Method and equipment for predicting tumor position under respiratory movement based on LSTM (Long Short Term Memory)

Abstract

The invention discloses a method and equipment for predicting a tumor position under respiratory movement based on LSTM (Long Short Term Memory), and the method comprises the steps: collecting a 4DCT (4 Discrete Cosine Transform) image sequence of a tumor part of a plurality of patients under free respiration; based on the acquired 4DCT image sequence, extracting a movement track of a tumor centroid point position along with time change; sequentially inputting all tumor motion data points in each fixed time window on the motion trail into a CNN-Attention-LSTM model for offline training, wherein the CNN-Attention-LSTM model is used for predicting the tumor position at the next time; and in a radiotherapy stage, performing online real-time updating on the trained CNN-Attention-LSTM model by using the latest respiratory movement data of the current patient, and predicting the tumor position at the next time. The real-time position of the tumor is provided for the radiotherapy robot, the time delay of the whole system is compensated, and the tumor position tracking is accurately realized in real time.

Description

Technical Field

[0001] The present invention belongs to the technical field of tumor radiotherapy, and specifically relates to a method and device for predicting tumor position under respiratory motion based on LSTM. Background Art

[0002] During radiotherapy, breathing movements can cause displacement of chest and abdominal tumors, causing them to move periodically with the breathing frequency. The amplitude of chest and abdominal tumor movement caused by breathing is about 10 to 20 mm, and the amplitude of movement of the lower lobe of the lung can even reach 30 mm. This movement may cause the tumor to move outside the treatment radiation area, thus affecting the treatment effect.

[0003] Image Guided Radiation Therapy (IGRT) is a new tumor radiotherapy technology that has gradually developed in the past decade. It uses advanced imaging equipment and image processing methods to accurately locate the patient's target area before treatment and conduct real-time positioning and tracking during treatment, thereby achieving precise radiotherapy of tumors and reducing damage to normal tissues and key organs around the tumor.

[0004] The core technology of real-time positioning and tracking of target area respiratory motion includes two parts:

[0005] (1) Predict the real-time movement of the target area in the body through the correlation between the position of the target area in the body and the respiratory movement of the surface of the body. The respiratory signal on the surface of the body has a good correlation with the movement of the chest and abdominal tumor in the body. The tumor position can be estimated through the correlation model between the surface marker and the tumor in the body.

[0006] (2) The delayed displacement of the tumor caused by respiratory movement is inferred and compensated through the time series signal prediction algorithm. There is a system delay of tens to hundreds of milliseconds between the acquisition of image information and the adjustment of the radiation field. The existence of the radiotherapy robot system delay causes the phase of the radiation field beam to lag behind the phase of tumor movement, affecting the accuracy of radiotherapy.

[0007] Therefore, in precise tumor radiotherapy, it is necessary to predict tumor respiratory motion and take into account system delays for corresponding compensation. System delays include the time taken by factors such as image acquisition, image processing, communication delays, control system processing, motor inductance, and mechanical damping.

[0008] The complexity and uncertainty of respiratory motion make it difficult to predict the location of tumors. The respiratory motion of the human body varies from person to person, and for the same patient, the period, amplitude, and baseline position of the respiratory motion also gradually change and drift during the treatment process.

[0009] In clinical practice, metal markers or other markers are implanted in or near the tumor to track the movement of the tumor. For example, the CyberKnife radiotherapy system can measure the movement of markers through X-rays and electromagnetic positioning devices, which is highly correlated with tumor movement. However, this method has certain limitations, such as the need for puncture to implant markers.

[0010] Since breathing is an irregular and difficult-to-predict physiological movement, and radiotherapy is usually divided into 20-30 sessions, the patient's breathing pattern may be different during each treatment. Therefore, a real-time, online, non-invasive and radiation-free tumor position prediction method is urgently needed during radiotherapy. Summary of the invention

[0011] In order to solve the above technical problems, the present invention proposes a method and device for predicting tumor position under respiratory motion based on LSTM.

[0012] In order to achieve the above object, the technical solution of the present invention is as follows:

[0013] In a first aspect, the present invention discloses a method for predicting tumor position under respiratory motion based on LSTM, comprising:

[0014] Step S1: acquiring 4DCT image sequences of the tumor site of several patients under free breathing;

[0015] Step S2: based on the acquired 4DCT image sequence, extract the motion trajectory of the tumor centroid position changing with time;

[0016] Step S3: All tumor motion data points within each fixed time window on the motion trajectory are sequentially input into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time.

[0017] Step S4: During the radiotherapy stage, the trained CNN-Attention-LSTM model is updated online in real time using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

[0018] Based on the above technical solution, the following improvements can be made:

[0019] As a preferred solution, step S2 includes:

[0020] Step S2.1: extracting the tumor target boundary contour from the 4DCT image using a segmentation algorithm;

[0021] Step S2.2: Calculate the pixel position of the centroid of the tumor in the image based on the extracted target boundary;

[0022] Step S2.3: repeating steps S2.1 to S2.2 until the tumor centroid positions of all 4DCT image sequences are extracted;

[0023] Step S2.4: After the extraction is completed, the motion trajectory curve of the tumor centroid position changing with time is calculated according to the pixel resolution and sampling frequency of the CT;

[0024] Step S2.5: performing normalization and / or smoothing filtering processing on the obtained motion trajectory curve.

[0025] As a preferred solution, the CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module and an LSTM module;

[0026] The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory;

[0027] The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element and fused as the input of the LSTM module.

[0028] The LSTM module is used to predict the tumor location at the next time.

[0029] As a preferred solution, the importance of the high-dimensional features output by the CNN module is proportional to the value of the output result of the corresponding attention mechanism module.

[0030] In addition, in a second aspect, the present invention further discloses a device for predicting tumor position under respiratory motion based on LSTM, comprising:

[0031] An acquisition module is used to acquire 4DCT image sequences of the tumor site of several patients under free breathing;

[0032] An extraction module is used to extract the motion trajectory of the tumor centroid position changing with time based on the acquired 4DCT image sequence;

[0033] The offline training module is used to sequentially input all tumor motion data points within each fixed time window on the motion trajectory into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time.

[0034] The online update module is used to update the trained CNN-Attention-LSTM model online in real time during the radiotherapy stage using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

[0035] As a preferred solution, the extraction module includes:

[0036] A segmentation unit, used for extracting a tumor target boundary contour from a 4DCT image using a segmentation algorithm;

[0037] A centroid calculation unit, used to calculate the centroid pixel position of the tumor in the image based on the extracted target boundary;

[0038] A repeated execution unit, used for repeatedly executing the methods in the segmentation unit and the centroid calculation unit, until the centroid point positions of the tumors in all 4DCT image sequences are extracted;

[0039] A curve calculation unit is used to calculate the motion trajectory curve of the tumor centroid position changing with time according to the pixel resolution and sampling frequency of the CT after the extraction is completed;

[0040] The curve processing unit is used to perform normalization and / or smoothing filtering on the obtained motion trajectory curve.

[0041] As a preferred solution, the CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module and an LSTM module;

[0042] The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory;

[0043] The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element and fused as the input of the LSTM module.

[0044] The LSTM module is used to predict the tumor location at the next time.

[0045] As a preferred solution, the importance of the high-dimensional features output by the CNN module is proportional to the value of the output result of the corresponding attention mechanism module.

[0046] The present invention discloses a method and device for predicting tumor position under respiratory motion based on LSTM, which provides the real-time position of the tumor for a radiotherapy robot, compensates for the overall system delay, guides the rays to be more accurately concentrated on the target area, and achieves real-time and accurate tumor position tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 This is a flow chart of a method for predicting tumor position under respiratory motion provided by an embodiment of the present invention.

[0049] Figure 2 A block diagram of a CNN module and an attention mechanism module provided in an embodiment of the present invention.

[0050] Figure 3 A block diagram of an LSTM module provided in an embodiment of the present invention.

[0051] Figure 4 A training block diagram of the CNN-Attention-LSTM model provided in an embodiment of the present invention.

[0052] Figure 5 A comparison chart of the prediction results of the CNN-Attention-LSTM model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The expression of “comprising” an element is an “open” expression, which merely means that a corresponding component or step exists, and should not be interpreted as excluding additional components or steps.

[0056] In order to achieve the purpose of the present invention, in some embodiments of a method and device for predicting tumor position under respiratory motion based on LSTM, as Figure 1 As shown, the method for predicting tumor position under respiratory motion includes:

[0057] Step S1: acquiring 4DCT image sequences of the tumor site of several patients under free breathing;

[0058] Step S2: based on the acquired 4DCT image sequence, extract the motion trajectory of the tumor centroid position changing with time;

[0059] Step S3: All tumor motion data points within each fixed time window on the motion trajectory are sequentially input into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time.

[0060] Step S4: During the radiotherapy stage, the trained CNN-Attention-LSTM model is updated online in real time using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

[0061] The present invention helps to more accurately predict tumor movement and compensate for the overall delay of the robotic system, thereby improving the efficacy of tumor treatment.

[0062] In order to understand the present invention in detail, each step is described below.

[0063] Step S1 acquires a 4DCT image sequence of each patient's tumor site. The CT sampling frequency is 5 Hz. During the acquisition phase, the patient breathes freely.

[0064] Step S2 extracts the motion trajectory of the tumor centroid position changing with time from the 4DCT image sequence, including:

[0065] Step S2.1: Using RGB threshold segmentation to extract the tumor target boundary contour from the 4DCT image using the watershed segmentation algorithm in channel B;

[0066] Step S2.2: Calculate the pixel position of the centroid of the tumor in the image based on the extracted target boundary;

[0067] Step S2.3: repeating steps S2.1 to S2.2 until the tumor centroid positions of all 4DCT image sequences are extracted;

[0068] Step S2.4: After the extraction is completed, the motion trajectory curve of the tumor centroid position changing with time is calculated according to the pixel resolution and sampling frequency of the CT;

[0069] Step S2.5: performing normalization and / or smoothing filtering processing on the obtained motion trajectory curve.

[0070] Step S2.5 is specifically as follows: use maximum value normalization to normalize the motion trajectory curve to between -1 and +1, and normalization facilitates subsequent neural network training. Then, use a sliding window with a window size of 3 to remove outliers and use mean smoothing filtering on them. A sliding window size of 3 means that the current point is taken as the center, and the average of the three adjacent values ​​is calculated as the value of the current point, which can smooth the curve.

[0071] Furthermore, the above CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module and an LSTM module;

[0072] The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory;

[0073] The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element and fused as the input of the LSTM module.

[0074] The LSTM module is used to predict the tumor location at the next time.

[0075] like Figure 2 As shown in the figure, the CNN module is constructed by stacking multiple layers of one-dimensional networks, each of which contains a convolutional layer, a batch normalization layer, and a nonlinear layer. In addition, the pooling layer is used to achieve sampling aggregation, and the stacking of convolutional layers is used to build a hierarchical structure to gradually extract more abstract features. The output of this module is a feature sequence of m length n, whose size can be expressed as (n×m).

[0076] The attention mechanism module consists of two parts: feature aggregation and scale recovery.

[0077] In the feature aggregation part, key features are extracted from sub-sequences across scales by stacking multiple layers of convolution and pooling layers, while the last layer uses a 1×1 convolution kernel to mine linear relationships.

[0078] The scale recovery part restores the key features to a size of (n×m), which is consistent with the feature size output by the CNN module, and limits the value to between 0 and 1 through the sigmoid function.

[0079] The contextual features extracted in this way are used as the saliency feature map of the basic features. The input of each parallel module is designed with an adjustable overlap. By adjusting the overlapping step length of the input, the model can better adapt to data with different patterns and further expand the length of the overall input of the joint model at one time to capture more accurate long-term features. The non-dense overlapping design of the input between parallel modules helps to avoid overfitting.

[0080] The high-dimensional features output by the CNN module are multiplied element by element with the significant features output by the corresponding attention mechanism module. Specifically, the higher the importance of the output features of the CNN module, the closer the output of the corresponding attention mechanism module is to 1; conversely, the lower the importance of the output features of the CNN module, the closer the output of the corresponding attention mechanism module is to 0. The value reflects the importance of the feature, and the result of the identification and fusion of the important features is used as the input of a node of the LSTM module, and the final prediction result is obtained by using LSTM for sequence modeling.

[0081] like Figure 3 As shown, the storage unit state c is introduced into the LSTM module t , increasing the gradient stability during back propagation and the information flow during forward reasoning.t Used to save or discard c t-1 The historical information in, input gate i t Used to filter new storage units The information in the previous gates and memory is then combined to construct the final memory cell state c t The last storage unit uses output gate o t Filter to retain / discard some information and build a hidden state h t Mathematically, given a time step t, the LSTM module is described as follows:

[0082] Forget Gate:

[0083] Input Gate:

[0084] New storage unit status:

[0085] Final storage unit status:

[0086] Output gate: o t =σ(W o x t +U o h t-1 +b o );

[0087] Hidden state: h t =o t ⊙tanh(c t ).

[0088] in:

[0089] f t is the retention degree value;

[0090] c t is the current state;

[0091] is an intermediate state;

[0092] o t is the output degree value;

[0093] h t is the output at the current moment;

[0094] x t is the input at the current moment;

[0095] i t Add the degree value of the current state;

[0096] tanh() is the hyperbolic tangent function;

[0097] b, W and U represent the bias, input window weight, and recursive weight learned during the optimization process;

[0098] ⊙ represents element-wise multiplication between matrices / vectors.

[0099] The gate function uses the Sigmoid function σ(), and the state generation uses the hyperbolic tangent function tanh(). For each time step, the hidden state of one LSTM layer is used as the input of the next LSTM layer.

[0100] For the last hidden layer of the LSTM, the hidden state at the last time step is fed into a fully connected layer to build the predicted output window:

[0101] Prediction output:

[0102] Where: W FC and b FC Represents the weight matrix and bias vector of the fully connected layer.

[0103] The present invention adds a CNN+Attention fusion module before the LSTM module to increase the short-term local prediction capability.

[0104] The single LSTM module constitutes the LSTM prediction sequence, which corresponds to the tumor motion signal.

[0105] In the present invention, a stateless LSTM is implemented, i.e., during the optimization process, the hidden state and the cell state are cleared after each batch of data.

[0106] like Figure 4 As shown, in step S3, all tumor motion data points in each fixed time window on the motion trajectory are sequentially input into the CNN-Attention-LSTM model for offline training.

[0107] The LSTM model uses motion signal xi as input data and calculates the MSE loss function between the real output and the predicted output to update the model weights. The above xi is all tumor motion data points within a fixed time window. Assuming that the given time window is 20s, according to the sampling frequency of CT, xi has a total of about 100 data points representing the respiratory motion signal. In the offline training stage, the data within the sliding window with a time length of 20s is used for training.

[0108] The training, validation, and test sets were split in a 3:1:1 ratio, and the optimal hyperparameters were found using a grid search. Offline LSTM optimization was performed following the typical machine learning train / validation / test segmentation. The model was iteratively optimized on the training set for 600 epochs while monitoring the training and validation losses. If the validation loss improved, the model weights were saved to disk. If the validation loss did not improve within 100 epochs, the iteration was stopped.

[0109] In step S4, during the treatment process, the latest respiratory movement data of the current patient collected in real time (data within a time window of 20 seconds in length is used as the input of the model) is used to update the trained CNN-Attention-LSTM model online in real time to further optimize the model.

[0110] During radiotherapy robot treatment, the CNN-Attention-LSTM model is updated online in real time to adapt to the current patient's respiratory movement pattern during treatment.

[0111] First, load the CNN-Attention-LSTM model weights that were previously trained offline based on hundreds of patients’ respiratory motion data, then update the patient’s respiratory motion data for the last 20 seconds in the sliding window in a first-in-first-out manner, and then update 10 epochs, which takes a total of 150ms. Since the sampling frequency of CT is 200ms, the online optimization of the CNN-Attention-LSTM model can be completed within the sampling interval of CT. The predicted results are as follows: Figure 5 shown.

[0112] The present invention has the following beneficial effects:

[0113] First, the present invention combines offline and online methods to update the weights of the CNN-Attention-LSTM model in real time, making it more accurate and more adaptable to changes in breathing.

[0114] Second, the model of the present invention is small and fast, with a prediction time of only 5ms, so that the model weights can be updated online in real time with the patient's most recent respiratory data.

[0115] Third, the present invention combines the short sequence feature abstraction capability of the CNN module to extract high-dimensional features, and then uses LSTM to integrate the short sequence high-dimensional features for time series prediction, which is suitable for processing tumor respiratory motion time series data with local correlation.

[0116] Fourth, the attention mechanism module of the present invention effectively addresses the deficiency that the model cannot distinguish the differences in the importance of temporal features by enhancing the influence of important temporal features in the model and reducing the influence of unimportant features in the model.

[0117] The embodiment of the present invention further discloses a device for predicting tumor position under respiratory motion based on LSTM, comprising:

[0118] An acquisition module is used to acquire 4DCT image sequences of the tumor site of several patients under free breathing;

[0119] An extraction module is used to extract the motion trajectory of the tumor centroid position changing with time based on the acquired 4DCT image sequence;

[0120] The offline training module is used to sequentially input all tumor motion data points within each fixed time window on the motion trajectory into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time.

[0121] The online update module is used to update the trained CNN-Attention-LSTM model online in real time during the radiotherapy stage using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

[0122] Further, the extraction module includes:

[0123] A segmentation unit, used for extracting a tumor target boundary contour from a 4DCT image using a segmentation algorithm;

[0124] A centroid calculation unit, used to calculate the centroid pixel position of the tumor in the image based on the extracted target boundary;

[0125] A repeated execution unit, used for repeatedly executing the methods in the segmentation unit and the centroid calculation unit, until the centroid point positions of the tumors in all 4DCT image sequences are extracted;

[0126] A curve calculation unit is used to calculate the motion trajectory curve of the tumor centroid position changing with time according to the pixel resolution and sampling frequency of the CT after the extraction is completed;

[0127] The curve processing unit is used to perform normalization and / or smoothing filtering on the obtained motion trajectory curve.

[0128] Furthermore, the CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module, and an LSTM module;

[0129] The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory;

[0130] The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element and fused as the input of the LSTM module.

[0131] The LSTM module is used to predict the tumor location at the next time.

[0132] Furthermore, the importance of the high-dimensional features output by the CNN module is proportional to the value of the output result of the corresponding attention mechanism module.

[0133] In this embodiment, the specific content of the LSTM-based tumor position prediction device under respiratory motion is similar to the content of the LSTM-based tumor position prediction method under respiratory motion disclosed in the above embodiment, and will not be repeated here.

[0134] The present invention discloses a method and device for predicting tumor position under respiratory motion based on LSTM, which provides the real-time position of the tumor for a radiotherapy robot, compensates for the overall system delay, guides the rays to be more accurately concentrated on the target area, and achieves real-time and accurate tumor position tracking.

[0135] It should be understood that the various techniques described herein can be implemented in combination with hardware or software, or a combination thereof. Thus, the method and apparatus of the present invention, or some aspects or portions of the method and apparatus of the present invention, can take the form of program codes (i.e., instructions) embedded in a tangible medium, such as a floppy disk, a CD-ROM, a hard disk drive, or any other machine-readable storage medium, wherein when the program is loaded into a machine such as a computer and executed by the machine, the machine becomes a device for practicing the present invention.

[0136] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for illustrating the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which shall fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The LSTM-based tumor location prediction method under respiratory motion is characterized by: include: Step S1: acquiring 4DCT image sequences of the tumor site of several patients under free breathing; Step S2: based on the acquired 4DCT image sequence, extract the motion trajectory of the tumor centroid position changing with time; Step S3: All tumor motion data points within each fixed time window on the motion trajectory are sequentially input into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time. Step S4: During the radiotherapy stage, the trained CNN-Attention-LSTM model is updated online in real time using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

2. The method for predicting tumor position under respiratory motion according to claim 1, characterized in that: The step S2 comprises: Step S2.1: extracting the tumor target boundary contour from the 4DCT image using a segmentation algorithm; Step S2.2: Calculate the pixel position of the centroid of the tumor in the image based on the extracted target boundary; Step S2.3: repeating steps S2.1 to S2.2 until the tumor centroid positions of all 4DCT image sequences are extracted; Step S2.4: After the extraction is completed, the motion trajectory curve of the tumor centroid position changing with time is calculated according to the pixel resolution and sampling frequency of the CT; Step S2.5: performing normalization and / or smoothing filtering processing on the obtained motion trajectory curve.

3. The method for predicting tumor position under respiratory motion according to claim 1, characterized in that: The CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module and an LSTM module; The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory; The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element, and the fused features are used as the input of the LSTM module; The LSTM module is used to predict the tumor location at the next time.

4. The method for predicting tumor position under respiratory motion according to claim 3, characterized in that: The importance of the high-dimensional features output by the CNN module is proportional to the value of the output result of the corresponding attention mechanism module.

5. A device for predicting tumor position under respiratory motion based on LSTM, characterized in that: include: An acquisition module is used to acquire 4DCT image sequences of the tumor site of several patients under free breathing; An extraction module is used to extract the motion trajectory of the tumor centroid position changing with time based on the acquired 4DCT image sequence; The offline training module is used to sequentially input all tumor motion data points within each fixed time window on the motion trajectory into the CNN-Attention-LSTM model for offline training. The CNN-Attention-LSTM model is used to predict the tumor position at the next time. The online update module is used to update the trained CNN-Attention-LSTM model online in real time during the radiotherapy stage using the latest respiratory motion data of the current patient to predict the tumor location at the next time.

6. The device for predicting tumor position under respiratory motion according to claim 5, characterized in that: The extraction module comprises: A segmentation unit, used for extracting a tumor target boundary contour from a 4DCT image using a segmentation algorithm; A centroid calculation unit, used to calculate the centroid pixel position of the tumor in the image based on the extracted target boundary; A repeated execution unit, used for repeatedly executing the methods in the segmentation unit and the centroid calculation unit, until the centroid point positions of the tumors in all 4DCT image sequences are extracted; A curve calculation unit is used to calculate the motion trajectory curve of the tumor centroid position changing with time according to the pixel resolution and sampling frequency of the CT after the extraction is completed; The curve processing unit is used to perform normalization and / or smoothing filtering on the obtained motion trajectory curve.

7. The device for predicting tumor position under respiratory motion according to claim 5, characterized in that: The CNN-Attention-LSTM model includes: a CNN module, an attention mechanism module and an LSTM module; The inputs of the CNN module and the attention mechanism module are all tumor motion data points within each fixed time window on the motion trajectory; The high-dimensional features output by the CNN module and the saliency features output by the attention mechanism module are multiplied element by element, and the fused features are used as the input of the LSTM module; The LSTM module is used to predict the tumor location at the next time.

8. The device for predicting tumor position under respiratory motion according to claim 7, characterized in that: The importance of the high-dimensional features output by the CNN module is proportional to the value of the output result of the corresponding attention mechanism module.

Images