system, n represents the number of actuators, Same as Similarly; It represents the starting point of the point on the upper platform;

[0062] S6.4.2: Convert the upper hinge point in the upper platform coordinate system to the lower platform coordinate system to obtain the position matrix, specifically:

[0063]

[0064]

[0065] In the formula, the pose of the upper platform coordinate system in the lower platform coordinate system is , where x, y, and z are the displacement amounts along the coordinate axes, and rx, ry, and rz are the rotations about the x-axis, y-axis, and z-axis in sequence; c is cos the abbreviation of s is sin the abbreviation of; is expressed as the coordinates of the point in the lower platform coordinate system, that is, the position matrix; represents the initial position of the origin of the upper platform coordinate system P-XYZ in the lower platform B-XYZ coordinate system .

[0066] S6.4.3: Calculate the execution amount of the actuator according to the position matrix , specifically:

[0067]

[0068]

[0069]

[0070] In the formula, represents the vector of the i-th actuator from to , and is expressed as then represents the unit vector of the i-th actuator from to , for example ; is expressed as the length of the connecting rod of the i-th actuator.

[0071] A real-time target area tracking and dynamic regulation system, including:

[0072] A model establishment module, used to establish a phase prediction model and a two-stage model, and superimpose the phase prediction model and the two-stage model into a target area prediction model;

[0073] A data collection module, used to collect 4DCT reconstruction data and three-dimensional point cloud body surface contour data and input them into the model establishment module;

[0074] A real-time data processing module, used to extract the target area prediction model, used to receive the real-time monitored body surface contour data and respiratory signal and input them into the target area prediction model, used to convert the prediction result of the target area prediction model into the target pose of the accelerator treatment couch;

[0075] A data calculation and output module, used to convert the target pose into the target position of the actuator and output it to the accelerator treatment couch;

[0076] The accelerator treatment couch receives the target position of the execution mechanism and adjusts its own target pose.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the target area prediction model, the real-time target area position in the body can be obtained from the real-time input body surface data and breathing signals, which can respond to the dynamic changes of the target area in real time and achieve dynamic regulation. The target area prediction model is an artificial intelligence model based on 4DCT reconstruction data, and the predicted target area position has high accuracy. In addition, the target area position of the target area prediction model only needs to be converted into the target pose of the accelerator treatment couch, and then the accelerator treatment couch is allowed to perform pose conversion so that the ray is aligned with the target area. There are fewer restrictions on the accelerator treatment couch, and it can be achieved as long as the accelerator treatment couch has the ability to change the patient's pose, without relying on specific equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a flowchart of a real-time target area tracking and dynamic regulation method of the present invention;

[0079] Figure 2 It is a flowchart of step S6 of a real-time target area tracking and dynamic regulation method of the present invention;

[0080] Figure 3 It is a schematic structural diagram of the accelerator treatment couch of the present invention;

[0081] Figure 4 It is a schematic structural diagram of the upper platform, lower platform and execution mechanism of the accelerator treatment couch of the present invention;

[0082] Figure 5 It is a geometric schematic diagram of the execution amount of the execution mechanism of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0083] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for explaining the present invention, rather than for limiting the protection scope of the present invention.

[0084] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the components related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape and size of the components in actual implementation. The type, quantity and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0085] Embodiment 1

[0086] As Figure 1 shown in Embodiment 1 of a real-time target area tracking and dynamic regulation method, which includes the following steps:

[0087] S1: Obtain a phase prediction model based on 4DCT body surface reconstruction data and three-dimensional point cloud body surface contour data. The specific process is as follows;

[0088] S1.1: Obtain respiratory signals and corresponding image information within at least one respiratory cycle through 4DCT. The image information includes 4DCT three-dimensional point cloud contours and target area position information; divide the respiratory phase into N time phases on average according to the respiratory signals, and associate the corresponding 4DCT three-dimensional point cloud contours and target area position information for each time phase; the image information obtained by 4DCT is 4DCT reconstruction data, and 4DCT includes body surface reconstruction data (i.e., 4DCT three-dimensional point cloud contours) and 4DCT in-vivo target area reconstruction data (i.e., target area position information).

[0089] S1.2: Obtain the three-dimensional point cloud body surface contour of at least one respiratory cycle, and the three-dimensional point cloud body surface contour is obtained through a 3D camera;

[0090] S1.3: Align the three-dimensional point cloud body surface contour with the 4DCT three-dimensional point cloud contour to obtain the phase corresponding to the three-dimensional point cloud body surface contour and perform annotation; specifically, calculate the average vector distance between the point cloud data in the three-dimensional point cloud body surface contour and the point cloud data in the N 4DCT three-dimensional point cloud contours, and use the time phase corresponding to the 4DCT three-dimensional point cloud contour with the smallest average vector distance as the phase of the three-dimensional point cloud body surface contour.

[0091] S1.4: Establish a data set based on the data and phase annotation obtained in S1.1 - S1.3; in this embodiment, obtain patient data of multiple different ages, treatment sites, and genders, repeat steps S1.1 - S1.3 to obtain multiple groups of data, and merge the multiple groups of data to establish a data set.

[0092] S1.5: Input the data set into the first neural network model and train it to obtain the phase prediction model, and the phase prediction model outputs a predicted phase. In this embodiment, the first neural network includes a PointNet layer, a self-attention mechanism layer, and a fully connected layer, which are used to extract the features of the point cloud and perform classification.

[0093] S2: Obtain a two-stage model based on the result output by the phase prediction model and the 4DCT in-vivo target area reconstruction data. The specific steps are as follows:

[0094] The point cloud data of the 4DCT three-dimensional point cloud contour corresponding to the predicted phase and the point cloud data of the three-dimensional point cloud body surface contour obtained by the phase prediction model are used as input values and input into the second neural network model for training to obtain the two-stage model, and the two-stage model outputs the predicted target area position information.

[0095] In this embodiment, the second neural network model includes a DGCNN layer, a Transformer layer, and a Pointer layer. First, the DGCNN layer extracts high-dimensional feature vectors, and then the Transformer layer is used to extract global feature correlations. The correlation residual term obtained by the Transformer layer is input into the Pointer layer for Pointer Generation, and finally a feature matrix of the target area position is obtained. Among them, Pointer Generation is generated using the probability method. The loss function of the two-stage model is defined as the difference between the predicted target area position information and the target area position information in S1.1; the loss function has an L2 regularization term. The specific form of this loss function is:

[0096]

[0097] In the formula, represents the second loss function; represents the transpose matrix of the rotation matrix that transforms the three-dimensional point cloud body surface contour of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the rotation matrix that transforms the three-dimensional point cloud body surface contour of each breathing phase to the point cloud data corresponding to the corresponding true phase; represents the identity matrix; represents the translation matrix that transforms the three-dimensional point cloud body surface contour of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the translation matrix that transforms the body surface point cloud contour of each breathing phase to the point cloud data corresponding to the corresponding true phase, represents the regularization coefficient, represents the training parameter.

[0098] S3: Superimpose the phase prediction model and the two-stage model, and fuse them to obtain the target area prediction model;

[0099] S4: Real-time monitor the body surface contour data and breathing signal of the patient through the body surface optical guidance system;

[0100] S5: Input the body surface contour data and breathing signal into the target area prediction model, and the target area prediction model predicts the current position of the target area in the body;

[0101] S6: Calculate the target pose of the accelerator treatment couch according to the current position of the target area in the body, obtain the target position of the actuator of the accelerator treatment couch based on the target pose of the accelerator treatment couch, and drive the actuator to move to the target position so as to adjust the accelerator treatment couch to the target pose.

[0102] Working principle of this embodiment: The 4DCT mentioned in this embodiment, the full name is Four-Dimensional Computed Tomography, which adds a time dimension on the basis of traditional three-dimensional CT scanning and can capture and record the changes of tumors and surrounding tissues with respiratory movement. This is particularly important for the precise positioning of radiotherapy, because many tumors, especially those in the chest and abdomen, move with respiratory movement, which may lead to errors during treatment.

[0103] The detailed process of 4DCT usually includes the following steps:

[0104] Data acquisition: First, use a CT scanning device to continuously acquire images during the entire respiratory cycle of the patient. Use a respiratory monitoring device (such as optical abdominal tracking, abdominal pressure belt, etc.) to record the patient's respiratory cycle to synchronize the acquisition of images and respiratory signals to ensure that the images correspond to specific stages of the respiratory cycle. Image reconstruction: According to the respiratory signal, classify and reconstruct the acquired image data according to different respiratory phases to form an image sequence that changes over time. Data integration: Integrate the 3D images of all respiratory phases into a 4D data set, which can reflect the dynamic changes of the target area during the entire respiratory cycle. Therefore, the 4D data set includes respiratory signals, target area position information corresponding to the phases of the respiratory signals, and the 4DCT three-dimensional point cloud contour of the target area.

[0105] In this embodiment, in the dataset for training the phase prediction model, the three-dimensional point cloud body surface contour and the 4DCT three-dimensional point cloud contour are used as the inputs of the model, and the predicted phase corresponding to the three-dimensional point cloud body surface contour is output. Therefore, after obtaining the phase prediction model, the phase associated with the respiratory signal corresponding to the body surface data and the point cloud data of the in-vivo target area corresponding to the phase can be predicted based on the body surface data through this model. In the dataset for training the two-stage model, the point cloud data of the 4DCT three-dimensional point cloud contour corresponding to the predicted phase and the point cloud data of the three-dimensional point cloud body surface contour are used as the model inputs, and the target area position (i.e., the 4DCT in-vivo target area reconstruction data) information in the 4DCT reconstruction data is used as the model reference value. After obtaining the two-stage model, the target area position information can be predicted based on the predicted phase and the point cloud data of the three-dimensional point cloud body surface contour. After superimposing the phase prediction model and the two-stage model, the target area prediction model is obtained. When it is necessary to track the target area, only the real-time body surface data and the respiratory signal need to be input into the target area prediction model. The input data is calculated by the phase prediction model and the two-stage model, and finally the current position of the predicted in-vivo target area is output. After obtaining the current position of the in-vivo target area, the corresponding target pose of the accelerator treatment couch is obtained through conversion. The accelerator treatment couch enables the actuator of the accelerator treatment couch to reach the target position according to the target pose, thereby realizing the transformation of the pose. After the accelerator treatment couch is converted to the target pose, the ray irradiates the current position of the in-vivo target area. The in-vivo target area changes with breathing, and the real-time body surface data and the respiratory signal also change accordingly. The predicted position of the in-vivo target area by the target area prediction model also changes accordingly. Finally, the accelerator treatment couch can follow the position of the target area in real time, so that the ray can always irradiate only on the target area.

[0106] Advantages of this embodiment: Through the target area prediction model, the real-time position of the in-vivo target area can be obtained from the real-time input body surface data and respiratory signal, which can respond to the dynamic changes of the target area in real time and achieve dynamic regulation. The target area prediction model is based on artificial intelligence modeling using 4DCT reconstruction data, and the predicted target area position has high accuracy. In addition, the target area position of the target area prediction model only needs to be converted into the target pose of the accelerator treatment couch, so that the external accelerator treatment couch can perform pose conversion to align the ray with the target area, with less restrictions on the accelerator treatment couch. As long as the accelerator treatment couch can change the patient's pose, it can be realized without relying on specific equipment.

[0107] In this embodiment, the target area prediction model is established and trained by dividing it into a phase prediction model and a two-stage model, and then superimposed for use. According to different usage scenarios, the parameters of the directly corresponding phase prediction model and two-stage model can be adjusted, which is more convenient for adjustment, enabling the entire target area prediction model to better adapt to different users and further improving the prediction accuracy.

[0108] Example 2

[0109] Example 2 of a real-time target area tracking and dynamic regulation method, which is different from Example 1 in that before step S4, the target area prediction model is further optimized. Specifically:

[0110] Obtain data of the in-vivo target area through CBCT scanning, including the shape, size, position of the target area, and the correlation with body surface and in-vivo movements; the correlation between the body surface and in-vivo movements is extracted through convolutional operations of a convolutional neural network; input the data of the in-vivo target area into the phase prediction model for training through the transfer learning method.

[0111] The working principle of this example: Before putting the target area prediction model into use, first input the data obtained by CBCT scanning of the patient before treatment as training data into the phase prediction model in the target area prediction model, and adjust the phase prediction model again to be more suitable for the patient's own situation, so that the prediction result of the target area prediction model is more accurate.

[0112] The remaining features and technical effects of this implementation are the same as those of Example 1.

[0113] Example 3

[0114] Example 2 of a real-time target area tracking and dynamic regulation method, which is different from Example 1 in that step S6 is further defined. The specific steps are as Figure 2 shown, including:

[0115] S6.1: Calculate the conversion deviation corresponding to the target pose of the accelerator treatment couch when converting the current position of the in-vivo target area;

[0116] S6.2: Obtain the current pose of the accelerator treatment couch, and the current pose includes the current position and the current angle;

[0117] S6.3: Based on the current pose and the coordinate system of the accelerator treatment couch, convert the conversion deviation into the pose relationship of the upper platform of the accelerator treatment couch with respect to the lower platform;

[0118] S6.4: Convert the pose relationship of the upper platform with respect to the lower platform into the position matrix corresponding to the target position of each actuator. Through the position matrix, obtain the execution amount of the actuator. After the actuator moves the corresponding execution amount, the upper platform reaches the target pose.

[0119] In this example, the accelerator treatment couch is as Figure 3 and Figure 4 shown, where the upper platform 6 and the lower platform 4 are the six-dimensional couch 3 on the treatment couch, and the upper platform 6 and the lower platform 4 are connected by six actuators 5. The layout form of the actuator 5 is as Figure 4As shown, the specific forms of the upper platform 6, the lower platform 4, and the actuator 5 can be seen in "A Radiotherapy Bed" with the publication number CN108031016B.

[0120] In step S6.1, the conversion deviation is ( , , , , , ); in step S6.3, the conversion of the conversion deviation into the pose relationship of the upper platform of the accelerator treatment bed relative to the lower platform includes:

[0121] Spatial translation transformation:

[0122] X-axis rotation transformation:

[0123] Y-axis rotation transformation:

[0124] Z-axis rotation transformation:

[0125] The matrix corresponding to the conversion deviation is:

[0126]

[0127] In the formula, represents the transformation of the tumor target area at the isocenter; represents the pose relationship of the lower platform coordinate system B-XYZ under the isocenter coordinate system O-XYZ. The accelerator treatment bed has only 4 degrees of freedom, translation in the X, Y, and Z directions and rotation around the Z axis. Assuming its initial position is Lat, Lng, Vrt and the initial angle is Rtn, the following can be obtained:

[0128] = *

[0129] represents the initial pose transformation relationship of the upper platform P-XYZ coordinate system under the lower platform coordinate system B-XYZ. Assuming its initial position is , , , and the initial attitude is , , , then it can be expressed as follows:

[0130]

[0131] Based on the above known conditions, the pose relationship after correcting the target area can be obtained:

[0132]

[0133]

[0134] It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system O-XYZ after correcting the target area; It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system B-XYZ after correcting the target area.

[0135] The specific steps of step S6.4 are as follows:

[0136] S6.4.1: Establish the coordinates of the lower hinge point of the actuator with respect to the lower platform coordinate system and the coordinates of the upper hinge point of the actuator with respect to the upper platform coordinate system, respectively, as follows:

[0137]

[0138]

[0139] In the formula, It represents the coordinates of the lower hinge point of the actuator with respect to the lower platform coordinate system; It represents The X-axis coordinate of the point in the lower platform coordinate system, n represents the number of actuators, Similar to ; It represents the starting point of the point on the lower platform;

[0140]

[0141]

[0142] In the formula, It represents the coordinates of the upper hinge point of the actuator with respect to the upper platform coordinate system; It represents The X-axis coordinate of the point in the upper platform coordinate system, n represents the number of actuators, Similar to ; It represents the starting point of the point on the upper platform;

[0143] S6.4.2: Convert the upper hinge point in the upper platform coordinate system to the lower platform coordinate system to obtain the position matrix, specifically:

[0144]

[0145]

[0146] In the formula, the pose of the upper platform coordinate system in the lower platform coordinate system is , where x, y, and z are the displacement amounts along the coordinate axes, and rx, ry, and rz are the rotations about the x-axis, y-axis, and z-axis in sequence; c is cos an abbreviation of s is sin an abbreviation of; is expressed as the coordinates of point in the lower platform coordinate system, that is, the position matrix; represents the initial position of the origin of the upper platform coordinate system P-XYZ in the lower platform B-XYZ coordinate system .

[0147] S6.4.3: Calculate the execution amount of the actuator according to the position matrix , specifically:

[0148]

[0149]

[0150]

[0151] In the formula, represents the vector of the i-th actuator from to , while is expressed as then represents the unit vector of the i-th actuator from to , for example ; is expressed as the length of the connecting rod of the i-th actuator.

[0152] Specifically, the kinematic decomposition of the execution amount is as shown in Figure 5 , is the coordinate of the lower hinge point of the actuator relative to the lower platform coordinate system, representing the starting point in the figure; is the coordinate of the lower hinge point of the actuator relative to the lower platform coordinate system, representing the ending point in the figure, that is, the target position, represents the length of the i-th connecting rod; represents the movement distance of the i-th actuator, that is, the execution amount; represents the angle between the connecting rod and the movement direction of the actuator.

[0153] Working principle of this embodiment: The patient lies on the upper platform. In order to ensure that the ray can always irradiate the target area of the patient, it is necessary to adjust the pose of the patient. However, during the treatment, the patient will not actively change the pose. Therefore, in fact, what needs to be changed is the pose of the upper platform. The change in the pose of the upper platform is achieved through the actions of the actuators. Different actuators execute corresponding execution amounts, and the upper platform will undergo corresponding changes. Therefore, by converting the predicted target area position into the target pose of the patient, that is, the conversion deviation between the target poses of the upper platform, this conversion deviation is converted into the pose relationship and coordinates of the upper platform coordinate system with respect to the lower platform coordinate system after calculation. Then, the inverse solution of the actuators is performed based on this pose relationship and coordinates to obtain the execution amounts of each actuator. After each actuator completes its own execution amount, it reaches its own target position, and the pose of the upper platform also changes to the target pose that keeps the target area of the patient within the ray irradiation range.

[0154] Embodiment 3

[0155] An embodiment of a real-time target area tracking and dynamic regulation system, which can be used to implement Embodiments 1 - 3, includes:

[0156] A model establishment module, used to establish a phase prediction model and a two-stage model, and superimpose the phase prediction model and the two-stage model into a target area prediction model;

[0157] A data collection module, used to collect 4DCT reconstruction data and three-dimensional point cloud body surface contour data and input them into the model establishment module;

[0158] A real-time data processing module, used to extract the target area prediction model, receive the real-time monitored body surface contour data and respiratory signal and input them into the target area prediction model, and convert the prediction result of the target area prediction model into the target pose of the accelerator treatment couch;

[0159] A data calculation and output module, used to convert the target pose into the target position of the actuator and output it to the accelerator treatment couch;

[0160] The accelerator treatment couch receives the target position of the actuator and adjusts itself to the target pose.

[0161] Working principle or workflow of this embodiment: In the model establishment module, the frameworks of the phase prediction model and the two-stage model are built and saved respectively. After the data collection module inputs the collected data into the model establishment module, the model establishment module trains the phase prediction model and the two-stage model in sequence, and superimposes the trained phase prediction model and two-stage model into a target area prediction model block and stores it. When in actual use, the real-time data processing module extracts the target area prediction model stored in the model establishment module, and inputs the real-time data into the trained target area prediction model. The target area prediction model outputs the prediction result in real time. The prediction result is the target area position in the patient's body. The real-time data processing module converts the target area position in the patient's body into the target pose of the accelerator treatment couch. After receiving the target pose of the accelerator treatment couch, the data calculation and output module calculates the target position of the actuator in combination with the current pose and coordinate system. After receiving the target position, the actuator of the accelerator treatment couch moves the corresponding execution amount respectively and reaches the target position. The upper platform of the accelerator treatment couch reaches the target pose, and the patient lies on the upper platform, that is, the patient is in the target pose, and the target area position is within the irradiation range of the ray.

[0162] The accelerator treatment couch of this embodiment is an existing treatment couch, and its specific structure can be seen in Figure 3 and Figure 4 . It includes an accelerator 1, a treatment couch 2 located below the accelerator, a six-degree-of-freedom couch 3 is arranged on the treatment couch, and the six-degree-of-freedom couch includes a lower platform 4, an actuator 5 installed on the lower platform, and an upper platform 6 connected to the moving end of the actuator 5.

[0163] The above embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention.

Claims

1. A real-time target area tracking and dynamic regulation method, characterized in that: The method includes the following steps: S1: Obtain a phase prediction model based on 4DCT body surface reconstruction data and three-dimensional point cloud body surface contour data; the 4DCT body surface reconstruction data includes a respiration signal and a 4DCT three-dimensional point cloud contour corresponding to the respiration signal; S2: Obtain a two-stage model according to the result output by the phase prediction model and 4DCT in-vivo target area reconstruction data; the 4DCT in-vivo target area reconstruction data includes target area position information corresponding to the respiration signal; S3: Superimpose the phase prediction model and the two-stage model, and fuse them to obtain a target area prediction model; S4: Obtain real-time body surface contour data and respiration signal; S5: Input the body surface contour data and the respiration signal into the target area prediction model, and the target area prediction model predicts the current position of the in-vivo target area; S6: Calculate the target pose of the accelerator treatment couch according to the current position of the in-vivo target area, obtain the target position of the actuator of the accelerator treatment couch based on the target pose of the accelerator treatment couch, and drive the actuator to move to the target position so as to adjust the accelerator treatment couch to the target pose.

2. The real-time target area tracking and dynamic regulation method according to claim 1, wherein, In step S1, the specific process is as follows: S1.1: Obtain a respiration signal and corresponding image information within at least one respiration cycle through 4DCT, where the image information includes a 4DCT three-dimensional point cloud contour and target area position information; divide the respiration phase into N time phases on average according to the respiration signal, and associate the 4DCT three-dimensional point cloud contour and target area position information corresponding to each time phase; S1.2: Obtain the three-dimensional point cloud body surface contour of at least one respiration cycle; S1.3: Align the three-dimensional point cloud body surface contour with the 4DCT three-dimensional point cloud contour, obtain the phase corresponding to the three-dimensional point cloud body surface contour and perform annotation; S1.4: Establish a data set based on the data and phase annotation obtained in S1.1 - S1.3; S1.5: Input the data set into the first neural network model for training to obtain the phase prediction model, and the phase prediction model outputs a predicted phase.

3. The real-time target area tracking and dynamic regulation method according to claim 2, characterized in that, In step S1.3, calculate the average vector distance between the point cloud data in the three-dimensional point cloud body surface contour and the point cloud data in the N 4DCT three-dimensional point cloud contours, and use the time phase corresponding to the 4DCT three-dimensional point cloud contour with the smallest average vector distance as the phase of the three-dimensional point cloud body surface contour.

4. A real-time target area tracking and dynamic regulation method according to claim 2, characterized in that, The first neural network includes a PointNet layer, a self-attention mechanism layer, and a fully connected layer, and is used to extract the features of the point cloud and perform classification.

5. The real-time target area tracking and dynamic regulation method according to claim 2, characterized in that In step S2, the point cloud data of the 4DCT three-dimensional point cloud contour corresponding to the predicted phase obtained by the phase prediction model and the point cloud data of the three-dimensional point cloud body surface contour are used as input values and input into the second neural network model for training to obtain the two-stage model, and the two-stage model outputs predicted target area position information.

6. The real-time target area tracking and dynamic regulation method according to claim 5, wherein The second neural network model includes a DGCNN layer, a Transformer layer, and a Pointer layer.

7. A real-time target area tracking and dynamic regulation method according to claim 5, characterized in that, The loss function of the two-stage model is defined as the difference between the predicted target area position information and the target area position information in S1.1; the loss function has an L2 regularization term.

8. A real-time target area tracking and dynamic regulation method according to claim 7, characterized in that The loss function of the two-stage model is specifically: Wherein, represents the second loss function; represents the transposed matrix of the rotation matrix for transforming the three-dimensional point cloud body surface contour number of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the rotation matrix for transforming the three-dimensional point cloud body surface contour of each breathing phase to the point cloud data corresponding to the corresponding true phase; represents the identity matrix; represents the translation matrix for transforming the three-dimensional point cloud body surface contour of each breathing phase to the predicted point cloud data corresponding to the corresponding final predicted phase; represents the translation matrix for transforming the body surface point cloud contour of each breathing phase to the point cloud data corresponding to the corresponding true phase, represents the regularization coefficient, represents the training parameter.

9. A real-time target area tracking and dynamic regulation method according to any one of claims 1-8, characterized in that, Before step S4, optimize the target area prediction model. The specific steps are as follows: Obtain the data of the target area in the body through CBCT scanning as training data and input it into the phase prediction model for training to adjust the parameters of the phase prediction model.

10. A real-time target area tracking and dynamic regulation method according to claim 9, characterized in that, The data of the target area in the body obtained through CBCT scanning includes the shape, size, position of the target area, and the correlation with the body surface and in-body movement; The correlation between the body surface and in-body movement is extracted through convolution operations of the convolutional neural network; Input the data of the target area in the body into the phase prediction model for training through the transfer learning method.

11. A real-time target area tracking and dynamic regulation method according to any one of claims 1-8, characterized in that, In step S6, the specific steps include: S6.1: Calculate the conversion deviation corresponding to the conversion of the current position of the target area in the body to the target pose of the accelerator treatment couch. S6.2: Obtain the current pose of the accelerator treatment couch, where the current pose includes the current position and the current angle; S6.3: Based on the current pose and the coordinate system of the accelerator treatment couch, convert the conversion deviation into the pose relationship of the upper platform of the accelerator treatment couch relative to the lower platform; S6.4: Convert the pose relationship of the upper platform relative to the lower platform into the position matrix corresponding to the target position of each actuator. Obtain the execution amount of the actuator through the position matrix. After the actuator moves the corresponding execution amount, the upper platform reaches the target pose.

12. A real-time target area tracking and dynamic regulation method according to claim 11, characterized in that In step S6.1, the conversion deviation is ( , , , , , ); In step S6.3, the conversion of the conversion deviation into the pose relationship of the upper platform of the accelerator treatment couch with respect to the lower platform includes: Spatial translation transformation: X-axis rotation transformation: Y-axis rotation transformation: Z-axis rotation transformation: The matrix corresponding to the conversion deviation is: In the formula, represents the transformation of the tumor target area at the isocenter; represents the pose relationship of the lower platform coordinate system B-XYZ under the isocenter coordinate system O-XYZ; x, y, and z are the displacement amounts along the coordinate axes, and rx, ry, and rz are the rotations around the x-axis, y-axis, and z-axis in sequence; c is cos an abbreviation of s is sin an abbreviation of; The accelerator treatment couch has only 4 degrees of freedom, namely translation in the three directions of XYZ and rotation around the Z-axis. Assuming its initial position is Lat, Lng, Vrt and the initial angle is Rtn, the following representation can be obtained: = * It represents the initial pose transformation relationship of the upper platform P-XYZ coordinate system under the lower platform coordinate system B-XYZ. Assuming its initial position is , , , and the initial attitude is , , , then it is represented as follows: Based on the above known conditions, the corrected pose relationship of the target area can be obtained: * It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system O-XYZ after correcting the target area; It represents the pose relationship of the P-XYZ coordinate system with respect to the isocenter coordinate system B-XYZ after correcting the target area.

13. A real-time target area tracking and dynamic regulation method according to claim 12, characterized in that, The specific steps in step S6.4 are: S6.4.1: Establish the coordinates of the lower hinge point of the actuator relative to the coordinate system of the lower platform and the coordinates of the upper hinge point of the actuator relative to the coordinate system of the upper platform, which are respectively: Wherein, represents the coordinates of the lower hinge point of the actuator relative to the lower platform coordinate system; represents the X-axis coordinate of point in the lower platform coordinate system, and n represents the number of actuators, is the same as ; represents the starting point of point on the lower platform ; In the formula, represents the coordinates of the upper hinge point of the actuator relative to the upper platform coordinate system; represents the X-axis coordinate of point in the upper platform coordinate system, n represents the number of actuators, is the same as by the same token; represents the starting point of point on the upper platform ; S6.4.2: Convert the upper hinge point in the upper platform coordinate system to the lower platform coordinate system to obtain the position matrix, specifically: wherein, the pose of the upper platform coordinate system in the lower platform coordinate system is , where x, y, and z are the displacement amounts along the coordinate axes, and rx, ry, and rz are the rotations about the x-axis, y-axis, and z-axis in sequence; c is cos an abbreviation of s is sin an abbreviation of; is expressed as the coordinates of the point in the lower platform coordinate system, that is, the position matrix; represents the initial position of the origin of the upper platform coordinate system P-XYZ in the lower platform B-XYZ coordinate system ; S6.4.3: Calculate the execution amount of the actuator according to the position matrix , specifically as follows: In the formula, represents the vector of the i-th actuator from to , while represents the unit vector of the i-th actuator from to ; represents the length of the connecting rod of the i-th actuator.

14. A real-time target area tracking and dynamic regulation method according to any one of claims 1-8, characterized in that, In step S4, the body surface contour data and respiratory signal of the patient are monitored in real time through the body surface optical guidance system.

15. A system for real-time target area tracking and dynamic regulation, characterized in that, It includes: The model establishment module is used to establish the phase prediction model and the two-stage model, and superimpose the phase prediction model and the two-stage model into the target area prediction model; The data collection module is used to collect 4DCT reconstruction data and three-dimensional point cloud body surface contour data and input them into the model establishment module; The real-time data processing module is used to extract the target area prediction model, receive the body surface contour data and respiratory signal monitored in real time and input them into the target area prediction model, and convert the prediction result of the target area prediction model into the target pose of the accelerator treatment couch; The data calculation and output module is used to convert the target pose into the target position of the actuator and output it to the accelerator treatment couch; The accelerator treatment couch receives the target position of the actuator and adjusts its own target pose.

Images