A method and apparatus for predicting organ motion changes during radiotherapy

Abstract

The application discloses a method and device for predicting organ motion changes in a radiotherapy process, and relates to the technical field of medical radiotherapy. The method solves the problem that the future motion of an organ in a radiotherapy process cannot be accurately predicted in the prior art. The method comprises the following steps: extracting deformation field data of a historical CT image, and using the deformation field data as input to predict the deformation field data at a current time by using a pre-trained image deformation field prediction model; generating a CT image at a next time; and evaluating displacement by comparing the similarity of the CT images at the current time and the next time. The method uses deformation field as an intermediate medium, can more accurately capture the changes between images, and thus improves the prediction accuracy. By analyzing the deformation field data, the pre-trained image deformation field prediction model can predict the possible position and shape of the organ at a future time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical radiotherapy, and in particular to a method and device for predicting organ motion changes during radiotherapy. BACKGROUND

[0002] During radiotherapy treatment, organ displacement occurs due to human motion or physiological changes, which affects the radiotherapy effect. For example, the filling condition of the bladder causes changes in the position of the organ, and the risk of increasing damage to normal tissues has always been a concern in clinical practice. In order to reduce the impact of organ motion on radiotherapy, ensure that the radiation dose is accurately delivered to the tumor tissue, and at the same time minimize the damage to the surrounding healthy tissue. The commonly used techniques are:

[0003] Respiratory gating technology: By monitoring the patient's breathing pattern, radiotherapy is only performed at specific stages of the respiratory cycle (such as exhalation or inhalation), reducing organ movement due to respiratory motion.

[0004] Body position fixation technology: Special fixation devices (such as vacuum pads, thermoplastic films, etc.) are used to keep the patient's body position unchanged during treatment. This helps to reduce organ motion due to patient movement.

[0005] Real-time tracking system: External sensors or internal markers are used to track the position of the organ in real time, and the radiotherapy equipment is adjusted in real time to compensate for the motion of the organ.

[0006] Image-guided radiotherapy (IGRT): Imaging technology (such as CT, MRI or X-ray) is used during treatment to obtain real-time organ position information, enabling precise adjustment of radiotherapy equipment to ensure accurate delivery of radiation dose.

[0007] In the above-mentioned existing technologies, the first two are open-loop control methods before radiotherapy, and their accuracy cannot be guaranteed; the last two are real-time monitoring methods during radiotherapy, which cannot predict future changes in organs and require real-time adjustment of radiotherapy equipment. SUMMARY

[0008] The present application aims to overcome the problem that the future motion of the organ during radiotherapy cannot be accurately predicted in the prior art, and provides a method and device for predicting organ motion changes during radiotherapy.

[0009] In a first aspect, a method for predicting organ motion changes during radiotherapy is provided, comprising:

[0010] obtaining a CT image at a historical time;

[0011] extracting deformation field data of the historical CT image;

[0012] Input the deformation field data of the historical CT image into a pre-trained image deformation field prediction model to predict the deformation field data at the current time;

[0013] Calculate the CT image at the next time according to the deformation field data at the current time and the CT image at the current time;

[0014] By comparing the similarity of the CT images at the current time and the next time, the displacement situation is evaluated.

[0015] In some possible implementations, the historical CT image is a CT image taken at time 0 to time t after the patient is fixed in position using a body position fixing device, wherein time 0 is the initial time and time t is the current time.

[0016] In some possible implementations, the deformation field data of the historical CT image is extracted, including:

[0017] An image registration network is constructed, wherein the image registration network includes an encoder and a decoder, the encoder is used to extract image features of an input CT image, and the decoder is used to decode and reconstruct the image features output by the encoder into deformation field data;

[0018] CT image I0 and CT image I1 are input into the image registration network to obtain deformation field data f0, CT image I1 and CT image I2 are input into the image registration network to obtain deformation field data f1, and so on, CT image I t-1 and CT image I t are input into the image registration network to obtain deformation field data f t-1 , until deformation field data f0 to f t-1 are obtained, wherein CT image I0 is a CT image taken at initial time 0, CT image I t is a CT image taken at the current time t, and CT image I t-1 is a CT image taken at the previous time t-1 of the current time t.

[0019] In some possible implementations, the encoder adopts a block stack of Swin-transformer structure to realize extraction and fusion of image information at two times; and the decoder is composed of deconvolution blocks in a conventional convolutional neural network to reconstruct the fused feature information into deformation field data of images at two times.

[0020] In some possible implementation manners, the pre-trained image deformation field prediction model adopts a CA_RNN structure network capable of directly inputting two-dimensional matrix data and outputting two-dimensional matrix data, wherein the input gate in each unit processes data as follows: the two-dimensional data h t-1 of the hidden state transmitted at the last moment and the two-dimensional deformation field data x t input at the current moment are both subjected to feature extraction by a convolution block, then subjected to channel information fusion by a channel attention mechanism operation after channel splicing to obtain two-dimensional data h t of the hidden state t for transmission to the input gate of the next unit or the current unit.

[0021] In some possible implementation manners, the calculation formula of the CT image at a moment is as follows:

[0022] I t+1 =I t *f t

[0023] wherein I t+1 is the CT image at the next moment t+1, I t is the CT image at the current moment t, f t is the predicted deformation field data at the current moment, and * is matrix point multiplication.

[0024] In some possible implementation manners, in the process of evaluating the displacement, a structural similarity index is used to measure the similarity between images, and if the evaluated displacement is greater than a preset threshold, a prewarning prompt is given.

[0025] In a second aspect, a device for predicting organ motion changes in a radiotherapy process is provided, which comprises:

[0026] an image acquisition module configured to acquire a CT image at a historical moment;

[0027] an image deformation field extraction module configured to extract deformation field data of the historical CT image;

[0028] an image deformation field prediction module configured to input the deformation field data of the historical CT image into a pre-trained image deformation field prediction model to predict deformation field data at a current moment;

[0029] an image generation module configured to calculate a CT image at a next moment according to the deformation field data at the current moment and the CT image at the current moment;

[0030] a motion displacement evaluation module configured to evaluate a displacement by comparing the similarity between the CT image at the current moment and the CT image at the next moment.

[0031] In a third aspect, a computer readable storage medium is provided, the computer readable medium storing program codes for execution by a device, the program codes comprising steps for performing the method in any one of the implementation manners of the first aspect.

[0032] In a fourth aspect, an electronic device is provided, the electronic device comprising a processor, a memory, and a program or instructions stored on the memory and executable on the processor, the program or instructions, when executed by the processor, implementing the method in any one of the implementation manners of the first aspect.

[0033] The present application has the following beneficial effects:

[0034] 1. The present application uses deformation field as an intermediate medium, which can more accurately capture the changes between images, thereby improving the prediction accuracy. Through the analysis of deformation field data, the pre-trained image deformation field prediction model can predict the possible position and shape of the organ at the future time, thereby providing more accurate reference for treatment.

[0035] 2. The present application adopts a new CA_RNN (Convolutional Recurrent Neural Network) structure, which enables the original RNN unit to effectively process two-dimensional input data, thereby better adapting to the needs of medical image processing. Through this improvement, the pre-trained image deformation field prediction model can more accurately capture and analyze the spatio-temporal features in the data, and can more accurately estimate the future movement of the organ during radiotherapy. BRIEF DESCRIPTION OF DRAWINGS

[0036] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and their

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0038] Figure 1 is a flowchart of the method for predicting the movement of organs during radiotherapy in embodiment 1 of the present application;

[0039] Figure 2 is a schematic diagram of extracting deformation field data in the method for predicting the movement of organs during radiotherapy in embodiment 1 of the present application;

[0040] Figure 3 is a schematic diagram of the original RNN unit;

[0041] Figure 4 is a CA_RNN unit input gate calculation schematic diagram of the method for predicting organ motion changes in a radiotherapy process in Embodiment 1 of the present application;

[0042] Figure 5 is a whole flowchart of the method for predicting organ motion changes in a radiotherapy process in Embodiment 1 of the present application;

[0043] Figure 6 is an expanded flowchart of the method for predicting organ motion changes in a radiotherapy process in Embodiment 1 of the present application;

[0044] Figure 7 is a schematic diagram of calculating a CT image at a next time in the method for predicting organ motion changes in a radiotherapy process in Embodiment 1 of the present application;

[0045] Figure 8 is a structure block diagram of the device for predicting organ motion changes in a radiotherapy process in Embodiment 2 of the present application;

[0046] Figure 9 is an internal structure schematic diagram of the electronic device in Embodiment 4 of the present application.

[0047] Reference signs:

[0048] 100, image acquisition module; 200, image deformation field extraction module; 300, image deformation field prediction module; 400, image generation module; 500, motion displacement evaluation module. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0050] Embodiment 1

[0051] As shown in the method for predicting organ motion changes in a radiotherapy process in Embodiment 1 of the present application, the method comprises the following steps: Figure 1

[0052] S100, acquiring a CT image at a historical time;

[0053] ​It should be noted that CT (Computed Tomography) is an electronic computer tomography, which is to scan a certain thickness of a layer of a human body examination site by an X-ray beam, receive the X-rays through the human body tissues in each different direction of the layer by a detector, input the X-rays into a computer after analog / digital conversion, obtain a digital matrix of the scanned tomographic tissue attenuation coefficient after computer processing, and display the values in the matrix on a fluorescent screen by digital / analog conversion with different gray levels, that is, to form a CT image. The CT image can be obtained from a CT machine. In order to reduce the organ movement amplitude in the process of radiotherapy as much as possible, the historical CT image is a CT image taken at time 0 to time t after the patient is fixed in a body position by using a body position fixing device, so as to help reduce the organ movement caused by the patient movement, and further reduce the influence of the organ movement on radiotherapy, ensure that the radiation dose is accurately delivered to the tumor tissue, and at the same time, try to reduce the damage to the surrounding healthy tissues. Wherein, time 0 is the initial time, and time t is the current time.

[0054] S200, extract deformation field data of the historical CT image;

[0055] In this embodiment, an image registration network based on an encoder-decoder architecture is used to extract the deformation field data, as shown in the following figure. Figure 2 As shown in the figure, the image registration network includes an encoder and a decoder. The encoder is used to extract image features of the input CT image, and the decoder is used to decode and reconstruct the image features output by the encoder into deformation field data. When the deformation field data is extracted, first, the current time t image and the next time t+1 image are channel spliced and input into the encoder to extract image features, and then the above features are decoded and reconstructed into image deformation field by the decoder. Wherein, the encoder is stacked by blocks with Swin-transformer structure, which realizes the extraction and fusion of image information in two time points; the decoder is composed of deconvolution blocks in a conventional convolutional neural network, which reconstructs the fused feature information into deformation field data related to the images in two time points.

[0056] Specifically, first, the CT image I0 and the CT image I1 are input into the image registration network to obtain the deformation field data f0, the CT image I1 and the CT image I2 are input into the image registration network to obtain the deformation field data f1, and so on. The CT image I t-1 and the CT image I t are input into the image registration network to obtain the deformation field data f t-1 , until the deformation field data f0 to f t-1 are obtained, wherein the CT image I0 is a CT image taken at the initial time 0, the CT image I t is a CT image taken at the current time t, and the CT image I t-1a CT image taken at a previous time t-1 of the current time t.

[0057] It should be noted that Swin Transformer is a deep learning model architecture for computer vision tasks, proposed by a research team of Microsoft Research Asia in 2021. It is based on the Transformer architecture, but unlike traditional Transformers, Swin Transformer is particularly suitable for processing image data to extract and fuse context feature information of different scales due to its unique hierarchical feature map and window self-attention mechanism, and performs well in the field of computer vision (such as image classification, object detection, and semantic segmentation). And the deconvolution network is a way to reconstruct low-dimensional feature maps in image models with encoder-decoder architecture.

[0058] S300, input the deformation field data of the historical CT image into the pre-trained image deformation field prediction model to predict the deformation field data of the current time;

[0059] In time series data analysis, recurrent neural network (RNN) and long short-term memory network (LSTM) are common structures. RNN structure is simple and suitable for processing sequences of different lengths, but long sequences may cause gradient problems. LSTM solves this problem through a gating mechanism and is suitable for learning long-term dependencies. However, in specific scenarios such as the field involved in the present application, only recent data needs to be focused on, so the gradient problem of RNN is not significant. This makes RNN more advantageous in terms of simplifying structure and improving computational efficiency, and easier to train. Therefore, the present application uses the CA_RNN structure improved based on RNN for spatiotemporal sequence prediction.

[0060] As shown in Figure 3 , in the original RNN network, each unit input gate and output gate uses a fully connected linear layer structure, specifically as follows:

[0061]

[0062] where h is the hidden state, x is the input, and b h , b y are one-dimensional vectors, W hh , W xh , W hy are linear layer weight matrices, and f i , f0 are input gate and output gate nonlinear activation functions.

[0063] The RNN original unit is only suitable for processing single-dimensional time series data and is not suitable for processing space-time sequence image type data. In the present application, the input gate and the output gate in the RNN network are improved to obtain a CA_RNN structure network, so that it can directly input two-dimensional matrix data and output two-dimensional data. As shown in Figure 4 , the CA_RNN structure network can directly input two-dimensional matrix data and output two-dimensional matrix data, wherein the processing process of the input gate in each unit for data is as follows: the two-dimensional data h t-1 of the hidden state transmitted at the last moment and the two-dimensional deformation field data x t input at the present moment are both subjected to feature extraction by a convolution block, then subjected to channel splicing and then subjected to channel information fusion by a channel attention mechanism operation to obtain two-dimensional data h t of the hidden state, h t is used for transmission to the next unit or the input gate of the present unit. The calculation process can be represented by the following formula:

[0064]

[0065] wherein Convb() represents a convolution operation, Attention() represents an operation subjected to a channel attention mechanism, and Conv1*1() represents a convolution operation with a convolution kernel size of 1*1.

[0066] Similarly, the CA_RNN unit output gate converts the original linear layer operation in the RNN into a convolution operation as shown in Figure 4 , which adapts to the requirements in the present embodiment.

[0067] As shown in Figure 5 and Figure 6 , a "zipper" structure is adopted, which can be unfolded according to the length of the input image sequence and does not limit the length of the input sequence, and is flexible in the actual application process. Among them, the deformation field extraction module and the deformation field prediction module share the model weight in each time step. First, the deformation field data is extracted according to the current moment image and the next moment image to obtain the current moment deformation field matrix data, such as the deformation field matrix f1 in Figure 5 and Figure 6 . In the pre-trained image deformation field prediction model, the current moment deformation field matrix data f1 is taken as the input, the hidden state data matrix information h0 transmitted from the last moment is fused, and the pre-trained image deformation field prediction model transmitted to the next moment is calculated. In the last moment of the image sequence, the hidden state data matrix information h3 is subjected to convolution operation to predict the deformation field matrix information f4 of the next moment. Finally, f4 and I4 image are subjected to matrix point multiplication calculation to obtain the to-be-predicted image I5.

[0068] In the aforementioned end-to-end neural network, the similarity between the image predicted by the model and the labeled image is used as the training loss to learn the weight parameters in the deformation field extraction module and the prediction module. The specific loss function is shown in Equation I below. f Neutralization I m Let f be the image at the next time step in the time series corresponding to the fixed image and the moving image, and f be the deformation field corresponding to the image at the current time step. The loss function includes a similarity loss L. sim and smoothness L smooth It consists of two parts, L sim The loss is represented by the root mean square error of the deformation field of the fixed image and the moving image, L smooth The loss effect is to limit the smoothness of the deformation field. The deformation in the time series should be relatively small and smooth, which is represented by the first norm of the deformation field and the second norm of its gradient.

[0069]

[0070]

[0071]

[0072] Furthermore, the model employs a self-supervised training method during training, which has lower requirements for training data. For example, in the image sequence (I1,I2,I3,I4,I5), I3 is first predicted by (I1,I2) as a training sample, then (I1,I2,I3) is predicted as a training sample, and finally (I1,I2,I3,I4) is predicted, and so on forward.

[0073] S400. Calculate the CT image at the next moment based on the deformation field data and the CT image at the current moment.

[0074] This step calculates the image for the next moment based on the model-predicted deformation field data and the image data at the current moment. For example... Figure 7 As shown, the calculation process involves multiplying the corresponding elements of the two matrices element by element to generate the image data for the next time step. The calculation process is shown in the following formula:

[0075] I t+1 =I t *f t

[0076] Among them, I t+1 For the CT image at the next time step t+1, I t f is the CT image at the current time t. t The data represents the predicted deformation field at the current moment, and * represents matrix dot product.

[0077] S500, evaluate the displacement condition by comparing the similarity of the CT images of the current time and the next time.

[0078] In this step, the displacement condition is evaluated by comparing the similarity between the initial organ image and the predicted image. If a large displacement is detected, the system will trigger a warning prompt. In this embodiment, the structural similarity (SSIM) index is used to measure the similarity between images. SSIM is a comprehensive image quality evaluation method that considers brightness, contrast and structure as three key visual dimensions to accurately evaluate and compare the similarity between images. Through this multi-dimensional comparison, SSIM can provide more accurate and reliable image similarity evaluation results, effectively monitoring and warning displacement abnormalities.

[0079] For example, in the treatment of bladder cancer, the target area and organ at risk are determined before treatment, the dose distribution is calculated, and the radiotherapy equipment parameters are determined. However, during radiotherapy, the bladder may move due to physiological reasons, causing the original equipment parameters to be unsuitable, the target area dose to be insufficient, and the organ at risk dose to be too high, increasing the side effects. Using the method of the present application, the displacement condition can be predicted by historical CT image information during treatment, and a warning can be given. Specifically, the first, second and third time CT images are input, the first and second time deformation field data are obtained by extracting the deformation field data, and the third time deformation field data is predicted by the pre-trained image deformation field prediction model. The fourth time CT image is obtained by spatial transformation of the third time CT image + deformation field, and the similarity between the fourth CT image and the first CT image is calculated to determine whether there is a displacement risk. If there is, a warning is given, otherwise, the fifth, sixth, seventh, … time CT images are predicted to continue to determine whether there is a displacement risk.

[0080] During radiotherapy, the motion monitoring of patient organs is crucial for ensuring the accuracy and safety of treatment. Traditional real-time monitoring methods can capture organ motion in real time during treatment, but cannot predict future motion states. In order to improve the accuracy of prediction and achieve more effective warning, an innovative end-to-end whole network structure is proposed. Deformation field is used as an intermediate medium instead of directly predicting the specific image of the future time. The advantage of this method is that the deformation field can more accurately capture the changes between images, thereby improving the accuracy of prediction. By analyzing the deformation field data, the network can predict the possible position and shape of the organ at the future time, thereby providing more accurate reference for treatment.

[0081] Secondly, a new CA_RNN (Convolutional Recurrent Neural Network) structure is innovatively proposed in the present application. This structure enables the original RNN unit to effectively process two-dimensional input data, thereby better adapting to the needs of medical image processing. Through this improvement, the spatiotemporal features in the data can be more accurately captured and analyzed, providing more advanced and effective technical means for patient monitoring during radiotherapy.

[0082] Embodiment 2

[0083] As shown in Figure 8 the present application embodiment 2 relates to a device for predicting organ motion changes during radiotherapy, comprising:

[0084] An image acquisition module 100 is configured to acquire CT images at historical time points.

[0085] An image deformation field extraction module 200 is configured to extract deformation field data of the historical CT images.

[0086] An image deformation field prediction module 300 is configured to input the deformation field data of the historical CT images into a pre-trained image deformation field prediction model to predict the deformation field data at the current time point.

[0087] An image generation module 400 is configured to calculate the CT images at the next time point based on the deformation field data at the current time point and the CT images at the current time point.

[0088] A motion displacement evaluation module 500 is configured to evaluate the displacement by comparing the similarity of the CT images at the current time point and the next time point.

[0089] It should be noted that other specific embodiments of the device for predicting organ motion changes during radiotherapy in the present embodiment can refer to the specific embodiments of the method for predicting organ motion changes during radiotherapy described above, and to avoid redundancy, details are not repeated here.

[0090] Embodiment 3

[0091] The computer readable storage medium involved in the present application embodiment 3 stores program code for device execution, and the program code includes steps for executing the method in any one of the implementation manners in the present application embodiment 1.

[0092] The computer-readable storage medium may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM); the computer-readable storage medium may store program code, and when the program stored in the computer-readable storage medium is executed by a processor, the processor is used to perform the steps of the method in any of the implementations of Embodiment 1 of this application.

[0093] Example 4

[0094] like Figure 9 As shown, an electronic device according to Embodiment 4 of this application includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the method in any of the implementations in Embodiment 1 of this application.

[0095] The processor can be a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), graphics processing unit (GPU), or one or more integrated circuits, used to execute related programs to implement the method in any of the implementations of Embodiment 1 of this application.

[0096] The processor can also be an integrated circuit electronic device with signal processing capabilities. In implementation, each step of the method in any of the implementations of Embodiment 1 of this application can be completed by the integrated logic circuitry in the processor's hardware or by software instructions.

[0097] The above processor can also be a general processor, a digital signal processor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can be any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware code processor execution or executed by hardware and software module combination in the code processor. The software module can be located in the random access memory, the flash memory, the read only memory, the programmable read only memory or the electrically erasable programmable memory, the register or the mature storage medium in the art. The storage medium is located in the memory, the processor reads the information in the memory, combines the hardware to complete the function required by the unit in the data processing device of the embodiments of the present application, or executes the method in any one of the implementation manners in the embodiments of the present application.

[0098] The above is only the preferred specific embodiment of the present application; however, the protection scope of the present application is not limited to this. Any skilled person in the art can make equivalent replacement or change according to the technical range disclosed in the present application and the improvement concept of the technical solution and the embodiments of the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for predicting organ motion changes during radiotherapy, characterized in that, The method comprises the following steps: acquiring CT images at historical time points, wherein the CT images at the historical time points are CT images taken at time 0 to time t after a patient is fixed in position using a body position fixing device, wherein time 0 is an initial time point and time t is a current time point; extracting deformation field data of the historical CT images, specifically comprising: constructing an image registration network, wherein the image registration network comprises an encoder and a decoder, the encoder is used to extract image features of an input CT image, and the decoder is used to decode and reconstruct the image features output by the encoder into deformation field data; inputting CT image I0 and CT image I1 into the image registration network to obtain deformation field data f0, inputting CT image I1 and CT image I2 into the image registration network to obtain deformation field data f1, and inputting CT image It-1 and CT image It into the image registration network to obtain deformation field data ft-1 in a similar manner until deformation field data f0 to ft-1 are obtained, wherein CT image I0 is a CT image taken at initial time point 0, CT image It is a CT image taken at current time point t, and CT image It-1 is a CT image taken at the previous time point t-1 of current time point t, the encoder is stacked by blocks of Swin-transformer structure to realize extraction and fusion of image information at two time points, and the decoder is composed of deconvolution blocks in a conventional convolutional neural network to reconstruct the fused feature information into deformation field data of images at two time points; The deformation field data of the CT image at the historical moment is input into the pre-trained image deformation field prediction model to predict the deformation field data at the current moment, wherein the pre-trained image deformation field prediction model adopts a CA_RNN structure network, the CA_RNN structure network can directly input and output two-dimensional matrix data, wherein the input gate in each unit processes data as follows: the two-dimensional hidden state data h t-1 transmitted at the last moment and the two-dimensional deformation field data x t input at the current moment are both subjected to feature extraction by a convolution block, then subjected to inter-channel splicing and then subjected to channel information fusion by a channel attention mechanism operation to obtain two-dimensional hidden state data h t , h t is used for transmission to the input gate of the next unit or the current unit; calculating a CT image at a next time point according to the deformation field data at the current time point and the CT image at the current time point, wherein the calculation formula of the CT image at the next time point is: I t+1 =I t *f t where I t+1 is the CT image of next time t+1, I t is the CT image of current time t, f t is the predicted deformation field data of current time. evaluating an organ displacement condition by comparing the similarity of the CT images at the current time point and the next time point, wherein a structural similarity index is used to measure the similarity between images during the evaluation of the displacement condition, and a warning prompt is given if the evaluated displacement is greater than a preset threshold.

2. A device for predicting organ motion changes during radiotherapy, characterized in that, The method comprises the following steps: an image acquisition module is configured to acquire CT images at historical time points, wherein the CT images at the historical time points are CT images taken at time 0 to time t after a patient is fixed in position using a body position fixing device, wherein time 0 is an initial time point and time t is a current time point; An image deformation field extraction module is configured to extract deformation field data of historical CT images, and specifically includes: constructing an image registration network, wherein the image registration network includes an encoder and a decoder, the encoder is configured to extract image features of an input CT image, and the decoder is configured to decode and reconstruct the image features output by the encoder into deformation field data; inputting CT image I0 and CT image I1 into the image registration network to obtain deformation field data f0, inputting CT image I1 and CT image I2 into the image registration network to obtain deformation field data f1, and inputting CT image It-1 and CT image It into the image registration network to obtain deformation field data ft-1 in a similar manner, until deformation field data f0 to ft-1 are obtained, wherein CT image I0 is a CT image taken at initial time 0, CT image It is a CT image taken at current time t, CT image It-1 is a CT image taken at previous time t-1 of current time t, the encoder is stacked by blocks with a Swin-transformer structure to realize extraction and fusion of image information at two times; the decoder is composed of deconvolution blocks in a conventional convolutional neural network to reconstruct the fused feature information into deformation field data of images at two times; The image deformation field prediction module is configured to input the deformation field data of the CT image at the historical time into a pre-trained image deformation field prediction model to predict the deformation field data at the current time, wherein the pre-trained image deformation field prediction model adopts a CA_RNN structure network, the CA_RNN structure network can directly input two-dimensional matrix data and output two-dimensional matrix data, wherein the input gate in each unit processes the data as follows: the two-dimensional hidden state data h t-1 transmitted at the last time and the two-dimensional deformation field data x t input at the current time are both subjected to feature extraction by a convolution block, then subjected to channel splicing and channel information fusion by a channel attention mechanism operation to obtain two-dimensional hidden state data h t , h t is used for transmitting to the input gate of the next unit or the current unit. An image generation module is configured to calculate a CT image at a next time according to deformation field data at the current time and a CT image at the current time, wherein a calculation formula of the CT image at the next time is: I t+1 =I t *f t where I t+1 is the CT image of next time t+1, I t is the CT image of current time t, f t is the predicted deformation field data of current time. A motion displacement evaluation module is configured to evaluate an organ displacement condition by comparing similarities of CT images at the current time and at the next time, and a structural similarity index is used to measure the similarity between images during the evaluation of the displacement condition, and a warning is given if the evaluated displacement is greater than a preset threshold.

3. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program codes for device execution, and the program codes include steps for executing the method of claim 1.

4. An electronic device, comprising: The electronic device includes a processor, a memory, and a program or instruction stored on the memory and executable on the processor, and the program or instruction is executed by the processor to implement the method of claim 1.

Images