Fluence map estimation method and estimation model generation method, learned model, fluence map estimation system, control program, and recording medium

Abstract

To generate a fluence map having a high possibility of pass in an exposure dose test.SOLUTION: A fluence map estimation method in a radiation therapy includes: an acquisition step (S3) of acquiring a first dose distribution map of a first patient being a target of the radiation therapy; and estimation steps (S4, S5) of inputting the first dose distribution map to a learned model that has performed machine learning using learning data in which a specific fluence map whose exposure dose test result was preferable and a corresponding specific dose distribution map in a second patient being a past target of radiation therapy are associated with each other and outputting an estimation result of a first fluence map.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a fluence map estimation method, an estimation model generation method, a trained model, a fluence map estimation system, a control program, and a recording medium. [Background technology]

[0002] In recent years, radiation therapy has attracted attention for methods that can concentrate a dose on a treatment target area and reduce the dose to surrounding normal areas, such as Intensity Modulated Radiation Therapy (IMRT). Because these methods involve complex irradiation techniques, it is important to create a highly accurate irradiation plan. For example, Non-Patent Document 1 proposes a method that uses a machine learning model to create a dose distribution map, which is one step in creating an irradiation plan.

[0003] In radiation therapy, a dose test is usually performed on the prepared irradiation plan before the actual irradiation to the patient. In the dose test, it is confirmed whether the error between the target dose in the prepared irradiation plan and the measured dose when actually irradiated based on the irradiation plan is within a tolerance. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Kadoya et al., Journal of Radiation Research 2023; 64(5), 842-849 Summary of the Invention [Problem to be solved by the invention]

[0005] If the dose error in the irradiation dose test exceeds the allowable value, the irradiation plan must be re-created. Creating an irradiation plan takes time and places a heavy burden on the person in charge of creating the irradiation plan, so a method for creating an irradiation plan that has a high probability of passing the irradiation dose test is needed.

[0006] Furthermore, the dose distribution map cannot be directly used to set the irradiation conditions of the irradiation device. In order to set the irradiation conditions of the irradiation device, it is necessary to create a fluence map based on the dose distribution map. The conventional method described in Non-Patent Document 1 could not reduce the burden of creating a fluence map.

[0007] An object of one aspect of the present invention is to provide a method for estimating a fluence map that has a high probability of passing an exposure dose test. [Means for solving the problem]

[0008] In order to solve the above problem, a method for estimating a fluence map according to one embodiment of the present invention is a method for estimating a fluence map in radiation therapy, executed by a computer, which causes the computer to execute an acquisition step of acquiring a first dose distribution map created for a first patient who is the subject of the radiation therapy, and an estimation step of inputting the first dose distribution map into a first trained model machine-learned using first training data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map.

[0009] In order to solve the above problem, a method for generating an estimation model according to one embodiment of the present invention is a method for generating an estimation model for estimating a fluence map in radiation therapy, executed by a computer, in which the computer performs machine learning of the estimation model using first learning data that associates a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, where the results of an irradiation dose test based on the second fluence map were better than a predetermined standard, with a specific dose distribution map used to create the specific fluence map.

[0010] In order to solve the above problem, a trained model according to one embodiment of the present invention is a trained model for estimating a fluence map in radiation therapy, and is obtained by machine learning using first learning data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and the trained model is configured to cause a computer to function to output an estimated result of a first fluence map corresponding to the first dose distribution map from the first dose distribution map created for the first patient who is the subject of the radiation therapy.

[0011] In order to solve the above problem, a fluence map estimation system according to one embodiment of the present invention includes an acquisition unit that acquires a first dose distribution map created for a first patient who is a target of radiation therapy, and an estimation unit that inputs the first dose distribution map into a first trained model that has been machine-learned using first learning data that associates, among second fluence maps created for a second patient who was previously the target of the radiation therapy, a specific fluence map whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard with a specific dose distribution map used to create the specific fluence map, and outputs an estimated result of the first fluence map corresponding to the first dose distribution map.

[0012] In order to solve the above problem, a control program according to one embodiment of the present invention is a control program for controlling a computer, and causes the computer to execute an acquisition step of acquiring a first dose distribution map created for a first patient who is a target of radiation therapy, and an estimation step of inputting the first dose distribution map into a first trained model that has been machine-learned using first learning data that corresponds a specific fluence map, among second fluence maps created for a second patient who was previously the target of radiation therapy, in which the results of an irradiation dose test based on the second fluence map were better than a predetermined standard, to a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map. [Effects of the Invention]

[0013] According to one aspect of the present invention, it is possible to realize a method for estimating a fluence map that has a high probability of passing an exposure dose test. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is a block diagram showing an example of a schematic configuration of an estimation system according to a first embodiment. [Figure 2] FIG. 1 is a schematic diagram illustrating an example of machine learning for generating a first trained model. [Figure 3] 4 is a flowchart showing an example of the flow of processing executed by the estimation device according to the first embodiment. [Figure 4] FIG. 10 is a block diagram showing an example of a schematic configuration of an estimation system according to a second embodiment. [Figure 5] FIG. 1 is a schematic diagram illustrating an example of machine learning for generating a second trained model. [Figure 6] 10 is a flowchart showing an example of the flow of processing executed by the estimation device according to the second embodiment. [Figure 7] FIG. 10 is a diagram illustrating an example of a schematic configuration of an estimation system according to a third embodiment. [Figure 8]FIG. 10 is a block diagram showing an example of a schematic configuration of an estimation device according to a third embodiment. [Figure 9] 10A and 10B are diagrams showing the results of generating dose distribution maps according to an example of the present invention and a comparative example. [Figure 10] FIG. 10 is a diagram showing the results of comparing dose volume histograms for dose distribution maps according to an example of the present invention and a comparative example. [Figure 11] FIG. 10 is a diagram showing the results of comparing the average doses to the rectum wall and bladder wall for dose distribution maps according to an example of the present invention and a comparative example. [Figure 12] FIG. 10 is a diagram showing the gamma pass rate obtained in patient dose verification performed using fluence maps according to an example of the present invention and a comparative example. [Figure 13] FIG. 10 is a diagram showing MCS (Modulation Complexity Score) in fluence maps of an example of the present invention and a comparative example. DETAILED DESCRIPTION OF THE INVENTION

[0015] [Embodiment 1] <Summary of the Invention> An embodiment of the present invention provides a method and system for estimating a fluence map that has a high probability of passing an irradiation dose test in radiation therapy.

[0016] Radiation therapy is a method of treating a patient by irradiating a lesion with radiation. Preferably, radiation therapy is a method capable of varying the intensity of radiation within an irradiation field, such as intensity-modulated radiation therapy (IMRT). IMRT is not particularly limited, but may be, for example, volumetric modulated arc therapy (VMAT), also known as rotational IMRT.

[0017] The type of disease targeted by radiation therapy is not particularly limited, but may be, for example, cancer. The cancer may be a solid cancer, such as prostate cancer, brain tumor, head and neck cancer, lung cancer, kidney cancer, liver cancer, digestive cancer, skin cancer, bladder cancer, breast cancer, uterine cancer, ovarian cancer, skin cancer, and sarcoma.

[0018] In order to carry out such radiation therapy, a treatment plan is created to distinguish between the diseased area and normal tissue and irradiate each with the intended dose. The creation of a treatment plan includes the creation of a dose distribution map and a fluence map. An example of a conventional method for creating a treatment plan is shown below.

[0019] First, a contour of each tissue is created based on a medical image of the lesion site. The lesion site is set appropriately depending on the type of target disease. For example, if the target disease is prostate cancer, the lesion site may be the entire prostate or a specific part of the prostate. Examples of medical images of the lesion site include CT (Computed Tomography) images, MRI (Magnetic Resonance Imaging) images, X-ray images, endoscopic images, and tissue staining images. It is preferable that the contour of each tissue created indicates at least the boundary between the tissue containing the lesion site and the remaining normal tissue in such a medical image.

[0020] Next, a dose distribution map is created based on the contours of each tissue so that the dose is concentrated at the lesion site and the dose irradiated to normal tissue is reduced as much as possible. The dose distribution map may be, for example, an image showing the contours of each tissue and the radiation dose at each position on the medical image, or an image in which such an image showing the radiation dose distribution is superimposed on the medical image, etc.

[0021] It is usually difficult to use the created dose distribution map to set the irradiation dose by a radiation irradiation device. A fluence map based on the dose distribution map is created so that the radiation irradiation device can irradiate a dose based on the dose distribution map. Here, the fluence of radiation refers to the time integral value of the radiant flux or radiation energy passing through a unit area. The fluence map may be an image showing the magnitude of the fluence at each position on a medical image, or may be an image in which such an image showing the fluence distribution is superimposed on a medical image or the like.

[0022] The fluence map obtained in this way is verified by an exposure dose test before actually irradiating a patient. The exposure dose test is also called patient dose verification. In the exposure dose test, for example, radiation based on the fluence map is irradiated to a test irradiation target called a phantom. Then, it is verified whether the error between the dose irradiated to the lesion site and each position around it and the planned dose is within an allowable range, in other words, whether the result of the exposure dose test is better than a predetermined standard.

[0023] The dose test may be performed using a gamma analysis method that measures the dose difference at the same position, the distance between points of equal dose (DTA; Distance-to-Agreement), or a combination of these. If the results of these dose tests are better than a predetermined standard, the test passes and the patient can be irradiated. "Better than the predetermined standard" may mean, for example, that the gamma pass rate in the gamma analysis is equal to or greater than a predetermined threshold. On the other hand, if the results do not meet the predetermined standard and fail, the treatment plan must be re-created.

[0024] Such treatment plans are typically created carefully by medical professionals such as physicians or medical physicists, taking, for example, four hours or more per case. Furthermore, irradiation dose tests are also performed by multiple medical professionals, taking, for example, approximately two hours per case. When an irradiation dose test fails, conventional methods only provide a numerical value, such as the gamma pass rate, which is the result of gamma analysis. Therefore, it is unclear what the problem with the treatment plan was, and it is necessary to create a new treatment plan without knowing what improvements need to be made.

[0025] Therefore, there is a strong demand in clinical practice for an efficient method for creating treatment plans with a high pass rate in irradiation dose tests. In one embodiment of the present invention, a fluence map is estimated using a trained model that is machine-learned using fluence maps that have actually produced good results in irradiation dose tests in the past and the corresponding dose distribution maps. As a result, an estimated fluence map result that is highly likely to produce good results in irradiation dose tests can be obtained.

[0026] This method can improve the efficiency of radiation therapy treatment planning and reduce the burden of re-planning treatments due to failed radiation dose tests. This can increase radiation therapy capacity, enabling more patients to receive radiation therapy. This effect will also contribute to the achievement of Goal 3 of the United Nations' Sustainable Development Goals (SDGs), "Ensure good health and promote well-being for all."

[0027] <General configuration of estimation system 100> The schematic configuration of the estimation system 100 will be described below with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the schematic configuration of the estimation system 100.

[0028] As shown in Fig. 1, the estimation system 100 may include an estimation device 1 and a display device 4. Fig. 1 shows the estimation system 100 including one estimation device 1 and one display device 4. However, the configuration of the estimation system 100 is not limited to this. For example, the estimation system 100 may not include a display device 4, or may include multiple display devices 4.

[0029] In the estimation system 100, the estimation device 1 and the display device 4 are connected to each other so that they can communicate with each other. The estimation device 1 and the display device 4 may be connected directly, by wire or wirelessly, or may be connected via a communication network. The type of the communication network is not limited, and may be a local area network (LAN) or the Internet.

[0030] The estimation device 1 is a computer that uses a first dose distribution map 21 created for a first patient who is a target of radiation therapy to output an estimation result of a first fluence map corresponding to the first dose distribution map 21. The estimation result by the estimation device 1 may be transmitted from the estimation device 1 to a display device 4.

[0031] The display device 4 may be a computer, smartphone, tablet terminal, or the like used by a user of the estimation system 100. Note that FIG. 1 shows the estimation system 100 in which the display device 4 is separate from the estimation device 1. However, the configuration of the estimation system 100 is not limited to this. For example, the display device 4 may be a device integrated with the estimation device 1, and in this case, the display device 4 may be a display unit (such as a display) provided in the estimation device 1.

[0032] <Configuration of Estimation Device 1> Next, a description will be given of the configuration of the estimation device 1. The estimation device 1 includes a control unit 10, a storage unit 20, and an input unit 30.

[0033] The storage unit 20 may be, for example, a hard disk drive (HDD) or a solid state drive (SSD). The storage unit 20 stores information transmitted from the control unit 10, and the stored information may be read by the control unit 10. Examples of such information include a first dose distribution map 21, first training data 22, and a first trained model 23, which will be described later.

[0034] The input unit 30 is configured to allow a user of the estimation device 1 to input information. The input unit 30 may be, for example, at least one of a keyboard, a mouse, or a touchpad. Alternatively, the input unit 30 may be an interface that accepts input of information from an external storage device such as a USB memory.

[0035] The control unit 10 is a control device that controls all the components of the estimation device 1. The control unit 10 may be, for example, a central processing unit (CPU) or a graphics processing unit (GPU). The control unit 10 includes an acquisition unit 11, a first model generation unit 12, and a first estimation unit (estimation unit) 13.

[0036] The control unit 10 may read a control program, which is software stored in the storage unit 20, and develop it in a memory such as a RAM (Random Access Memory), and execute the functions of each unit.

[0037] (Acquisition part 11) The acquisition unit 11 acquires a first dose distribution map 21 created for a first patient. The acquisition unit 11 may further acquire various other information used by the estimation apparatus 1. In this specification, the first patient is a patient who is a target of radiation therapy, for example, a patient for whom a fluence map needs to be created in order to perform radiation therapy.

[0038] In this specification, a second patient is a patient who has previously been the subject of radiation therapy and for whom a fluence map created for the radiation therapy exists. That is, the second patient is a patient for whom information on a treatment plan created for the radiation therapy is known. Such information on the known treatment plan may include, in addition to the fluence map, information on medical images and dose distribution maps used to create the fluence map.

[0039] In this specification, various information such as a dose distribution map relating to a first patient will be labeled "first," and various information relating to a second patient will be labeled "second." In addition, when describing general matters that do not specify the type of patient, the names of these various information may not be labeled "first" or "second."

[0040] The first dose distribution map 21 acquired by the acquisition unit 11 is a dose distribution map created based on a first medical image of a lesion site of a first patient. The first dose distribution map 21 may be created using, for example, a conventional simulation method.

[0041] The acquisition unit 11 may acquire various types of information from the storage unit 20. Furthermore, the acquisition unit 11 may acquire at least a part of the various types of information as input information from the input unit 30. The various types of input information acquired by the acquisition unit 11 from the input unit 30 may be stored in the storage unit 20.

[0042] (First model generation unit 12) The first model generation unit 12 generates a first trained model 23. The first trained model 23 is an estimation model that estimates a first fluence map corresponding to the first dose distribution map 21 in accordance with input of the first dose distribution map 21. Here, the first dose distribution map 21 and the first fluence map "correspond" to each other means that information on the dose distribution shown in the first dose distribution map 21 corresponds to information on the fluence shown in the first fluence map. In other words, the first fluence map corresponding to the first dose distribution map 21 includes information obtained by converting the dose distribution of the first dose distribution map 21 into fluence.

[0043] The first trained model 23 is a model trained by machine learning using the first training data 22 in which a specific fluence map is associated with a specific dose distribution map used to create the specific fluence map. In other words, the first trained model 23 is a trained model for causing a computer to function to estimate the first fluence map from the first dose distribution map 21.

[0044] A specific fluence map refers to a fluence map, among second fluence maps of a second patient, for which the results of an irradiation dose test were better than a predetermined standard. Furthermore, a specific dose distribution map refers to a dose distribution map, among second dose distribution maps of a second patient, that was used to create the specific fluence map. When information on treatment plans for multiple second patients exists, the specific fluence map and the specific dose distribution map can be said to be included in the information on treatment plans for at least some of the second patients, that is, specific patients.

[0045] The first trained model 23 is an estimation model for estimating a fluence map, which is machine-learned using the first training data 22 in which specific fluence maps that have yielded favorable results in an exposure dose test are associated with corresponding specific dose distribution maps. Therefore, the first fluence map estimated by the first trained model 23 is likely to result in favorable results in an exposure dose test, similar to the specific fluence map. In other words, the first model generation unit 12 can generate a first trained model 23 that significantly increases the likelihood of estimating a first fluence map that yields favorable results in an exposure dose test, compared to conventional methods.

[0046] The algorithm used for machine learning of the first trained model 23 is not particularly limited. For example, a Generative Adversarial Network (GAN), a Vision Transformer (ViT), a Convolutional Neural Network (CNN), etc. may be used, and a combination of these may also be used. An example of a flow of machine learning of the first trained model 23 will be described below with reference to FIG. 2. FIG. 2 is a schematic diagram showing an example of machine learning for generating the first trained model 23.

[0047] As shown in FIG. 2, a GAN including a generator and a discriminator may be used to generate the first trained model 23. The generator generates a new image by modifying an input image. The first model generation unit 12 inputs a specific dose distribution map to the generator. The generator generates a generated fluence map, which is an image modified based on the input specific dose distribution map, and sends the generated fluence map to the discriminator.

[0048] The classifier identifies the authenticity of the input image. A specific fluence map (reference fluence map) corresponding to the specific dose distribution map input to the generator is also input to the classifier as condition data. Here, the "true" of the authenticity identified by the classifier refers to the specific fluence map input as condition data, and the "false" refers to a generated fluence map generated by the generator by modifying the specific dose distribution map. Thus, the GAN used to generate the first trained model 23 is preferably a conditional generative adversarial network (conditional GAN).

[0049] Next, the classifier feeds back information about whether the classification result was correct or incorrect to itself and the generator. The generator optimizes the modification conditions for the dose distribution map so that the classification result of the classifier is incorrect. Meanwhile, the classifier optimizes the classification conditions so that the classification result is correct. The machine learning algorithm that generates the first trained model 23 is trained by repeated optimization in this way so that the generator can generate fluence map images corresponding to the dose distribution map with high accuracy.

[0050] In this way, the first trained model 23 may be a generator that has learned the relationship between the specific dose distribution map and the specific fluence map. The generator has a function of generating the first fluence map from the first dose distribution map 21. By using GAN for machine learning, the generator and the classifier can autonomously proceed with machine learning even when the amount of first training data 22 is small. Therefore, the accuracy of the first fluence map generated by the first trained model 23 can be efficiently improved.

[0051] When a CNN or the like is used to generate the first trained model 23, the first model generation unit 12 may construct the first training data 22 using the specific dose distribution map as an explanatory variable and the specific fluence map as a target variable. Then, the first model generation unit 12 may perform machine learning using the first training data 22 to generate the first trained model 23.

[0052] The first model generation unit 12 may store the generated first trained model 23 in the storage unit 20.

[0053] The control unit 10 may not include the first model generation unit 12. The estimation device 1 may be a device that performs processing to output an estimation result of a first fluence map corresponding to the first dose distribution map 21, using a first trained model 23 that is generated in advance. In this case, the estimation device 1 may store the first trained model 23 in the storage unit 20 in advance.

[0054] (1st estimation part 13) The first estimation unit 13 inputs the first dose distribution map 21 to the first trained model 23 and outputs an estimation result of a first fluence map corresponding to the first dose distribution map 21.

[0055] The first estimating unit 13 may output the estimation result of the first fluence map to the display device 4 and display it on the display device 4. Furthermore, the first estimating unit 13 may output the estimation result of the first fluence map to the storage unit 20 to store it, or may output it to a printing device, a recording medium other than the storage unit 20, a communication device, or the like.

[0056] The estimation device 1 having such a first estimation unit 13 can input a first dose distribution map 21 and output an estimation result of a first fluence map corresponding to the first dose distribution map 21. Furthermore, the first trained model 23 used to generate the first fluence map is machine-learned using first training data 22 in which a specific dose distribution map and a specific fluence map are associated with each other. The specific dose distribution map and the specific fluence map are part of the information of a treatment plan that has a track record of passing an irradiation dose test. By using the first trained model 23 machine-learned using such first training data 22, the first estimation unit 13 can output an estimation result of a first fluence map that is highly likely to pass an irradiation dose test.

[0057] <Method of estimating fluence maps> A method for estimating a fluence map in radiation therapy according to one embodiment of the present invention will be described with reference to Fig. 3, taking as an example a case where the estimation method is executed by an estimation device 1. Fig. 3 is a flowchart showing an example of the flow of processing executed by the estimation device 1. Fig. 3 also shows the flow of processing executed by an estimation system 100 including the estimation device 1. Note that the content already explained in the section on the estimation system 100 above will not be explained here.

[0058] First, a method for generating an estimation model (first trained model 23) according to an embodiment of the present invention will be described. As shown in Fig. 3, the acquisition unit 11 acquires a specific dose distribution map and a specific fluence map (S1).

[0059] Next, the first model generation unit 12 generates a first trained model 23 by machine learning using first training data 22 in which a specific dose distribution map, which is an input image, is associated with a specific fluence map, which is a condition (S2, learning step). Here, a case is illustrated in which the machine learning algorithm used by the first model generation unit 12 is a conditional GAN. However, the machine learning algorithm used by the first model generation unit 12 is not particularly limited.

[0060] Next, a method for estimating a fluence map according to an embodiment of the present invention will be described. The estimation device 1 may perform only the subsequent processes without performing the above-described processes S1 and S2.

[0061] The acquisition unit 11 acquires the first dose distribution map 21 (S3, acquisition step).

[0062] Next, the first estimation unit 13 inputs the first dose distribution map 21 to the first trained model 23 (S4). Then, the first estimation unit 13 outputs an estimation result of a first fluence map corresponding to the first dose distribution map 21 (S5, estimation step).

[0063] [Embodiment 2] Other embodiments of the present invention will be described below. For ease of explanation, the same reference numerals will be used to designate components having the same functions as those described in the previous embodiment, and the description thereof will not be repeated.

[0064] <Estimation system 100a> 4, unlike the estimation system 100 according to the first embodiment, the estimation system 100a according to this embodiment may include an estimation device 1a having a control unit 10a. The control unit 10a further includes a second model generation unit 14 and a second estimation unit 15 in addition to the control blocks of the control unit 10 according to the first embodiment. Accordingly, the storage unit 20 included in the estimation device 1a may store a specific medical image 24, a second trained model 25, and a first medical image 26, which will be described later.

[0065] The estimation device 1a has the function of outputting an estimation result of a first dose distribution map 21 from a first medical image 26, in addition to the functions of the estimation device 1 according to the first embodiment. That is, the estimation device 1a can output an estimation result not only of the first fluence map, but also of the first dose distribution map 21 used to generate the first fluence map. According to the estimation device 1a, it is also possible to successively estimate the first dose distribution map 21 and the first fluence map from the first medical image 26.

[0066] (Acquisition part 11) In this embodiment, the acquisition unit 11 may acquire a first medical image 26. The first medical image 26 is a medical image of a first patient. The first medical image 26 is used by the second estimator 15 to generate an estimation result of the first dose distribution map 21.

[0067] (Second model generation unit 14) The second model generation unit 14 generates a second trained model 25. The second trained model 25 is a model that estimates a first dose distribution map 21 corresponding to a first medical image 26 in response to an input of the first medical image 26. The first medical image 26 and the first dose distribution map 21 "correspond" to each other means that the dose distribution shown in the first dose distribution map 21 is created based on the position of each tissue shown in the first medical image 26.

[0068] The second trained model 25 may be a model generated by machine learning using a machine learning algorithm that combines a generative adversarial network including a generator that has learned the relationship between a specific medical image 24 and a specific dose distribution map and a Vision Transformer. Such a second trained model 25 is a model that causes a computer to function so as to output the first dose distribution map 21 as an estimation result of the second trained model 25 from an input of a first medical image 26.

[0069] The specific medical image 24 is a medical image of a second patient, and is a medical image used to create a specific dose distribution map for which the results of the irradiation dose test were favorable. The specific medical image 24 and the specific dose distribution map can be considered as second training data for machine learning the second trained model 25.

[0070] Fig. 5 is a schematic diagram showing an example of machine learning for generating the second trained model 25. With reference to Fig. 5, a machine learning algorithm that combines GAN and ViT, which is used for generating the second trained model 25, will be described.

[0071] 5, a GAN including a generator and a classifier may be used to generate the second trained model 25, as in the generation of the first trained model 23. In addition, it is preferable to use a machine learning algorithm that combines GAN and ViT.

[0072] Vision Transformer (ViT) is an algorithm that divides an input image into multiple patches, performs feature vectorization for each patch while taking into account the positional relationship between the patches, and processes the resulting image using a Transformer Encoder with an attention layer. The ViT used by the second model generation unit 14 in combination with GAN is not particularly limited as long as it is an algorithm that uses an image as input and processes the image using a Transformer Encoder. The ViT may be, for example, an algorithm disclosed in a reference (Alexey Dosovitskiy et al., An Image Is Worth 16x16 Words: Transformers For Image Recognition At Scale, conference paper at ICLR 2021, 2021), or may be a modified version of the algorithm.

[0073] An example of a ViT with a modified algorithm is the Swin Transformer described in the reference (Ze Liu et al., Proceedings of the IEEE / CVF International Conference on Computer Vision (ICCV), 2021, pp. 10012-10022). The Swin Transformer divides patches into smaller windows and calculates the positional relationship between patches within the same window. It also performs calculations for windows divided at different positions (shifted windows) to reduce the loss of feature information for the positional relationship. Furthermore, by having a hierarchical structure that allows for variable patch size, it is possible to consider feature quantities with different area sizes, from the entire image to the fine details. Compared to the unmodified ViT, this Swin Transformer can reduce computational costs while providing estimation with the same or higher accuracy.

[0074] In machine learning to generate the second trained model 25, medical images input to the generator are modified by ViT (to generate new images). As described above, ViT can perform feature vectorization in a way that includes a high degree of information about the positional relationships of feature points in image data, without performing convolution processing such as in CNN (convolutional neural network). Therefore, new images can be generated without excessively increasing the processing load while reducing the loss of features in the medical image, thereby achieving highly efficient machine learning processing.

[0075] The generator and classifier processes of the GAN have already been explained, so details will be omitted. The second model generation unit 14 inputs a specific medical image 24 to the generator. The generator generates a generated dose distribution map, which is an image modified from the input specific medical image 24, and sends the generated dose distribution map to the classifier. Here, the modification of the image by the generator may be performed by ViT.

[0076] The second model generation unit 14 also inputs a specific dose distribution map (reference dose distribution map) as condition data to the classifier. The classifier classifies the authenticity of the input image and feeds back information on whether the classification result was correct or incorrect to itself and the generator.

[0077] The second trained model 25 thus trained may be a generator that has learned the relationship between the specific medical image 24 and the specific dose distribution map. The generator has a function of generating the first dose distribution map 21 from the first medical image 26.

[0078] Note that only one of the GAN and ViT algorithms, or CNN, may be used to generate the second trained model 25. The second model generation unit 14 may store the generated second trained model 25 in the storage unit 20.

[0079] The control unit 10a may not include the second model generation unit 14. The estimation device 1a may perform processing to output an estimation result of the first dose distribution map 21 corresponding to the first medical image 26 using a second trained model 25 generated in advance. In this case, the estimation device 1a may store the second trained model 25 in the storage unit 20 in advance.

[0080] (Second estimation part 15) The second estimating unit 15 inputs the first medical image 26 into the second trained model 25 and outputs an estimation result of the first dose distribution map 21 corresponding to the first medical image 26. The second estimating unit 15 may output the estimation result of the first dose distribution map 21 to the first estimating unit 13. The first estimating unit 13 can output the estimation result of the first fluence map using the first dose distribution map 21 acquired from the second estimating unit 15.

[0081] The estimation device 1a having such a second estimation unit 15 can estimate both the first dose distribution map 21 and the first fluence map using a trained model that has been machine-learned using information on treatment plans that have produced good results in irradiation dose tests. Therefore, the estimation device 1a can output an estimated result of the first fluence map that is highly likely to produce good results in the irradiation dose test.

[0082] The second estimation unit 15 may output the estimation result of the first dose distribution map 21 to the display device 4 and display it on the display device 4. Furthermore, the second estimation unit 15 may output the estimation result of the first dose distribution map 21 to the storage unit 20 to be stored therein, or may output the estimation result to a printing device, a recording medium other than the storage unit 20, a communication device, or the like.

[0083] The second model generating unit 14 and the second estimating unit 15 may be realized by a computer separate from the estimation device 1 a. In this case, the estimation device 1 a may acquire the first dose distribution map 21 output by the separate computer and use it to generate the first fluence map.

[0084] <Method for estimating dose distribution maps> A method for estimating a dose distribution map in radiation therapy according to one embodiment of the present invention will be described with reference to Fig. 6, taking as an example a case where the estimation method is executed by an estimation device 1a. Fig. 6 is a flowchart showing an example of the flow of processing executed by the estimation device 1a. Fig. 6 also shows the flow of processing executed by an estimation system 100a including the estimation device 1. Note that the contents already explained in the above sections up to the estimation system 100a will not be explained again here.

[0085] First, a method for generating the second trained model 25 according to an embodiment of the present invention will be described. As shown in Fig. 6, the acquisition unit 11 acquires a specific medical image 24 and a specific dose distribution map (S11).

[0086] Next, the second model generation unit 14 generates a second trained model 25 by machine learning using second training data in which a specific medical image 24, which is an input image, is associated with a specific dose distribution map, which is a condition (S12). Here, the machine learning algorithm used by the second model generation unit 14 is a combination of conditional GAN ​​and ViT, as an example. However, the machine learning algorithm used by the second model generation unit 14 is not particularly limited.

[0087] Next, a method for estimating a dose distribution map according to one embodiment of the present invention will be described. The estimation device 1a may perform the following processes without performing the processes of S11 and S12 described above.

[0088] The acquisition unit 11 acquires a first medical image 26 (S13). Next, the second estimation unit 15 inputs the first medical image 26 to the second trained model 25 (S14). Then, the second estimation unit 15 outputs an estimation result of the first dose distribution map 21 corresponding to the first medical image 26 (S15).

[0089] Thereafter, the estimation device 1a may cause the first estimator 13 to acquire the first dose distribution map 21 output by the second estimator 15, thereby successively executing the processes from S3 to S5 shown in FIG.

[0090] [Embodiment 3] Further embodiments of the present invention are described below.

[0091] 1, the estimation system 100 includes an estimation device 1 that includes an input unit 30 that accepts inputs such as a first dose distribution map 21 from a user and outputs estimation results to a display device 4. However, the present invention is not limited to this. For example, as shown in FIG. 7, an estimation system 100b according to this embodiment may include an estimation device 1b that is communicably connected to communication terminals 5a and 5b used by users via a communication network 9.

[0092] 7, the estimation device 1b receives information such as a first dose distribution map 21 from each of the communication terminals 5a and 5b. Then, the estimation device 1b transmits an estimation result corresponding to the information received from the communication terminal 5a to the communication terminal 5a, and transmits an estimation result corresponding to the information received from the communication terminal 5b to the communication terminal 5b. The estimation result is a first fluence map estimated by the first trained model 23 by the first estimator 13.

[0093] 7 illustrates an estimation system 100b including the communication terminals 5a and 5b and the estimation device 1b, but is not limited to this. In the estimation system 100b, the estimation device 1b may be capable of communicating with, for example, three or more communication terminals.

[0094] (Configuration of estimation device 1b) The configuration of the estimation device 1b will be described with reference to Fig. 8. Fig. 8 is a functional block diagram showing an example configuration of an estimation system 100b according to one embodiment of the present invention. For ease of explanation, components having the same functions as those described in the previous embodiment are denoted by the same reference numerals, and their description will not be repeated.

[0095] 8, the estimation device 1b includes a communication unit 60 that functions as a communication interface with the communication terminals 5a and 5b. The acquisition unit 11 receives information such as the first dose distribution map 21 via the communication unit 60.

[0096] The first estimation unit 13 transmits the estimation results to each of the communication terminals 5a and 5b via the communication unit 60. The estimation device 1b may generate a web page that allows the estimation results related to the received information to be viewed or downloaded, and provide information for accessing the web page to the user who sent the information. The estimation device 1b included in the estimation system 100b may include the control unit 10a included in the estimation device 1a, instead of the control unit 10.

[0097] [Software implementation example] The functions of the estimation device 1 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each part included in the control unit 10).

[0098] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0099] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0100] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0101] 〔summary〕 A method for estimating a fluence map according to aspect 1 of the present invention is a method for estimating a fluence map in radiation therapy, executed by a computer, and causes the computer to execute an acquisition step of acquiring a first dose distribution map created for a first patient who is the subject of the radiation therapy, and an estimation step of inputting the first dose distribution map into a first trained model machine-learned using first training data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map.

[0102] A method for estimating a fluence map according to aspect 2 of the present invention may be such that, in aspect 1, the first learned model includes a generator that has learned the relationship between the specific dose distribution map and the specific fluence map, and the generator has a function of generating the first fluence map from the first dose distribution map.

[0103] A method for estimating a fluence map according to aspect 3 of the present invention may, in aspect 1 or 2, further cause the computer to generate the first dose distribution map from a first medical image of the first patient using a second trained model trained by a machine learning algorithm that combines a generative adversarial network including a generator that has learned the relationship between the specific medical image used to create the specific dose distribution map and the specific dose distribution map, and a Vision Transformer.

[0104] A method for generating an estimation model according to aspect 4 of the present invention is a method for generating an estimation model for estimating a fluence map in radiation therapy, executed by a computer, in which the computer is caused to perform machine learning of the estimation model using first learning data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map.

[0105] The trained model of aspect 5 of the present invention is a trained model for estimating a fluence map in radiation therapy, and is obtained by machine learning using first learning data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and is used to cause a computer to function to output an estimated result of a first fluence map corresponding to the first dose distribution map from the first dose distribution map created for the first patient who is the subject of the radiation therapy.

[0106] A fluence map estimation system according to aspect 6 of the present invention includes an acquisition unit that acquires a first dose distribution map created for a first patient who is a subject of radiation therapy, and an estimation unit that inputs the first dose distribution map into a first trained model that has been machine-learned using first learning data that associates, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, a specific fluence map whose irradiation dose test results based on the second fluence map were better than a predetermined standard with a specific dose distribution map used to create the specific fluence map, and outputs an estimated result of the first fluence map corresponding to the first dose distribution map.

[0107] A control program according to aspect 7 of the present invention is a control program for controlling a computer, and causes the computer to execute an acquisition step of acquiring a first dose distribution map created for a first patient who is the subject of radiation therapy, and an estimation step of inputting the first dose distribution map into a first trained model machine-learned using first learning data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map.

[0108] A recording medium according to an eighth aspect of the present invention is a computer-readable recording medium on which the control program of the seventh aspect is recorded.

[0109] [Additional Notes] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Example]

[0110] An embodiment of the present invention will be described below, but the present invention is not limited to the scope of each embodiment shown below.

[0111] (1. Generation of dose distribution map) A second trained model was generated to generate a first dose distribution map for the first patients who were to receive radiation therapy. 270 cases of patients who had been diagnosed with prostate cancer and had received radiation therapy for the prostate were selected as the second patients. These second patients had medical images of the prostates that were to be treated with radiation therapy, and treatment plan data including a dose distribution map that showed the contours and dose distribution of the prostate created based on the medical images.

[0112] The treatment plan data in question all had gamma pass rates (GPRs) above a predetermined threshold in the patient dose verification, which is an irradiation dose test, and were judged to be acceptable in clinical practice. That is, the treatment plan data in question include information on the specific dose distribution map and the specific fluence map.

[0113] CT images were used as medical images of the prostate. The specific dose distribution map was image data created based on the CT images using the dose calculation algorithm AcurosXB in the treatment planning system Eclipse (Varian Medical Systems) to perform radiation therapy using the treatment system TrueBeam (Varian Medical Systems). The radiation irradiation energy was 10MV-X ray, and the dose prescription was 2Gy × 39 fractions = 78Gy (D mean prescription).

[0114] To obtain the second trained model, we performed machine learning using a machine learning algorithm that combines GAN and ViT. CT images were input as input images into the GAN generator. Furthermore, specific dose distribution maps were used as conditions, and information on the correspondence with the corresponding CT images was also input into the GAN classifier. In other words, the GAN used here is a conditional GAN ​​that can input conditions corresponding to the input image. Furthermore, the ViT used in combination with the GAN uses a Swin Transformer.

[0115] A CT image was input into the second trained model, and a first dose distribution map corresponding to the CT image was obtained.

[0116] (2. Generating a fluence map) Next, a first trained model was generated from the obtained first dose distribution map to generate an estimated result of the first fluence map.

[0117] To obtain the first trained model, machine learning was performed using a conditional GAN ​​algorithm. The specific dose distribution map described above was used as an input image and input to the GAN generator. In addition, the specific fluence map corresponding to the specific dose distribution map was used as a condition and input to the GAN classifier, including information on the correspondence with the corresponding specific dose distribution map.

[0118] The specific fluence map, which is a condition for the training data, is image data created for each patient 2 from the specific dose distribution map by the treatment planning system (Eclipse). The specific fluence map also has a gamma pass rate (GPR) equal to or greater than a predetermined threshold in the patient dose verification, as described above, and is included in the treatment plan data that has been determined to be acceptable in clinical practice.

[0119] (3. Evaluation of dose distribution maps) Figure 9 shows a dose distribution map of a comparative example created using Eclipse and a first dose distribution map of an example of the present invention created using the second trained model. As shown in Figure 9, the second trained model was created using Eclipse and showed a dose distribution with the same accuracy as the dose distribution map actually used in radiation therapy.

[0120] Next, Fig. 10 shows the results of comparing the dose volume histograms (DVH) of the dose distribution maps of the present invention and the comparative example. Also, Fig. 11 shows the average dose (DVH) to the rectal wall and bladder wall, which are organs at risk (OAR). mean ) are compared in the dose distribution maps of the invention example and the comparative example.

[0121] 10 and 11, the dose to the planning target volume (PTV) indicating the treatment site and the rectal wall and bladder wall indicating organs at risk (OAR), which are normal tissues, was not significantly different in the example of the present invention from the comparative example. In other words, the first dose distribution map of the example of the present invention output by the second trained model was equivalent to the dose distribution map of the comparative example, and it was shown that it can be used for radiation therapy.

[0122] (4. Evaluation of Fluence Map) Patient dose verification was performed using the dose distribution map of the comparative example, the fluence map of the comparative example created by the conventional method, and the first fluence map of the present invention obtained by inputting the first dose distribution map of the present invention into the first trained model.

[0123] The tolerance for patient dose verification was set to three conditions: 3% / 2 mm, 2% / 2 mm, and 2% / 1 mm, based on the process-based tolerance in AAPM TG-218. Gamma analysis to obtain the gamma pass rate was performed under the following conditions. ·Measurement: EPID ·Analysis: In-house (Python) Relative dose mode Global normalization Cax normalization

[0124] Figure 12 shows the gamma pass rate obtained in patient dose verification under each condition. As shown in Figure 12, for all tolerances, a significant improvement in the gamma pass rate was observed when the first fluence map of the example of the present invention was used compared to when the fluence map of the comparative example was used. In other words, it was demonstrated that the first fluence map of the example of the present invention output by the first trained model achieves irradiation accuracy that exceeds that of the fluence map of the comparative example, and is suitable for use in radiation therapy.

[0125] One of the reasons why the irradiation accuracy of the first fluence map in the present invention was good is thought to be the reduced complexity of the fluence map. The reference (McNiven, et al. Med Phys 2010; 37(2)) explains the MCS (Modulation Complexity Score). The MCS is an index of the overall complexity of a radiation beam and can be calculated from the MLC (Multi-Leaf Collimator) operation and the irradiation field area. The MCS is a value between 0.0 and 1.0, and the smaller the value, the higher the complexity of the radiation beam.

[0126] FIG. 13 shows the MCS of the fluence maps of the present invention example and the comparative example. As shown in FIG. 13, the first fluence map of the present invention example exhibited a significantly larger MCS value than the fluence map of the comparative example. In other words, it was shown that the first fluence map of the present invention example is less complex than conventional ones and can achieve highly accurate irradiation. This is thought to indicate that the first trained model has learned the features of multiple specific fluence maps and is optimized by focusing on important features, thereby becoming a model capable of estimating a first fluence map with low complexity. [Explanation of symbols]

[0127] 11 Acquisition Department 13 1st estimation part (estimation part) 21 First Dose Distribution Map 22 First training data 23 First trained model 24 Specific medical images 25 Second trained model 26 First Medical Imaging 100, 100a, 100b Estimation Systems

Claims

1. 1. A computer-implemented method for estimating a fluence map in radiation therapy, comprising: The computer, acquiring a first dose distribution map created for a first patient who is a target of the radiation therapy; A fluence map estimation method that executes an estimation step of inputting the first dose distribution map into a first trained model that has been machine-learned using first learning data that associates a specific fluence map, among second fluence maps created for a second patient who was previously the subject of radiation therapy, where the result of an irradiation dose test based on the second fluence map was better than a predetermined standard, with a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map.

2. The first trained model is a generator that learns a relationship between the specific dose distribution map and the specific fluence map; The method of claim 1 , wherein the generator is operable to generate the first fluence map from the first dose distribution map.

3. The computer further comprises:

2. The method for estimating a fluence map according to claim 1, wherein the first dose distribution map is generated from a first medical image of the first patient using a second trained model trained by a machine learning algorithm that combines a generative adversarial network including a generator that has learned the relationship between the specific dose distribution map and the specific medical image used to create the specific dose distribution map, and a Vision Transformer.

4. 1. A computer-implemented method for generating an estimation model for estimating a fluence map in radiation therapy, comprising: The computer, A method for generating an estimation model, which performs machine learning of the estimation model using first learning data that corresponds a specific fluence map, among second fluence maps created for a second patient who was previously the subject of radiation therapy, in which the results of an irradiation dose test based on the second fluence map were better than a predetermined standard, to a specific dose distribution map used to create the specific fluence map.

5. A trained model for estimating a fluence map in radiation therapy, obtained by machine learning using first learning data in which a specific fluence map, which has a result of an irradiation dose test based on the second fluence map that is better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy; A trained model for causing a computer to function to output an estimated result of a first fluence map corresponding to a first dose distribution map created for a first patient who is a target of the radiation therapy, from the first dose distribution map.

6. an acquisition unit that acquires a first dose distribution map created for a first patient who is a target of radiation therapy; A system for estimating a fluence map in radiation therapy, comprising: an estimation unit that inputs a first dose distribution map into a first trained model that has been machine-learned using first learning data that associates a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, where the result of an irradiation dose test based on the second fluence map was better than a predetermined standard, with a specific dose distribution map used to create the specific fluence map; and outputs an estimated result of the first fluence map corresponding to the first dose distribution map.

7. A control program for controlling a computer, The computer, an acquiring step of acquiring a first dose distribution map created for a first patient who is a target of radiation therapy; a control program for executing an estimation step of inputting the first dose distribution map into a first trained model that has been machine-learned using first learning data in which a specific fluence map, among second fluence maps created for a second patient who was previously the subject of the radiation therapy, whose results of an irradiation dose test based on the second fluence map were better than a predetermined standard, is associated with a specific dose distribution map used to create the specific fluence map, and outputting an estimated result of the first fluence map corresponding to the first dose distribution map.

8. A computer-readable recording medium on which the control program according to claim 7 is recorded.

Images