Radiotherapy plan generation method, system and equipment based on ordinary differential equation, and medium

Abstract

The invention discloses a radiotherapy plan generation method, system and device based on an ordinary differential equation and a medium, belongs to automatic generation of a radiotherapy plan in the field of artificial intelligence medicine, and aims to solve the technical problems of low dose distribution precision and low radiotherapy quality in a generated radiotherapy scheme in the prior art. A dose prediction module and a flux prediction module are arranged in a radiotherapy plan generation model and both adopt a U-Net network structure, a projection drawing is used as the input of the dose prediction module, the output of the dose prediction module and the projection drawing are used as the input of the flux prediction module, and when the flux prediction module performs up-sampling, the flux prediction module performs up-sampling on the flux prediction module. A feature map obtained after up-sampling, a feature map of corresponding jump connection in the midway of the flux prediction module and a corresponding feature map generated in the midway of the dose prediction module are spliced and then subjected to convolution processing, so that the model can capture continuous changes of organs, and a finally generated radiotherapy scheme is higher in dose distribution precision and higher in radiotherapy quality.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence medical technology and relates to the generation of radiotherapy plans, and in particular to a method, system, device and medium for generating radiotherapy plans based on ordinary differential equations. Background Art

[0002] Radiation therapy (RT) is an important tumor treatment method, and its main way is to provide targeted radiation to cancer cells while minimizing exposure to healthy organs. Therefore, the development of high-quality radiotherapy plans is crucial in tumor cancer treatment. In the past few decades, intensity-modulated radiation therapy (IMRT) has been widely adopted as a treatment target. IMRT uses adjustable leaves of a multi-leaf collimato (MLC) to adjust the shape and intensity of the radiation beam. In clinical practice, MLC sequences are represented as flux maps, so the flux map can be regarded as the radiotherapy plan itself. However, the traditional method of creating IMRT treatment plans, known as inverse optimization, is a time-consuming and experience-dependent process.

[0003] In order to reduce the time cost of radiotherapy planning and further reduce the impact of human planning habits on treatment results. In recent years, various semi-automated treatment planning methods have been proposed. These methods can be roughly divided into two categories: knowledge-based planning (KBP) methods and deep learning-based dose prediction (DLDP) methods. The KBP method aims to predict the dose-volume histogram (DVH) for new patients based on historical data of patients with similar anatomical features, and use this as a reference to help physicists complete radiotherapy planning. The DLDP method uses a deep neural network trained with clinical cases to predict the dose distribution of new patients.

[0004] In recent years, with the rapid development of artificial intelligence technology, the application of dose prediction methods based on deep learning has become more and more extensive due to its fast calculation speed and high degree of automation.

[0005] The invention patent application with application number 202110142971.X discloses a method for automatically generating radiotherapy plans based on deep learning, which includes: step A, preprocessing the case database to obtain case feature images and case flux maps; step B, using case feature images and case flux maps to train flux map prediction neural networks, and obtaining the optimal flux map prediction neural network through cross-validation; step C, reading and preprocessing the medical digital image storage and communication standard files of patients undergoing radiotherapy to obtain feature images of the patients to be treated; step D, passing the feature images of the patients to be treated to the prediction neural network, obtaining the predicted flux map of the patients to be treated, and sending the predicted flux map of the patients to be treated to the planning system, thereby generating a predicted radiotherapy plan for the patients to be treated. By introducing deep learning of convolutional neural networks to generate radiotherapy plans that can be used directly, the dependence on the personal experience of physicists is reduced, the manual prediction errors are reduced, the time for formulating radiotherapy plans is shortened, and personalized treatment plans are provided.

[0006] Application number 202310869407.7 also discloses a method, system, electronic device and storage medium for generating dose for radiotherapy verification plan, the method comprising: obtaining a case verification plan data set, and obtaining projection information of the planned radiation field through data analysis; using the phantom CT image data, phantom contour structure data, and projection information corresponding to the case verification plan data set as input data of a deep learning model, and the dose distribution map as the output of the deep learning model, and training the deep learning model until a dose calculation network model is obtained; using the dose calculation network model to perform dose calculation on the case information to be calculated, the dose of the patient verification plan can be quickly generated, thereby improving the accuracy of plan verification; automatically calculating the dose through a deep learning model reduces interference from human factors, reduces error rates, and improves work efficiency.

[0007] As in the above-mentioned invention patent application, most of the prior art obtains projection maps and feature maps, and inputs them into the network model to finally obtain the radiotherapy plan. However, in the process of generating the radiotherapy plan, in addition to considering the projection map, or the combination of the projection map and the flux map, it lacks an important reference basis-the dose map, which results in low treatment quality and unsatisfactory radiotherapy effect when performing radiotherapy according to the generated radiotherapy plan. In addition, most of the existing network models use a single U-Net network, which cannot capture the continuous changes of organs, and the dose distribution accuracy in the generated radiotherapy plan is low, and the radiotherapy quality is not high. Summary of the invention

[0008] The purpose of the present invention is to provide a method, system, device and medium for generating a radiotherapy plan based on ordinary differential equations in order to solve the technical problems of low dose distribution accuracy and poor radiotherapy quality in the radiotherapy plans generated in the prior art.

[0009] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions: A method for generating a radiotherapy plan based on ordinary differential equations comprises the following steps: Step 1, obtaining sample image data; Obtain CT outlines, radiotherapy dose maps, and radiotherapy flux maps of the tumor site, and project the CT outlines onto a two-dimensional plane perpendicular to the direction of the radiotherapy rays to obtain a projection map; Step 2, constructing a radiotherapy plan generation model; Constructing a radiotherapy plan generation model, which includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; Step 3, training the radiotherapy plan generation model; Using the sample image data obtained in step 1 to train the radiotherapy plan generation model constructed in step 2; Step 4, radiotherapy plan generation; Obtain a CT outline of the tumor site for which a plan is to be generated, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be tested; input the projection image to be tested into the radiotherapy plan generation model trained in step 3, and the radiotherapy plan generation model outputs a radiotherapy plan.

[0010] Further, in step 2, the dose prediction module includes a first dose convolution submodule, a first dose downsampling submodule, a second dose convolution submodule, a second dose downsampling submodule, a third dose convolution submodule, a third dose downsampling submodule, a fourth dose convolution submodule, a first dose upsampling submodule, a fifth dose convolution submodule, a second dose upsampling submodule, a sixth dose convolution submodule, a third dose upsampling submodule, and a seventh dose convolution submodule, which are arranged in sequence; The projection image is obtained after the first dose convolution submodule, the first dose characteristic map is obtained after the first dose downsampling submodule to obtain the second dose characteristic map, the second dose characteristic map is obtained after the second dose convolution submodule to obtain the third dose characteristic map, the third dose characteristic map is obtained after the second dose downsampling submodule to obtain the fourth dose characteristic map, the fourth dose characteristic map is obtained after the third dose convolution submodule to obtain the fifth dose characteristic map, the fifth dose characteristic map is obtained after the third dose downsampling submodule to obtain the sixth dose characteristic map, the sixth dose characteristic map is obtained after the fourth dose convolution submodule to obtain the seventh dose characteristic map, the seventh dose characteristic map is obtained after the first dose After the dose upsampling submodule, the eighth dose characteristic map is obtained, the eighth dose characteristic map is spliced ​​with the fifth dose characteristic map and passed through the fifth dose convolution submodule to obtain the ninth dose characteristic map, the ninth dose characteristic map is passed through the second dose upsampling submodule to obtain the tenth dose characteristic map, the tenth dose characteristic map is spliced ​​with the third dose characteristic map and passed through the sixth dose convolution submodule to obtain the eleventh dose characteristic map, the eleventh dose characteristic map is passed through the third dose upsampling submodule to obtain the twelfth dose characteristic map, the twelfth dose characteristic map is spliced ​​with the first dose characteristic map and passed through the seventh dose convolution submodule to obtain the dose map.

[0011] Further, in step 2, the flux prediction module includes a first flux convolution submodule, a first flux downsampling submodule, a second flux convolution submodule, a second flux downsampling submodule, a third flux convolution submodule, a third flux downsampling submodule, a fourth flux convolution submodule, a first flux upsampling submodule, a fifth flux convolution submodule, a second flux upsampling submodule, a sixth flux convolution submodule, a third flux upsampling submodule, a seventh flux convolution submodule, an eighth flux convolution submodule, a fourth flux downsampling submodule, a ninth flux convolution submodule, a fifth flux downsampling submodule, a tenth flux convolution submodule, a sixth flux downsampling submodule, and an eleventh flux convolution submodule, which are arranged in sequence; The projection image is subjected to the eighth flux convolution submodule to obtain the thirteenth flux characteristic map, the thirteenth flux characteristic map is subjected to the fourth flux downsampling submodule to obtain the fourteenth flux characteristic map, the fourteenth flux characteristic map is subjected to the ninth flux convolution submodule to obtain the fifteenth flux characteristic map, the fifteenth flux characteristic map is subjected to the fifth flux downsampling submodule to obtain the sixteenth flux characteristic map, the sixteenth flux characteristic map is subjected to the tenth flux convolution submodule to obtain the seventeenth flux characteristic map, the seventeenth flux characteristic map is subjected to the sixth flux downsampling submodule to obtain the eighteenth flux characteristic map, and the eighteenth flux characteristic map is subjected to the eleventh flux convolution submodule to obtain the nineteenth flux characteristic map; The dose map output by the projection map and the dose prediction module is passed through the first flux convolution submodule to obtain a first flux characteristic map, the first flux characteristic map is passed through the first flux downsampling submodule to obtain a second flux characteristic map, the second flux characteristic map is passed through the second flux convolution submodule to obtain a third flux characteristic map, the third flux characteristic map is passed through the second flux downsampling submodule to obtain a fourth flux characteristic map, the fourth flux characteristic map is passed through the third flux convolution submodule to obtain a fifth flux characteristic map, the fifth flux characteristic map is passed through the third flux downsampling submodule to obtain a sixth flux characteristic map, the sixth flux characteristic map is passed through the fourth flux convolution submodule to obtain a seventh flux characteristic map, the seventh flux characteristic map is added to the nineteenth flux characteristic map and passed through the first flux upsampling submodule to obtain the ninth flux characteristic map. Eight flux characteristic graphs, the eighth flux characteristic graph, the fifth flux characteristic graph and the fifth dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the fifth flux convolution submodule to obtain a ninth flux characteristic graph, the ninth flux characteristic graph is passed through the second flux upsampling submodule to obtain a tenth flux characteristic graph, the tenth flux characteristic graph, the third flux characteristic graph and the third dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the sixth flux convolution submodule to obtain an eleventh flux characteristic graph, the eleventh flux characteristic graph is passed through the third flux upsampling submodule to obtain a twelfth flux characteristic graph, the twelfth flux characteristic graph, the first flux characteristic graph and the first dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the seventh flux convolution submodule to obtain a flux graph.

[0012] Furthermore, the dose prediction module and the flux prediction module are both provided with a NODE module and a 1*1 convolutional layer. The feature map output by the seventh dose convolution submodule of the dose prediction module is sequentially passed through the NODE module and the 1*1 convolutional layer to output the dose map. The feature map output by the seventh flux convolution submodule of the flux prediction module is sequentially passed through the NODE module and the 1*1 convolutional layer to output the flux map.

[0013] Furthermore, the ordinary differential equation used by the NODE module is:

[0014]

[0015] in, Indicates input, Represents the state value of the model, represents a learnable function, and Both represent learnable parameters.

[0016] Furthermore, the trained radiotherapy plan generation model was tested using the test set in the sample image data. The evaluation indicators in the test process included MAE, MSE, and PSNR. The calculation method of MAE indicator is:

[0017] The calculation method of MSE indicator is:

[0018] The calculation method of PSNR indicator is:

[0019] in, Indicates pixels, Represents the total number of pixels, Indicates the first The intensity of a pixel, Indicates the label flux map The intensity of a pixel, Indicates the maximum value of the pixel in the entire flux map.

[0020] A radiotherapy plan generation system based on ordinary differential equations, comprising: The sample image data acquisition module is used to obtain the CT outline, radiotherapy dose map and radiotherapy flux map of the tumor site, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection map; A radiotherapy plan generation model construction module is used to construct a radiotherapy plan generation model, and the radiotherapy plan generation model includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; A radiotherapy plan generation model training module is used to train the radiotherapy plan generation model constructed by the radiotherapy plan generation model construction module using the sample image data acquired by the sample image data acquisition module; The radiotherapy plan generation module is used to obtain the CT outline of the tumor site for which the plan is to be generated, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be tested; the projection image to be tested is input into the radiotherapy plan generation model trained by the radiotherapy plan generation model training module, and the radiotherapy plan generation model outputs the radiotherapy plan.

[0021] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0022] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above method.

[0023] The beneficial effects of the present invention are as follows: 1. In the present invention, a dose prediction module and a flux prediction module are set in the radiotherapy plan generation model, both of which adopt the U-Net network structure, and the projection map is used as the input of the dose prediction module, and the output of the dose prediction module and the projection map are used as the input of the flux prediction module. When the flux prediction module is upsampled, the feature map obtained after upsampling, the feature map corresponding to the jump connection in the middle of the flux prediction module, and the corresponding feature map generated in the middle of the dose prediction module are spliced ​​and then convolution processing is performed, so that the model can capture the continuous changes of the organ, and the dose distribution accuracy in the finally generated radiotherapy plan is higher, and the radiotherapy quality is higher.

[0024] 2. In the present invention, by introducing a neural network model based on neural ordinary differential equations, fully automated radiotherapy plan generation can be achieved, reducing the reliance on manual intervention in traditional methods, improving the efficiency and consistency of radiotherapy plan formulation, and reducing the workload of manual outlining and plan optimization.

[0025] 3. In the present invention, the characteristics of neural ordinary differential equations are utilized to better capture the continuous changes of tumor target areas and organs at risk in CT images, thereby generating a more accurate dose distribution; compared with traditional radiotherapy planning methods based on static images, the present application can automatically take into account the complexity of the anatomical structure of individual patients and provide personalized treatment plans.

[0026] 4. In the present invention, the outlined image is projected through a ray tracing algorithm to convert the three-dimensional data into a two-dimensional representation, thereby improving the computational efficiency while retaining key anatomical information; this data preprocessing method can better simulate the projection effect during actual radiotherapy, making the plan generated by the model more in line with clinical needs.

[0027] 5. In the present invention, the time for plan generation is significantly reduced through a fully automated process, so that the radiotherapy plan can be completed in a shorter time, thereby providing treatment to patients more quickly.

[0028] 6. In the present invention, through the training and evaluation process of the model, the dose limitation of the organs at risk can be automatically taken into account when generating the radiotherapy plan, thereby maximizing the protection of healthy tissues, optimizing the treatment dose of tumors, and improving the safety and effectiveness of treatment.

[0029] 7. The fully automated method of the present invention can reduce the dependence on the operator's experience, reduce the impact of human factors on the quality of radiotherapy plans, and make the formulation of radiotherapy plans more standardized and normalized.

[0030] 8. In the present invention, by introducing the NODE module with excellent ability to process complex and nonlinear features into the two prediction modules, the NODE module can capture complex dynamic behaviors, significantly enhance its expression ability in processing nonlinear systems, and improve the accuracy of dose distribution recognition. The dose distribution in the final generated radiotherapy plan is more accurate and the radiotherapy quality is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 is a schematic diagram of a radiotherapy plan generation model in the present invention; Among them, concatenation means merging two feature maps into one feature map along a specified dimension, so that its feature information is expanded in this dimension, thereby retaining and integrating all the information of the two sets of features; multiplication means performing element-by-element multiplication operations on the corresponding position elements of the two feature maps; addition means performing element-by-element addition operations on the corresponding position elements of the two feature maps, thereby fusing the two sets of feature information and realizing the superposition and comprehensive expression of features; Figure 3 is a schematic diagram of a dose prediction module in a radiotherapy plan generation model of the present invention; Figure 4 is a schematic diagram of a flux prediction module in a radiotherapy plan generation model of the present invention; Figure 5 It is a schematic diagram of projection of target area organs at risk in the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0033] Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0034] Example 1 This embodiment provides a method for generating a radiotherapy plan based on ordinary differential equations, which is used to automatically generate a radiotherapy plan with higher dose distribution accuracy and higher radiotherapy quality. Figure 1 As shown, it includes the following steps: Step 1, obtaining sample image data; Obtain CT outlines, radiotherapy dose maps, and radiotherapy flux maps of the tumor site, and project the CT outlines onto a two-dimensional plane perpendicular to the direction of the radiotherapy rays to obtain a projection map.

[0035] Real case data were collected from West China Hospital of Sichuan University as sample data, including patients’ CT outlines, radiotherapy dose maps, and radiotherapy flux maps, and the above data were organized according to the architecture.

[0036] Acquisition of CT data: The physicist obtains CT image data by using a CT machine to scan the patient's tumor area. One scan will generate multiple continuous CT images.

[0037] Organ delineation data acquisition: The physicist delineates the target area and organs at risk on the acquired CT images according to the pathological characteristics of the case.

[0038] Flux map data acquisition: The physicist uses the radiotherapy planning system to make the radiotherapy plan according to the outline structure of the pathology.

[0039] Dose data acquisition: The physicist simulates the radiotherapy process of the prepared radiotherapy plan to obtain dose data.

[0040] All types of data are organized with case as the primary key.

[0041] In order to improve the efficiency of model training and the accuracy of plan generation, the sample data needs to be processed to make it more suitable for the input format of the neural network. The specific data processing process is as follows: like Figure 5 As shown in the figure, the ray tracing algorithm is first used to project the three-dimensional outline of the target area and organs at risk in the CT image onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image. This projection result simulates the actual projection effect during radiotherapy, simplifies the processing difficulty of the three-dimensional image, and retains important structural information.

[0042] The projection images are then normalized to ensure that all data are of the same scale. This step can reduce the model’s sensitivity to data segmentation and allow it to focus more on feature extraction of target areas and organs at risk.

[0043] Finally, the standardized dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used to train the model and optimize the model parameters; the validation set is used to evaluate and adjust the hyperparameters; and the test set is used for the final model evaluation to verify its performance in practical applications.

[0044] Step 2, constructing a radiotherapy plan generation model; Constructing a radiotherapy plan generation model, which includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; The specific working process of the radiotherapy plan generation model, such as Figure 2 shown.

[0045] like Figure 3 As shown, the dose prediction module adopts a U-Net network structure, which includes a first dose convolution submodule, a first dose downsampling submodule, a second dose convolution submodule, a second dose downsampling submodule, a third dose convolution submodule, a third dose downsampling submodule, a fourth dose convolution submodule, a first dose upsampling submodule, a fifth dose convolution submodule, a second dose upsampling submodule, a sixth dose convolution submodule, a third dose upsampling submodule, a seventh dose convolution submodule, a NODE module, and a 1*1 convolution layer, which are arranged in sequence.

[0046] The projection image is obtained after the first dose convolution submodule, the first dose characteristic map is obtained after the first dose downsampling submodule, the second dose characteristic map is obtained after the second dose convolution submodule, the third dose characteristic map is obtained after the second dose downsampling submodule, the fourth dose characteristic map is obtained after the third dose convolution submodule, the fifth dose characteristic map is obtained after the third dose downsampling submodule, the sixth dose characteristic map is obtained after the fourth dose convolution submodule, the seventh dose characteristic map is obtained after the first dose upsampling submodule, the eighth dose characteristic map is spliced ​​with the fifth dose characteristic map (that is, the fifth dose characteristic map is connected to the eighth dose characteristic map by a jump connection). The dose characteristic map is connected and spliced) and passes through the fifth dose convolution submodule to obtain the ninth dose characteristic map, the ninth dose characteristic map passes through the second dose upsampling submodule to obtain the tenth dose characteristic map, the tenth dose characteristic map is spliced ​​with the third dose characteristic map (that is, the third dose characteristic map is connected and spliced ​​with the tenth dose characteristic map by a jump connection) and passes through the sixth dose convolution submodule to obtain the eleventh dose characteristic map, the eleventh dose characteristic map is passed through the third dose upsampling submodule to obtain the twelfth dose characteristic map, the twelfth dose characteristic map is spliced ​​with the first dose characteristic map (that is, the first dose characteristic map is connected and spliced ​​with the twelfth dose characteristic map by a jump connection) and passes through the seventh dose convolution submodule, the NODE module, and the 1*1 convolution layer to obtain the dose map.

[0047] Among them, the first dose convolution submodule, the second dose convolution submodule, the third dose convolution submodule, the fourth dose convolution submodule, the fifth dose convolution submodule and the sixth dose convolution submodule all include two layers of 3*3 convolution layers, and the seventh dose convolution submodule includes one layer of 3*3 convolution layer.

[0048] In the dose prediction module, the projection map undergoes three rounds of encoding, gradually reducing the projection map size from 192*192 to a 24*24 feature map, while increasing the number of channels to 1024. The focus of this encoding process is to capture abstract latent features of different scales; subsequently, the final output dose map is generated through three decoding stages, each of which includes upsampling and convolution operations to reconstruct the dose map based on the encoded features, and jump connections help retrieve features of the same size from the encoder, thereby facilitating the reconstruction process; finally, the NODE module is used for nonlinear remapping to obtain the predicted dose map.

[0049] like Figure 4 As shown, the flux prediction module also adopts the U-Net network structure, which includes a first flux convolution submodule, a first flux downsampling submodule, a second flux convolution submodule, a second flux downsampling submodule, a third flux convolution submodule, a third flux downsampling submodule, a fourth flux convolution submodule, a first flux upsampling submodule, a fifth flux convolution submodule, a second flux upsampling submodule, a sixth flux convolution submodule, a third flux upsampling submodule, a seventh flux convolution submodule, an eighth flux convolution submodule, a fourth flux downsampling submodule, a ninth flux convolution submodule, a fifth flux downsampling submodule, a tenth flux convolution submodule, a sixth flux downsampling submodule, and an eleventh flux convolution submodule, which are arranged in sequence.

[0050] The flux prediction module has two inputs. One is to input the projection image into the eighth flux convolution submodule of the flux prediction module. The other is to splice the projection image and the dose image output by the dose prediction module and then input them into the first flux convolution submodule of the flux prediction module.

[0051] The projection image is passed through the eighth flux convolution submodule to obtain the thirteenth flux feature map, the thirteenth flux feature map is passed through the fourth flux downsampling submodule to obtain the fourteenth flux feature map, the fourteenth flux feature map is passed through the ninth flux convolution submodule to obtain the fifteenth flux feature map, the fifteenth flux feature map is passed through the fifth flux downsampling submodule to obtain the sixteenth flux feature map, the sixteenth flux feature map is passed through the tenth flux convolution submodule to obtain the seventeenth flux feature map, the seventeenth flux feature map is passed through the sixth flux downsampling submodule to obtain the eighteenth flux feature map, and the eighteenth flux feature map is passed through the eleventh flux convolution submodule to obtain the nineteenth flux feature map.

[0052] The projection image and the dose image output by the dose prediction module (i.e., the two images are spliced ​​and input) are passed through the first flux convolution submodule to obtain the first flux characteristic image, the first flux characteristic image is passed through the first flux downsampling submodule to obtain the second flux characteristic image, the second flux characteristic image is passed through the second flux convolution submodule to obtain the third flux characteristic image, the third flux characteristic image is passed through the second flux downsampling submodule to obtain the fourth flux characteristic image, the fourth flux characteristic image is passed through the third flux convolution submodule to obtain the fifth flux characteristic image, the fifth flux characteristic image is passed through the third flux downsampling submodule to obtain the sixth flux characteristic image, the sixth flux characteristic image is passed through the fourth flux convolution submodule to obtain the seventh flux characteristic image, the seventh flux characteristic image is added to the nineteenth flux characteristic image and passed through the first flux upsampling submodule to obtain the eighth flux characteristic image, the eighth flux characteristic image, the fifth flux characteristic image and the fifth dose characteristic image outputted from the dose prediction module are spliced ​​(i.e., the fifth flux characteristic image, the fifth dose characteristic image). The flux characteristic map is connected and spliced ​​with the eighth flux characteristic map by a jump connection) and passes through the fifth flux convolution submodule to obtain a ninth flux characteristic map. The ninth flux characteristic map is passed through the second flux upsampling submodule to obtain a tenth flux characteristic map. The tenth flux characteristic map, the third flux characteristic map and the third dose characteristic map outputted midway by the dose prediction module are spliced ​​(that is, the third flux characteristic map and the third dose characteristic map are connected and spliced ​​with the tenth flux characteristic map by a jump connection) and pass through the sixth flux convolution submodule to obtain an eleventh flux characteristic map. The eleventh flux characteristic map is passed through the third flux upsampling submodule to obtain a twelfth flux characteristic map. The twelfth flux characteristic map, the first flux characteristic map and the first dose characteristic map outputted midway by the dose prediction module are spliced ​​(that is, the first flux characteristic map and the first dose characteristic map are connected and spliced ​​with the twelfth flux characteristic map by a jump connection) and pass through the seventh flux convolution submodule to obtain a flux map.

[0053] The structure of the flux prediction module is consistent with that of the dose prediction module. The difference is that it takes the dose map predicted by the dose prediction module and the projection map of the target area's organs at risk as input at the same time. After a series of encoding and decoding operations, the feature map obtained is sent to the NODE module for nonlinear remapping to obtain the predicted flux map, that is, the radiotherapy plan.

[0054] The ordinary differential equation used by the NODE module is:

[0055]

[0056] in, Indicates input, Represents the state value of the model, represents a learnable function, and Both represent learnable parameters.

[0057] Step 3, training the radiotherapy plan generation model; The sample image data obtained in step 1 is used to train the radiotherapy plan generation model constructed in step 2.

[0058] The training set is input into the radiotherapy plan generation model, and forward calculation and back propagation are performed in sequence, which are repeated many times until the model converges. The training can be stopped when the output on the validation set is the minimum value among multiple trainings, thereby obtaining the optimal model.

[0059] The model can be trained using existing technologies, and the parameter settings, loss functions, etc. in the training process can be directly applied to existing technologies without the need for creative work.

[0060] After training and verifying the model, the model needs to be tested. Use the test set to test the model. The evaluation indicators in the test process can be selected from the three indicators of MAE, MSE, and PSNR. The calculation method of MAE indicator is:

[0061] The calculation method of MSE indicator is:

[0062] The calculation method of PSNR indicator is:

[0063] in, Indicates pixels, Represents the total number of pixels, Indicates the first The intensity of a pixel, Indicates the label flux map The intensity of a pixel, Indicates the maximum value of the pixel in the entire flux map.

[0064] Step 4, radiotherapy plan generation; Obtain a CT outline of the tumor site for which a plan is to be generated, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be tested; input the projection image to be tested into the radiotherapy plan generation model trained in step 3, and the radiotherapy plan generation model outputs a radiotherapy plan.

[0065] Example 2 This embodiment provides a radiotherapy plan generation system based on ordinary differential equations, which is used to automatically generate a radiotherapy plan with higher dose distribution accuracy and higher radiotherapy quality. It includes: The sample image data acquisition module is used to obtain the CT outline, radiotherapy dose map and radiotherapy flux map of the tumor site, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection map.

[0066] Real case data were collected from West China Hospital of Sichuan University as sample data, including patients’ CT outlines, radiotherapy dose maps, and radiotherapy flux maps, and the above data were organized according to the architecture.

[0067] Acquisition of CT data: The physicist obtains CT image data by using a CT machine to scan the patient's tumor area. One scan will generate multiple continuous CT images.

[0068] Organ delineation data acquisition: The physicist delineates the target area and organs at risk on the acquired CT images according to the pathological characteristics of the case.

[0069] Flux map data acquisition: The physicist uses the radiotherapy planning system to make the radiotherapy plan according to the outline structure of the pathology.

[0070] Dose data acquisition: The physicist simulates the radiotherapy process of the prepared radiotherapy plan to obtain dose data.

[0071] All types of data are organized with case as the primary key.

[0072] In order to improve the efficiency of model training and the accuracy of plan generation, the sample data needs to be processed to make it more suitable for the input format of the neural network. The specific data processing process is as follows: like Figure 5 As shown in the figure, the ray tracing algorithm is first used to project the three-dimensional outline of the target area and organs at risk in the CT image onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image. This projection result simulates the actual projection effect during radiotherapy, simplifies the processing difficulty of the three-dimensional image, and retains important structural information.

[0073] The projection images are then normalized to ensure that all data are of the same scale. This step can reduce the model’s sensitivity to data segmentation and allow it to focus more on feature extraction of target areas and organs at risk.

[0074] Finally, the standardized dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used to train the model and optimize the model parameters; the validation set is used to evaluate and adjust the hyperparameters; and the test set is used for the final model evaluation to verify its performance in practical applications.

[0075] A radiotherapy plan generation model construction module is used to construct a radiotherapy plan generation model, and the radiotherapy plan generation model includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; The specific working process of the radiotherapy plan generation model, such as Figure 2 shown.

[0076] like Figure 3 As shown, the dose prediction module adopts a U-Net network structure, which includes a first dose convolution submodule, a first dose downsampling submodule, a second dose convolution submodule, a second dose downsampling submodule, a third dose convolution submodule, a third dose downsampling submodule, a fourth dose convolution submodule, a first dose upsampling submodule, a fifth dose convolution submodule, a second dose upsampling submodule, a sixth dose convolution submodule, a third dose upsampling submodule, a seventh dose convolution submodule, a NODE module, and a 1*1 convolution layer, which are arranged in sequence.

[0077] The projection image is obtained after the first dose convolution submodule, the first dose characteristic map is obtained after the first dose downsampling submodule, the second dose characteristic map is obtained after the second dose convolution submodule, the third dose characteristic map is obtained after the second dose downsampling submodule, the fourth dose characteristic map is obtained after the third dose convolution submodule, the fifth dose characteristic map is obtained after the third dose downsampling submodule, the sixth dose characteristic map is obtained after the fourth dose convolution submodule, the seventh dose characteristic map is obtained after the first dose upsampling submodule, the eighth dose characteristic map is spliced ​​with the fifth dose characteristic map (that is, the fifth dose characteristic map is connected to the eighth dose characteristic map by a jump connection). The dose characteristic map is connected and spliced) and passes through the fifth dose convolution submodule to obtain the ninth dose characteristic map, the ninth dose characteristic map passes through the second dose upsampling submodule to obtain the tenth dose characteristic map, the tenth dose characteristic map is spliced ​​with the third dose characteristic map (that is, the third dose characteristic map is connected and spliced ​​with the tenth dose characteristic map by a jump connection) and passes through the sixth dose convolution submodule to obtain the eleventh dose characteristic map, the eleventh dose characteristic map is passed through the third dose upsampling submodule to obtain the twelfth dose characteristic map, the twelfth dose characteristic map is spliced ​​with the first dose characteristic map (that is, the first dose characteristic map is connected and spliced ​​with the twelfth dose characteristic map by a jump connection) and passes through the seventh dose convolution submodule, the NODE module, and the 1*1 convolution layer to obtain the dose map.

[0078] Among them, the first dose convolution submodule, the second dose convolution submodule, the third dose convolution submodule, the fourth dose convolution submodule, the fifth dose convolution submodule and the sixth dose convolution submodule all include two layers of 3*3 convolution layers, and the seventh dose convolution submodule includes one layer of 3*3 convolution layer.

[0079] In the dose prediction module, the projection map undergoes three rounds of encoding, gradually reducing the projection map size from 192*192 to a 24*24 feature map, while increasing the number of channels to 1024. The focus of this encoding process is to capture abstract latent features of different scales; subsequently, the final output dose map is generated through three decoding stages, each of which includes upsampling and convolution operations to reconstruct the dose map based on the encoded features, and jump connections help retrieve features of the same size from the encoder, thereby facilitating the reconstruction process; finally, the NODE module is used for nonlinear remapping to obtain the predicted dose map.

[0080] like Figure 4 As shown, the flux prediction module also adopts the U-Net network structure, which includes a first flux convolution submodule, a first flux downsampling submodule, a second flux convolution submodule, a second flux downsampling submodule, a third flux convolution submodule, a third flux downsampling submodule, a fourth flux convolution submodule, a first flux upsampling submodule, a fifth flux convolution submodule, a second flux upsampling submodule, a sixth flux convolution submodule, a third flux upsampling submodule, a seventh flux convolution submodule, an eighth flux convolution submodule, a fourth flux downsampling submodule, a ninth flux convolution submodule, a fifth flux downsampling submodule, a tenth flux convolution submodule, a sixth flux downsampling submodule, and an eleventh flux convolution submodule, which are arranged in sequence.

[0081] The flux prediction module has two inputs. One is to input the projection image into the eighth flux convolution submodule of the flux prediction module. The other is to splice the projection image and the dose image output by the dose prediction module and then input them into the first flux convolution submodule of the flux prediction module.

[0082] The projection image is passed through the eighth flux convolution submodule to obtain the thirteenth flux feature map, the thirteenth flux feature map is passed through the fourth flux downsampling submodule to obtain the fourteenth flux feature map, the fourteenth flux feature map is passed through the ninth flux convolution submodule to obtain the fifteenth flux feature map, the fifteenth flux feature map is passed through the fifth flux downsampling submodule to obtain the sixteenth flux feature map, the sixteenth flux feature map is passed through the tenth flux convolution submodule to obtain the seventeenth flux feature map, the seventeenth flux feature map is passed through the sixth flux downsampling submodule to obtain the eighteenth flux feature map, and the eighteenth flux feature map is passed through the eleventh flux convolution submodule to obtain the nineteenth flux feature map.

[0083] The projection image and the dose image output by the dose prediction module (i.e., the two images are spliced ​​and input) are passed through the first flux convolution submodule to obtain the first flux characteristic image, the first flux characteristic image is passed through the first flux downsampling submodule to obtain the second flux characteristic image, the second flux characteristic image is passed through the second flux convolution submodule to obtain the third flux characteristic image, the third flux characteristic image is passed through the second flux downsampling submodule to obtain the fourth flux characteristic image, the fourth flux characteristic image is passed through the third flux convolution submodule to obtain the fifth flux characteristic image, the fifth flux characteristic image is passed through the third flux downsampling submodule to obtain the sixth flux characteristic image, the sixth flux characteristic image is passed through the fourth flux convolution submodule to obtain the seventh flux characteristic image, the seventh flux characteristic image is added to the nineteenth flux characteristic image and passed through the first flux upsampling submodule to obtain the eighth flux characteristic image, the eighth flux characteristic image, the fifth flux characteristic image and the fifth dose characteristic image outputted from the dose prediction module are spliced ​​(i.e., the fifth flux characteristic image, the fifth dose characteristic image). The flux characteristic map is connected and spliced ​​with the eighth flux characteristic map by a jump connection) and passes through the fifth flux convolution submodule to obtain a ninth flux characteristic map. The ninth flux characteristic map is passed through the second flux upsampling submodule to obtain a tenth flux characteristic map. The tenth flux characteristic map, the third flux characteristic map and the third dose characteristic map outputted midway by the dose prediction module are spliced ​​(that is, the third flux characteristic map and the third dose characteristic map are connected and spliced ​​with the tenth flux characteristic map by a jump connection) and pass through the sixth flux convolution submodule to obtain an eleventh flux characteristic map. The eleventh flux characteristic map is passed through the third flux upsampling submodule to obtain a twelfth flux characteristic map. The twelfth flux characteristic map, the first flux characteristic map and the first dose characteristic map outputted midway by the dose prediction module are spliced ​​(that is, the first flux characteristic map and the first dose characteristic map are connected and spliced ​​with the twelfth flux characteristic map by a jump connection) and pass through the seventh flux convolution submodule to obtain a flux map.

[0084] The structure of the flux prediction module is consistent with that of the dose prediction module. The difference is that it takes the dose map predicted by the dose prediction module and the projection map of the target area's organs at risk as input at the same time. After a series of encoding and decoding operations, the feature map obtained is sent to the NODE module for nonlinear remapping to obtain the predicted flux map, that is, the radiotherapy plan.

[0085] The ordinary differential equation used by the NODE module is:

[0086]

[0087] in, Indicates input, Represents the state value of the model, represents a learnable function, and Both represent learnable parameters.

[0088] The radiotherapy plan generation model training module is used to train the radiotherapy plan generation model constructed by the radiotherapy plan generation model construction module using the sample image data acquired by the sample image data acquisition module.

[0089] The training set is input into the radiotherapy plan generation model, and forward calculation and back propagation are performed in sequence, which are repeated many times until the model converges. The training can be stopped when the output on the validation set is the minimum value among multiple trainings, thereby obtaining the optimal model.

[0090] The model can be trained using existing technologies, and the parameter settings, loss functions, etc. in the training process can be directly applied to existing technologies without the need for creative work.

[0091] The radiotherapy plan generation module is used to obtain the CT outline of the tumor site for which the plan is to be generated, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be tested; the projection image to be tested is input into the radiotherapy plan generation model trained by the radiotherapy plan generation model training module, and the radiotherapy plan generation model outputs the radiotherapy plan.

[0092] Example 3 A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a radiotherapy plan generation method based on ordinary differential equations.

[0093] The computer device may be a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with the user through a keyboard, a mouse, a remote control, a touch pad, or a voice control device.

[0094] The memory includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or D interface display memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the memory can also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the radiotherapy plan generation method based on ordinary differential equations, etc. In addition, the memory can also be used to temporarily store various types of data that have been output or are to be output.

[0095] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data, such as running the program code of the radiotherapy plan generation method based on ordinary differential equations.

[0096] Example 4 A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of a method for generating a radiotherapy plan based on ordinary differential equations.

[0097] The computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor so that the at least one processor executes the steps of the above-mentioned method for generating a radiotherapy plan based on ordinary differential equations.

[0098] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server or network device, etc.) to execute the radiotherapy plan generation method based on ordinary differential equations described in the embodiment of the present application.

Claims

1. A method for generating a radiotherapy plan based on ordinary differential equations, characterized in that: The steps include: Step 1, obtaining sample image data; Obtain CT outlines, radiotherapy dose maps, and radiotherapy flux maps of the tumor site, and project the CT outlines onto a two-dimensional plane perpendicular to the direction of the radiotherapy rays to obtain a projection map; Step 2, constructing a radiotherapy plan generation model; Constructing a radiotherapy plan generation model, which includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; Step 3, training the radiotherapy plan generation model; Using the sample image data obtained in step 1 to train the radiotherapy plan generation model constructed in step 2; Step 4, radiotherapy plan generation; Obtaining a CT outline of the tumor site for which a plan is to be generated, and projecting the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be measured; The projection image to be tested is input into the radiotherapy plan generation model trained in step 3, and the radiotherapy plan generation model outputs the radiotherapy plan.

2. A method for generating a radiotherapy plan based on ordinary differential equations as claimed in claim 1, characterized in that: In step 2, the dose prediction module includes a first dose convolution submodule, a first dose downsampling submodule, a second dose convolution submodule, a second dose downsampling submodule, a third dose convolution submodule, a third dose downsampling submodule, a fourth dose convolution submodule, a first dose upsampling submodule, a fifth dose convolution submodule, a second dose upsampling submodule, a sixth dose convolution submodule, a third dose upsampling submodule, and a seventh dose convolution submodule, which are arranged in sequence; The projection image is obtained after the first dose convolution submodule, the first dose characteristic map is obtained after the first dose downsampling submodule to obtain the second dose characteristic map, the second dose characteristic map is obtained after the second dose convolution submodule to obtain the third dose characteristic map, the third dose characteristic map is obtained after the second dose downsampling submodule to obtain the fourth dose characteristic map, the fourth dose characteristic map is obtained after the third dose convolution submodule to obtain the fifth dose characteristic map, the fifth dose characteristic map is obtained after the third dose downsampling submodule to obtain the sixth dose characteristic map, the sixth dose characteristic map is obtained after the fourth dose convolution submodule to obtain the seventh dose characteristic map, the seventh dose characteristic map is obtained after the first dose After the dose upsampling submodule, the eighth dose characteristic map is obtained, the eighth dose characteristic map is spliced ​​with the fifth dose characteristic map and passed through the fifth dose convolution submodule to obtain the ninth dose characteristic map, the ninth dose characteristic map is passed through the second dose upsampling submodule to obtain the tenth dose characteristic map, the tenth dose characteristic map is spliced ​​with the third dose characteristic map and passed through the sixth dose convolution submodule to obtain the eleventh dose characteristic map, the eleventh dose characteristic map is passed through the third dose upsampling submodule to obtain the twelfth dose characteristic map, the twelfth dose characteristic map is spliced ​​with the first dose characteristic map and passed through the seventh dose convolution submodule to obtain the dose map.

3. A method for generating a radiotherapy plan based on ordinary differential equations as claimed in claim 1, characterized in that: In step 2, the flux prediction module includes a first flux convolution submodule, a first flux downsampling submodule, a second flux convolution submodule, a second flux downsampling submodule, a third flux convolution submodule, a third flux downsampling submodule, a fourth flux convolution submodule, a first flux upsampling submodule, a fifth flux convolution submodule, a second flux upsampling submodule, a sixth flux convolution submodule, a third flux upsampling submodule, a seventh flux convolution submodule, an eighth flux convolution submodule, a fourth flux downsampling submodule, a ninth flux convolution submodule, a fifth flux downsampling submodule, a tenth flux convolution submodule, a sixth flux downsampling submodule, and an eleventh flux convolution submodule, which are arranged in sequence; The projection image is subjected to the eighth flux convolution submodule to obtain the thirteenth flux characteristic map, the thirteenth flux characteristic map is subjected to the fourth flux downsampling submodule to obtain the fourteenth flux characteristic map, the fourteenth flux characteristic map is subjected to the ninth flux convolution submodule to obtain the fifteenth flux characteristic map, the fifteenth flux characteristic map is subjected to the fifth flux downsampling submodule to obtain the sixteenth flux characteristic map, the sixteenth flux characteristic map is subjected to the tenth flux convolution submodule to obtain the seventeenth flux characteristic map, the seventeenth flux characteristic map is subjected to the sixth flux downsampling submodule to obtain the eighteenth flux characteristic map, and the eighteenth flux characteristic map is subjected to the eleventh flux convolution submodule to obtain the nineteenth flux characteristic map; The dose map output by the projection map and the dose prediction module is passed through the first flux convolution submodule to obtain a first flux characteristic map, the first flux characteristic map is passed through the first flux downsampling submodule to obtain a second flux characteristic map, the second flux characteristic map is passed through the second flux convolution submodule to obtain a third flux characteristic map, the third flux characteristic map is passed through the second flux downsampling submodule to obtain a fourth flux characteristic map, the fourth flux characteristic map is passed through the third flux convolution submodule to obtain a fifth flux characteristic map, the fifth flux characteristic map is passed through the third flux downsampling submodule to obtain a sixth flux characteristic map, the sixth flux characteristic map is passed through the fourth flux convolution submodule to obtain a seventh flux characteristic map, the seventh flux characteristic map is added to the nineteenth flux characteristic map and passed through the first flux upsampling submodule to obtain the ninth flux characteristic map. Eight flux characteristic graphs, the eighth flux characteristic graph, the fifth flux characteristic graph and the fifth dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the fifth flux convolution submodule to obtain a ninth flux characteristic graph, the ninth flux characteristic graph is passed through the second flux upsampling submodule to obtain a tenth flux characteristic graph, the tenth flux characteristic graph, the third flux characteristic graph and the third dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the sixth flux convolution submodule to obtain an eleventh flux characteristic graph, the eleventh flux characteristic graph is passed through the third flux upsampling submodule to obtain a twelfth flux characteristic graph, the twelfth flux characteristic graph, the first flux characteristic graph and the first dose characteristic graph outputted from the dose prediction module are spliced ​​and passed through the seventh flux convolution submodule to obtain a flux graph.

4. The method for generating a radiotherapy plan based on ordinary differential equations according to claim 1, characterized in that: Both the dose prediction module and the flux prediction module are provided with a NODE module and a 1*1 convolutional layer. The feature map output by the seventh dose convolution submodule of the dose prediction module passes through the NODE module and the 1*1 convolutional layer in sequence to output the dose map. The feature map output by the seventh flux convolution submodule of the flux prediction module passes through the NODE module and the 1*1 convolutional layer in sequence to output the flux map.

5. A method for generating a radiotherapy plan based on ordinary differential equations as claimed in claim 4, characterized in that: The ordinary differential equation used by the NODE module is: in, Indicates input, Represents the state value of the model, represents a learnable function, and Both represent learnable parameters.

6. The method for generating a radiotherapy plan based on ordinary differential equations according to claim 1, characterized in that: The trained radiotherapy plan generation model was tested using the test set in the sample image data. The evaluation indicators during the test included MAE, MSE, and PSNR. The calculation method of MAE indicator is: The calculation method of MSE indicator is: The calculation method of PSNR indicator is: in, Indicates pixels, Represents the total number of pixels, Indicates the first The intensity of a pixel, Indicates the label flux map The intensity of a pixel, Indicates the maximum value of the pixel in the entire flux map.

7. A radiotherapy plan generation system based on ordinary differential equations, characterized in that: include: The sample image data acquisition module is used to obtain the CT outline, radiotherapy dose map and radiotherapy flux map of the tumor site, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection map; A radiotherapy plan generation model construction module is used to construct a radiotherapy plan generation model, and the radiotherapy plan generation model includes a dose prediction module and a flux prediction module; In the dose prediction module, the projection image is downsampled and upsampled to output the dose map; In the flux prediction module, the projection map and the dose map are downsampled, the projection map is spliced ​​after downsampling, and then the flux map is output after upsampling; A radiotherapy plan generation model training module is used to train the radiotherapy plan generation model constructed by the radiotherapy plan generation model construction module using the sample image data acquired by the sample image data acquisition module; A radiotherapy plan generation module is used to obtain a CT outline of the tumor site for which a plan is to be generated, and project the CT outline onto a two-dimensional plane perpendicular to the direction of the radiotherapy ray to obtain a projection image to be measured; The projection image to be tested is input into the radiotherapy plan generation model trained by the radiotherapy plan generation model training module, and the radiotherapy plan generation model outputs the radiotherapy plan.

8. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 6.

Images