A motion prediction method and motion prediction device based on deep learning

Abstract

The embodiment of the present invention discloses a motion prediction method and a motion prediction device based on deep learning, which are used to improve the accuracy and reliability of positioning a region of interest in the clinical diagnosis process of human organs, thereby improving the resolution and signal-to-noise ratio of medical images, and ultimately improving the accuracy of medical images in the diagnosis and treatment of diseases. The motion prediction method includes: in the model training stage, using a sequence-to-sequence seq2seq model based on an attention mechanism to perform data training to obtain a motion prediction target model, the sequence-to-sequence seq2seq model includes an encoding unit and a decoding unit, and the attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit; in the real-time prediction stage, the real-time motion data of the region of interest collected in real time is input into the motion prediction target model to obtain a motion prediction target curve of the region of interest, and the region of interest includes a body part that is diagnosed and treated in the clinical diagnosis process of human organs.

Description

Technical Field

[0001] The present invention relates to the technical field of respiratory motion prediction in medical diagnosis and treatment, and in particular to a motion prediction method based on deep learning. Background Art

[0002] In the clinical diagnosis of human organs such as the chest and abdomen, human respiratory motion can cause the heart, lungs and other organs to shift as a whole in three-dimensional space. When imaging internal organs such as the heart and lungs, respiratory motion can cause respiratory motion artifacts in the acquired target images, reducing the resolution and signal-to-noise ratio of medical images, and adversely affecting the diagnosis and treatment of diseases.

[0003] In current technology, breathing control methods are usually used to solve the above technical problems. However, the breathing control method is easily affected by the human body condition and control ability, and cannot ensure the reliability of the time and spatial resolution of medical images. It is a passive tracking method and has low timeliness. Summary of the invention

[0004] Based on this, it is necessary to address the above problems. The present invention proposes a motion prediction method based on deep learning, which is used to improve the accuracy and reliability of positioning the region of interest in the clinical diagnosis process of human organs, thereby improving the resolution and signal-to-noise ratio of medical images, and ultimately improving the accuracy of medical images in the diagnosis and treatment of diseases.

[0005] In a first aspect, the present invention provides a motion prediction method based on deep learning, comprising:

[0006] In the model training stage, a sequence-to-sequence (seq2seq) model based on an attention mechanism is used to perform data training to obtain a motion prediction target model. The sequence-to-sequence (seq2seq) model includes an encoding unit and a decoding unit. The attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit. In the real-time prediction stage, the real-time motion data of the region of interest collected in real time is input into the motion prediction target model to obtain a motion prediction target curve of the region of interest. The region of interest includes body parts that are diagnosed and treated during the clinical diagnosis of human organs.

[0007] In a second aspect, the present invention provides a motion prediction device, comprising: a training module and a motion prediction module, wherein the training module is used to perform data training using a sequence-to-sequence (seq2seq) model based on an attention mechanism to obtain a motion prediction target model during a model training phase, the seq2seq model comprising an encoding unit and a decoding unit, and the attention mechanism is used to determine the correlation between a hidden layer state of the encoding unit and a hidden layer state of the decoding unit; the motion prediction module is used to input real-time motion data of a region of interest collected in real time into the motion prediction target model to obtain a motion prediction target curve of the region of interest, wherein the region of interest includes a body part that is diagnosed and treated during clinical diagnosis of human organs.

[0008] In a third aspect, the present invention provides another motion prediction device, comprising: a processor; and a memory having executable codes stored thereon, wherein when the executable codes are executed by the processor, the processor executes the method described in the first aspect above.

[0009] The technical solution provided by the present invention can achieve the following beneficial effects:

[0010] In the clinical diagnosis process of human organs, the motion prediction curve of the region of interest can be actively predicted through the sequence-to-sequence seq2seq model based on the attention mechanism. Since the seq2seq model adopts the encoding-decoding structure and introduces the attention mechanism, its structure has the ability to simultaneously model long-term and short-term time series features, which can effectively increase the length of the predictable sequence and improve the prediction reliability. At the same time, the structure can enhance the model's ability to model local contextual information in the time series, thereby reducing the impact of abnormal points on the prediction results, improving the prediction accuracy, and making the predicted motion prediction curve of the region of interest more reliable and accurate, thereby improving the resolution and signal-to-noise ratio of medical images, and ultimately improving the accuracy of medical images in the diagnosis and treatment of diseases.

[0011] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0013] Figure 1 A schematic diagram of a flow chart of an embodiment of a motion prediction method provided in an embodiment of the present invention;

[0014] Figure 2A schematic diagram of the overall network structure of a seq2seq model composed of GRU provided in an embodiment of the present invention;

[0015] Figure 3 A schematic diagram of a unit structure of a GRU unit in an embodiment of the present invention;

[0016] Figure 4 A schematic diagram of a unit structure of an LSTM unit in an embodiment of the present invention;

[0017] Figure 5 A schematic diagram of the structure of a motion prediction device according to an embodiment of the present invention;

[0018] Figure 6 It is a schematic diagram of the structure of another embodiment of the motion prediction device in the embodiment of the present invention;

[0019] Figure 7 FIG. 4 is a schematic diagram of the structure of another embodiment of the motion prediction device in the embodiment of the present invention. DETAILED DESCRIPTION

[0020] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0021] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0022] It should be understood that although the terms "first", "second", "third", etc. may be used to describe various information in the present invention, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0023] Among the related technologies, the breathing control method is easily affected by the patient's physical condition and control ability, and the temporal and spatial resolution of the image cannot be reliably guaranteed; the motion tracking method, the respiratory gating method and the forced shallow breathing method are all passive tracking methods with low timeliness.

[0024] In order to overcome the defects of the above-mentioned method, the present invention provides a motion prediction method based on deep learning, which can improve the accuracy and reliability of positioning the region of interest, thereby improving the resolution and signal-to-noise ratio of medical images, and ultimately improving the accuracy of medical images in the diagnosis and treatment of diseases.

[0025] In order to facilitate the understanding of the technical solution of this application, some professional terms involved in this application are briefly introduced below:

[0026] Sequence-to-sequenc (seq2seq) model: It is a very important and popular model in natural language processing technology. Simply put, it is a translation model that translates a language sequence into another language sequence. The entire processing process is done by using deep neural networks such as Long Short-Term Memory Network (LSTM) or Recurrent Neural Network (RNN), among which Gate Recurrent Unit (GRU) is a type of RNN.

[0027] The attention mechanism, also known as the attention model, is widely used in various types of deep learning tasks such as natural language processing, image recognition, and speech recognition. It is one of the core technologies in deep learning technology that deserves the most attention and in-depth understanding. The essence of the attention mechanism is a series of attention allocation coefficients, that is, a series of weight parameters, which can be used to emphasize or select important information of the target processing object and suppress some irrelevant details.

[0028] To facilitate understanding of the technical solutions in the embodiments of the present invention, the technical solutions provided by the present invention will be described below in conjunction with the accompanying drawings and specific embodiments.

[0029] Figure 1 The present invention is a flow chart of an embodiment of a motion prediction method provided in an embodiment of the present invention.

[0030] like Figure 1 As shown, the motion prediction method based on deep learning in the present invention includes:

[0031] 101. In the model training stage, the seq2seq model based on the attention mechanism is used for data training to obtain the motion prediction target model.

[0032] The seq2seq model based on the attention mechanism includes: an encoding unit and a decoding unit. The encoding unit is also called an encoder, and the decoding unit is also called a decoder. The attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit. In the specific design, the function of the attention mechanism can be realized by adding an attention unit or module to the seq2seq model. The specific structure of the seq2seq model based on the attention mechanism can be found in the following Figure 2-Figure 4 The relevant description of the corresponding part.

[0033] Optionally, the present invention uses a seq2seq model based on an attention mechanism to perform data training to obtain a motion prediction target model, which can be: first, inputting the training set corresponding to the region of interest into the seq2seq model based on the attention mechanism to generate a motion prediction training curve; secondly, calculating the error value between the motion prediction training curve and the actual motion curve corresponding to the region of interest, wherein the actual motion curve corresponding to the region of interest is obtained by actual measurement using a monitoring device; finally, inputting the error value into the error training model to automatically adjust the model parameters of the seq2seq model based on the attention mechanism, and finally obtaining the motion prediction target model.

[0034] The training set corresponding to the region of interest includes data after noise values ​​and obvious outliers are removed from all motion data collected in the region of interest. Specifically, in the model training stage, the method further includes: 1) collecting motion data of the region of interest; 2) removing noise values ​​and obvious outliers from the collected motion data of the region of interest to obtain a training set corresponding to the region of interest, and removing noise values ​​and obvious outliers is a preprocessing operation for the motion data. Further optionally, after the motion data of the region of interest is preprocessed, feature data can also be selected from the preprocessed motion data according to the label, and finally the selected feature data is summarized to obtain the training set.

[0035] Further optionally, in the model training stage, the function of "using the attention mechanism to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit" can be implemented by the following operations:

[0036] 1) Based on the attention mechanism, the correlation coefficient between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit is calculated; 2) The calculated correlation coefficient is subjected to a softmax operation to obtain an attention weight vector, wherein the softmax operation can be specifically implemented by a normalized softmax function; 3) The attention output vector is calculated based on the attention weight vector and the hidden layer state of the encoding unit; 4) The attention output vector of each time step is concatenated with the hidden layer state of the decoding unit of all time steps to obtain the final output vector of the model training.

[0037] The above-mentioned function of "the attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit" can be realized by the attention module, and its detailed description can be found below Figure 2-Figure 4 Related description of the attention module in the corresponding embodiment.

[0038] Optionally, the encoding unit and the decoding unit in the seq2seq model based on the attention mechanism can be composed of multi-layer cascade GRNs. In other words, the encoding operation and the decoding operation can be performed using multi-layer cascade GRNs. Alternatively, the encoding unit and the decoding unit can also be composed of multi-layer cascade LSTMs. In other words, the encoding operation and the decoding operation can be performed using multi-layer cascade LSTMs.

[0039] 102. In the real-time prediction stage, the real-time motion data of the region of interest collected in real time is input into the motion prediction target model to obtain a motion prediction target curve of the region of interest.

[0040] After completing the training of the motion prediction target model in the model training phase, in the real-time prediction phase: the real-time collected data is input into the trained network model to obtain the motion prediction curve of the region of interest.

[0041] The region of interest in the present invention includes, but is not limited to, the body parts that are diagnosed and treated during the clinical diagnosis of human organs, and may also be organs of interest during the diagnosis and treatment of animals.

[0042] The above technical solution can actively predict the motion prediction curve of the region of interest through the seq2seq model based on the attention mechanism in the clinical diagnosis process of human organs. Since the seq2seq model with an encoding-decoding structure is adopted and the attention mechanism is introduced, its structure has the ability to simultaneously model long-term and short-term time series features, which can effectively increase the length of the predictable sequence and improve the prediction reliability. At the same time, the structure can enhance the modeling ability of the model for local context information in the time series, thereby reducing the influence of abnormal points on the prediction results, improving the prediction accuracy, and making the predicted motion prediction curve of the region of interest more reliable and accurate, thereby improving the resolution and signal-to-noise ratio of medical images, and ultimately improving the accuracy of medical images in the diagnosis and treatment of diseases. Furthermore, the encoding and decoding parts are composed of LSTM network or GRU network modules, which can further improve its prediction reliability and accuracy.

[0043] Furthermore, the attention mechanism enables the seq2seq model to focus on the target-related parts of the input sequence during training and prediction, making the prediction results more predictive and robust than methods that do not use the attention mechanism.

[0044] Optionally, in an implementation manner of the embodiment of the present invention, after obtaining the motion prediction curve of the region of interest, the method further includes:

[0045] 103. Use the motion prediction target curve of the region of interest to guide the probe of the inspection device to move to the predicted position, or guide the radiation of the radiation therapy device to locate the position of the region of interest at that moment.

[0046] The seq2seq model based on the attention mechanism was mentioned above, but its structure was not explained. Figure 2-Figure 4 Taking the seq2seq model composed of GRU as an example, the structure of the seq2seq model based on the attention mechanism in the present invention is explained.

[0047] Figure 2 A schematic diagram of the overall network structure of a seq2seq model composed of GRU provided in an embodiment of the present invention.

[0048] like Figure 2 As shown, the seq2seq model based on the attention mechanism in the embodiment of the present invention includes: an encoder, a decoder and an attention module. Figure 2 In the example, S represents a start marker, A and B modules represent the GRU units that constitute the encoder and decoder, and x n represents the input sequence, y n represents the output prediction sequence, h n represents the hidden layer state of the encoder, h′ mRepresents the hidden layer state of the decoder.

[0049] In the attention module, a score is calculated by the hidden states in the decoder and the hidden states in the encoder to calculate the attention weight, which is divided into the following two steps:

[0050] 1) Calculate the correlation between the hidden states in the decoder and the hidden states in the encoder. The calculation method is as follows:

[0051]

[0052] in, represents the transpose of the decoder output at each time step, h′ m It represents the mth hidden layer output of the decoder, score is the correlation coefficient, and t is the time step.

[0053] 2) Perform a softmax operation on the obtained score to obtain a weight vector α with a probability sum of 1 i .

[0054] 3) Finally, the attention module outputs a vector: c i =h t *α i ;

[0055] In the attention module, the input is all the hidden states of the encoder, with a dimension size of (hid dim, sequence length). The hidden state of the decoder at a time step has a dimension size of (hid dim, 1). The dimension size of the score is (sequence length, 1). The dimension of the final attention module output vector is (hid dim, 1).

[0056] Finally, the output of each time step of the attention module is concatenated with the output of the decoder hidden layer at the corresponding time step to obtain the final output (i.e., the final output vector).

[0057] The encoder and decoder use multi-layer cascaded GRU or LSTM units. This embodiment uses two layers, and three or more layers can also be used depending on different application scenarios.

[0058] Figure 3 Schematic diagram of a unit structure of a GRU unit in an embodiment of the present invention.

[0059] like Figure 3As shown, z t represents the update gate, r t represents the reset gate, h t-1 is the hidden layer output of the previous GRU unit, h t is the updated hidden layer output of the current GRU unit, W z , W r , W represents the corresponding weight matrix, x t Represents the current input, and the calculation formula is as follows:

[0060] z t =σ(W z ·[h t-1 ,x t ])

[0061] r t =σ(W r ·[h t-1 ,x t ])

[0062]

[0063]

[0064] Figure 4 Schematic diagram of a unit structure of an LSTM unit in an embodiment of the present invention.

[0065] like Figure 4 As shown, f t represents the forget gate, i t represents the input gate, o t represents the output gate, h t-1 is the hidden layer output of the previous GRU unit, C t-1 is the state of the previous GRU unit, h t is the updated hidden layer output of the current GRU unit, W f , W i , W C , W o represents the corresponding weight matrix, x t Represents the current input, and the calculation formula is as follows: t =σ(W f ·[h t-1 ,x t ]+b f )

[0066] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0067]

[0068]

[0069] o t =σ(W o [h t-1 ,x t ]+b o )

[0070] h t =o t *tanh(C t )

[0071] Corresponding to the aforementioned application function realization method embodiment, the present invention also provides a motion prediction device and corresponding embodiments.

[0072] Figure 5 FIG. 4 is a schematic diagram of the structure of an embodiment of a motion prediction device in an embodiment of the present invention.

[0073] like Figure 5 As shown, the motion prediction device 50 includes: a training module 501 and a motion prediction module 502, wherein the training module 501 is used to perform data training using a sequence-to-sequence seq2seq model based on an attention mechanism to obtain a motion prediction target model during a model training phase, the seq2seq model includes an encoding unit and a decoding unit, and the attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit;

[0074] The motion prediction module 502 is used to input the real-time motion data of the region of interest collected in real time into the motion prediction target model to obtain the motion prediction target curve of the region of interest. The region of interest includes the body part that is diagnosed and treated during the clinical diagnosis of human organs.

[0075] Optionally, the encoding unit and the decoding unit: the encoding and decoding operations can be performed using multi-layer cascaded gated recurrent units; or, the encoding and decoding operations can be performed using multi-layer cascaded long short-term memory network units.

[0076] In one implementation of the embodiment of the present invention, the training module 501 is specifically configured to perform the following operations:

[0077] 1) Input the training set corresponding to the region of interest into the sequence-to-sequence seq2seq model based on the attention mechanism to generate the motion prediction training curve;

[0078] 2) Calculate the error value between the motion prediction training curve and the actual motion curve corresponding to the region of interest;

[0079] 3) Input the error value into the error training model to automatically adjust the model parameters of the sequence-to-sequence seq2seq model based on the attention mechanism, and finally obtain the motion prediction target model.

[0080] In one implementation of the embodiment of the present invention, the training module 501 is further configured to perform the following operations:

[0081] Calculate the correlation coefficient between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit based on the attention mechanism;

[0082] Perform a softmax operation on the calculated correlation coefficient to obtain the attention weight vector;

[0083] 1) Calculate the attention output vector based on the attention weight vector and the hidden layer state of the coding unit;

[0084] 2) Concatenate the attention output vector of each time step with the hidden layer state of the decoding unit of all time steps to obtain the final output vector of the model training.

[0085] Figure 6 FIG. 4 is a schematic diagram of the structure of another embodiment of the motion prediction device in the embodiment of the present invention.

[0086] like Figure 6 As shown, Figure 5 Compared with the structure shown, the motion prediction device 50 may further include: a data processing module 503 and a guiding module 504;

[0087] The data processing module 503 is used to remove noise values ​​and obvious abnormal values ​​from the collected motion data of the region of interest to obtain a training set corresponding to the region of interest.

[0088] The guiding module 504 is used to guide the probe of the inspection device to move to the predicted position by using the motion prediction target curve of the region of interest, or to guide the rays of the radiation therapy device to be positioned at the position where the region of interest is at that moment.

[0089] It should be noted that the other functions and beneficial effects of the training module 501, the motion prediction module 502, the data processing module 503 and the guidance module 504 are the same as those described above. Figure 1 The corresponding method embodiments are the same, and the relevant descriptions can be found in the above Figure 1 The description of the relevant parts in the corresponding method embodiment will not be repeated here.

[0090] Figure 7 FIG. 4 is a schematic diagram of the structure of another embodiment of the motion prediction device in the embodiment of the present invention.

[0091] like Figure 7As shown, the motion prediction device 70 includes: a memory 701 and a processor 702. The memory stores executable codes, and when the executable codes are executed by the processor, the processor executes the motion prediction method based on deep learning described in any of the above embodiments.

[0092] The processor 702 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0093] The memory 701 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage. Among them, ROM can store static data or instructions required by the processor 702 or other modules of the computer. The permanent storage device may be a readable and writable storage device. The permanent storage device may be a non-volatile storage device that does not lose the stored instructions and data even after the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (such as a magnetic or optical disk, flash memory) as a permanent storage device. In some other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, an optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as a dynamic random access memory. The system memory may store some or all instructions and data required by the processor at run time. In addition, the memory 701 may include a combination of any computer-readable storage media, including various types of semiconductor memory chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, the memory 701 may include a readable and / or writable removable storage device, such as a laser disc (CD), a read-only digital versatile disc (such as a DVD-ROM, a double-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (such as an SD card, a mini SD card, a Micro-SD card, etc.), a magnetic floppy disk, etc. The computer-readable storage medium does not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0094] The memory 701 stores executable codes, and when the executable codes are processed by the processor 702 , the processor 702 can execute part or all of the methods described above.

[0095] In addition, the method according to the present invention may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing part or all of the steps in the above method of the present invention.

[0096] Alternatively, the present invention may also be implemented as a computer-readable storage medium (or machine-readable storage medium) on which executable code (or computer program, or computer instruction code) is stored. When the executable code (or computer program, or computer instruction code) is executed by a processor of an electronic device (or electronic device, server, etc.), the processor executes part or all of the steps of the above-mentioned method according to the present invention.

[0097] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein are selected to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A motion prediction method based on deep learning, characterized in that: include: In the model training stage, a sequence-to-sequence seq2seq model based on an attention mechanism is used to perform data training to obtain a motion prediction target model, wherein the sequence-to-sequence seq2seq model includes an encoding unit and a decoding unit, and the attention mechanism is used to determine the correlation between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit; In the real-time prediction stage, the real-time motion data of the region of interest collected in real time is input into the motion prediction target model to obtain the motion prediction target curve of the region of interest, wherein the region of interest includes the body part to be diagnosed and treated during the clinical diagnosis of human organs; The method of using a sequence-to-sequence seq2seq model based on an attention mechanism to perform data training to obtain a motion prediction target model includes: inputting a training set corresponding to the region of interest into the sequence-to-sequence seq2seq model based on an attention mechanism to generate a motion prediction training curve; calculating an error value between the motion prediction training curve and an actual motion curve corresponding to the region of interest; inputting the error value into an error training model to automatically adjust a model parameter of the sequence-to-sequence seq2seq model based on an attention mechanism, and finally obtaining the motion prediction target model; During the model training stage, the method also includes: calculating the correlation coefficient between the hidden layer state of the encoding unit and the hidden layer state of the decoding unit based on the attention mechanism; performing a softmax operation on the calculated correlation coefficient to obtain an attention weight vector; calculating an attention output vector based on the attention weight vector and the hidden layer state of the encoding unit; and concatenating the attention output vector of each time step with the hidden layer state of the decoding unit of all time steps to obtain the final output vector of the model training.

2. The method according to claim 1, characterized in that In the model training stage, the method further includes: Use multiple layers of cascaded gated recurrent units to perform encoding and decoding operations; Alternatively, multiple layers of cascaded long short-term memory network units are used to perform encoding and decoding operations.

3. The method according to claim 1, characterized in that In the model training stage, the method further includes: The collected motion data of the region of interest is subjected to noise values ​​and obvious abnormal values ​​to obtain a training set corresponding to the region of interest.

4. The method according to claim 1, characterized in that: The method further comprises: The motion prediction target curve of the region of interest is used to guide the probe of the inspection device to move to the predicted position, or to guide the rays of the radiation therapy device to be positioned at the position where the region of interest is located.

5. A motion prediction device, characterized in that: include: A training module, used for obtaining a motion prediction target model by performing data training using a sequence-to-sequence (seq2seq) model based on an attention mechanism in a model training phase, wherein the seq2seq model includes an encoding unit and a decoding unit, and the attention mechanism is used to determine a correlation between a hidden layer state of the encoding unit and a hidden layer state of the decoding unit; A motion prediction module, used to input the real-time motion data of the region of interest collected in real time into the motion prediction target model to obtain a motion prediction target curve of the region of interest, wherein the region of interest includes a body part to be diagnosed and treated during the clinical diagnosis of human organs; The training module is specifically used to input the training set corresponding to the region of interest into the sequence-to-sequence (seq2seq) model based on the attention mechanism to generate a motion prediction training curve; calculate the error value between the motion prediction training curve and the actual motion curve corresponding to the region of interest; Inputting the error value into the error training model to automatically adjust the model parameters of the sequence-to-sequence seq2seq model based on the attention mechanism, and finally obtaining the motion prediction target model; Calculating a correlation coefficient between a hidden layer state of the encoding unit and a hidden layer state of the decoding unit based on an attention mechanism; A softmax operation is performed on the calculated correlation coefficient to obtain an attention weight vector; an attention output vector is calculated based on the attention weight vector and the hidden layer state of the encoding unit; and the attention output vector of each time step is concatenated with the hidden layer state of the decoding unit of all time steps to obtain the final output vector of the model training.

6. A motion prediction device, characterized in that: include: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to execute the method according to any one of claims 1 to 4.

Images