Dose distribution generation method, program, deep learning device, trained model, and dose distribution generation system

Abstract

The present invention makes it possible to provide dose distribution information adapted to needs of clinical sites by correcting, on the basis of a facility treatment policy of a medical facility, standard relationship information that is a learning result obtained by means of deep learning from target dose distribution information and affected area description information, and creating facility relationship information suitable for each facility. With this method, it is possible to generate, in accordance with a medical facility and a patient, dose distribution candidate information of radiation emitted to an affected area. As a result, it is possible to form a treatment plan optimum for a patient at each medical facility.

Description

Dose distribution generation method, program, deep learning device, trained model, and dose distribution generation system

[0001] The present invention relates to a dose distribution generation method, a program, a deep learning device, a trained model, and a dose distribution generation system used in radiation therapy.

[0002] Radiation therapy technology is improving year by year, and in recent years, intensity-modulated radiation therapy (IMRT), which exhibits greater therapeutic effects than conventional methods, has become increasingly popular. IMRT modulates the intensity of radiation within the radiation field by moving a radiation-shielding multi-leaf collimator during radiation irradiation, concentrating radiation on the tumor and reducing exposure to surrounding normal tissue. This makes it possible to irradiate the tumor with stronger radiation without increasing side effects. IMRT is a technology that has a higher survival rate and reduces the occurrence of side effects compared to conventional radiation therapy, and is expected to significantly contribute to improving the quality of life of cancer patients.

[0003] Radiation treatment plans must be determined by calculating the dose distribution based on information about the patient's body obtained from CT images, etc. Normal tissue exists around the tumor to be irradiated, and among these normal tissues are organs at risk (OARs) that must be avoided from radiation exposure as much as possible. Medical professionals determine and select the optimal treatment plan that delivers the maximum dose to the target tumor without damaging the OAR. Because an increase in the dose to the target also leads to an increase in the dose to the OAR, there is a trade-off between the irradiation dose to the target and the reduction in the dose to the OAR. Medical professionals adjust radiation irradiation parameters and repeatedly operate the treatment planning system to determine a treatment plan that is deemed clinically appropriate.

[0004] The time required to develop IMRT treatment plans has been a problem due to the need to operate a treatment planning system and repeatedly optimize the plan. Furthermore, because radiation therapy plans are determined by medical professionals based on their individual knowledge and experience, it has been pointed out that the quality of treatment plans varies. Several methods have been proposed to address these issues (Patent Documents 1 to 3). Patent Document 1 proposes a method that shortens treatment time by searching a database of previously planned radiation treatments and developing a treatment plan based on a plan that closely matches patient parameters. Patent Document 2 describes creating a treatment plan tailored to the planner's (medical professional's) preferences by performing an optimization process based on a first treatment plan and determining a second treatment plan. Patent Document 3, developed by the present inventors, discloses a technology that reduces differences between medical professionals by determining dose distribution based on target diseased area information and pre-trained results obtained through predetermined deep learning.

[0005] JP 2016-523146 A JP 2023-519563 A JP 2020-178935 A

[0006] Radiation treatment plans must take into account not only differences in the treatment target but also various other factors. For example, differences in the radiation therapy equipment used at each medical facility, as well as treatment policies such as combinations with surgery and chemotherapy, often vary from facility to facility. As a result, the proposed treatment plan is not necessarily optimal, and adjustments must be made at the medical facility. In other words, there was a need for technology that could reduce the time required to create treatment plans and provide the best plan for each medical facility and each patient.

[0007] The method described in Patent Document 1 creates a treatment plan based on an existing radiation therapy plan that is highly consistent with patient parameters, which can shorten the treatment time, but it is not possible to create an optimal treatment plan for each medical facility. The method described in Patent Document 2 can create a plan that conforms to the treatment policy of medical professionals, but since it is determined based on a dose distribution function, it takes a certain amount of time to create the treatment plan. The method described in Patent Document 3 uses AI for optimization, so it can create a high-quality treatment plan in a short time, but it does not provide an optimal method for each medical facility or patient.

[0008] In view of this situation, the present invention aims to provide a method and system that can quickly create a dose distribution customized for each medical facility and each patient while optimizing the irradiation dose to the target and the dose reduction to the OAR.

[0009] The present invention relates to the following method and system. This dose distribution generation method uses relationship information created by deep learning to generate a radiation therapy dose distribution, and includes: a facility treatment policy acquisition step for acquiring a facility treatment policy for a medical facility; and a facility relationship information creation step for modifying standard relationship information, which is a learning result obtained by deep learning from explanatory affected area information and target dose distribution information, based on the facility treatment policy to create facility relationship information. The standard relationship information, which is a basic trained model obtained by deep learning, may not be suitable for the actual situation at some medical facilities. Therefore, by making modifications based on the input facility treatment policy, it is possible to generate candidate dose distribution information tailored to the facility. Similarly, it is also possible to obtain candidate dose distribution information suited to a patient in accordance with the patient treatment policy.

[0010] A dose distribution generation method having an output step of outputting dose distribution candidate information and reference information from an output unit. By displaying dose distribution candidate information when using related information used at a facility other than the patient's own facility, such as related information at other facilities, together with the patient's dose distribution candidate information, medical personnel can visually consider a dose distribution appropriate for the patient.

[0011] 1 is a diagram showing an example of the hardware configuration of a dose distribution generation system 100 according to an embodiment; FIG. 2 is a flowchart showing an example of a processing flow in which a deep learning device acquires facility-related information according to an embodiment; FIG. 3 is a diagram showing an example of an input screen for a facility treatment policy; FIG. 4 is a flowchart showing another example of a processing flow in which a deep learning device acquires facility-related information according to an embodiment; FIG. 5 is a diagram showing a schematic diagram of a method for correcting standard relationship information serving as a basic model in accordance with a facility policy; FIG. 6 is a diagram showing a schematic diagram of a method for correcting standard relationship information serving as a basic model; FIG. 7 is a flowchart showing an example of a processing flow in which a dose distribution generation device generates a dose distribution for each patient according to an embodiment; FIG. 8 is a diagram showing a schematic diagram of a dose distribution candidate output together with reference information.

[0012] FIG. 1 illustrates an example of the hardware configuration of a dose distribution generation system 100 according to an embodiment. The dose distribution generation system 100 generates a dose distribution of radiation irradiated to a target affected area during radiation therapy based on target affected area information, learning results previously acquired through a predetermined deep learning process, and an institutional treatment policy and a patient treatment policy. Here, the target affected area refers to the affected area to be treated in radiation therapy, i.e., the target volume to be irradiated with radiation. The target affected area information is information indicating the position and cross-sectional contours of the target affected area, specifically, information indicating the positions and contours of multiple cross-sections of the same target affected area obtained from CT images or the like. The multiple cross-sections of the same target affected area may be, for example, cross-sections of the target affected area at different positions on a predetermined axis. The predetermined deep learning is deep learning using training data in which explanatory affected area information is used as an explanatory variable and target dose distribution information is used as a target variable. The explanatory affected area information is information used in the learning process of the predetermined deep learning, indicating the position and cross-sectional contours of the affected area. The target dose distribution information is information used in the learning process of a specified deep learning, is information that is previously associated with explanatory affected area information, and is information that indicates the dose distribution in the cross section indicated by the corresponding explanatory affected area information. Deep learning that learns the relationship between the position of the affected area, the cross-sectional contour, and the dose distribution is called dose distribution learning. The reference relationship information among the information that indicates the relationship between the position of the affected area, the cross-sectional contour, and the dose distribution, and that indicates the learning results of dose distribution learning, is called standard relationship information (learned model).

[0013] The dose distribution generation system 100 includes a deep learning device 1 and a dose distribution generation device 2. The deep learning device 1 includes a control unit 11 including a processor 91 such as a CPU and a memory 92 connected by a bus, and executes a program. By executing the program, the deep learning device 1 functions as a device including the control unit 11, an input unit 12, a storage unit 13, and a communication unit 14. Note that all or part of the functions of the deep learning device 1 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built into a computer system. The program may also be transmitted via a telecommunications line.

[0014] The control unit 11 (deep learning unit) controls the operation of each functional unit of the deep learning device 1. The control unit 11 acquires standard relationship information as a reference through dose distribution learning based on the explanation affected area information and target dose distribution information, which are teacher data. The control unit 11 may create one standard relationship information or multiple standard relationship information based on the explanation affected area information and target dose distribution information of one facility or multiple facilities. Furthermore, the control unit 11 selects facility relationship information suitable for the facility performing treatment from the multiple standard relationship information based on the acquired facility treatment policy (basic information for creating a treatment plan, such as information on treatment devices, protocols, and OAR protection), which will be described later, or modifies the standard relationship information to suit the facility.

[0015] The input unit 12 includes input devices such as a mouse, keyboard, and touch panel. Alternatively, input may be made by voice input, or input may be made using a chatbot. The input unit 12 may be configured by connecting these input devices to its own device. The input unit 12 accepts input of various information for its own device. The various information includes explanation affected area information, target dose distribution information, and facility treatment policy.

[0016] The storage unit 13 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 13 stores various information related to the deep learning device 1. For example, the storage unit 13 stores explanatory affected area information and target dose distribution information input to the input unit 12, standard relationship information which is the learning result of dose distribution learning, and further, an input facility treatment policy and facility-related information corrected by the facility treatment policy.

[0017] The communication unit 14 includes a communication interface for communication between the communication unit 14 and the dose distribution generation device 2, and communicates with the dose distribution generation device 2 via a wired or wireless connection. The communication unit 14 also functions as an acquisition unit that receives various information including dose distribution information and target dose distribution information. The communication unit 14 transmits related information to the dose distribution generation device 2 via the network 9. The communication unit 14 may communicate with an external storage device such as a USB (Universal Serial Bus) and acquire various information including target affected area information from the external storage device.

[0018] The dose distribution generation device 2 generates dose distribution candidates (hereinafter referred to as "dose distribution candidates") for each cross section of the target affected area indicated by the target affected area information based on the target affected area information, facility-related information, and a patient treatment policy set for each patient. The dose distribution generation device 2 includes a control unit 21 including a processor 93 such as a CPU and a memory 94 connected via a bus, and executes a program. By executing the program, the dose distribution generation device 2 functions as a device including a communication unit 20, a control unit 21, an input unit 22, a storage unit 23, and an output unit 24. Note that, similar to the deep learning device 1, all or part of the functions of the dose distribution generation device 2 may be realized using known hardware, and the program may be recorded on a known computer-readable recording medium and can be transmitted via an electric communication line.

[0019] The communication unit 20 includes a communication interface for communication between the communication unit 20 and the deep learning device 1. The communication unit 20 communicates with the deep learning device 1 via a wired or wireless connection. The communication unit 20 receives facility-related information transmitted by the deep learning device 1 via the network 9. The communication unit 20 also functions as an acquisition unit that receives various information including target affected area information from an external device via the network 9. The communication unit 20 may communicate with an external storage device such as a USB and acquire various information including target affected area information from the external storage device. The communication unit 20 also transmits information indicating dose distribution candidates in each cross section of the target affected area (hereinafter referred to as "dose distribution candidate information") to the external device via the network 9. The communication unit 20 may transmit the dose distribution candidate information to an external storage device such as a USB.

[0020] The control unit 21 controls the operation of each functional unit of the dose distribution generation device 2. The control unit 21 acquires candidate dose distribution information for each cross section of the target affected area based on the facility-related information corrected by the patient treatment policy.

[0021] The input unit 22 includes a known input device, similar to the input device of the deep learning device 1. The input unit 22 may be configured by connecting these input devices to the input unit 22. The input unit 22 accepts user operations on the input unit 22.

[0022] The storage unit 23 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 23 stores various information related to the dose distribution generation device 2. For example, the storage unit 23 stores target affected area information input to the input unit 22, facility-related information received via the communication unit 20, patient treatment policies, patient-related information, dose distribution candidate information, etc.

[0023] The output unit 24 outputs various information. For example, the output unit 24 outputs composite result information, which is the sum of doses at each position within a plane, based on target affected area information, dose distribution candidate information, and dose distributions within the same plane indicated by a plurality of dose distribution candidate information. The output unit 24 can also output and display dose distribution candidate information generated based on standard relationship information and relationship information from other facilities as reference information. The output unit 24 includes a display device, such as a CRT (Cathode Ray Tube) display, a liquid crystal display, or an organic EL (Electro-Luminescence) display. The output unit 24 may be configured by connecting such a display device to itself.

[0024] Here, the deep learning device and the dose distribution generation device are described as different devices, but the deep learning device 1 and the dose distribution generation device 2 do not necessarily need to be configured in different housings and may be configured in a single housing. Furthermore, the deep learning device 1 and the dose distribution generation device 2 may be implemented using multiple information processing devices connected to each other so as to be able to communicate via a network. In this case, the functional units of the deep learning device 1 and the dose distribution generation device 2 may be distributed and implemented in multiple information processing devices. For example, the input unit 22 and the control unit 21, and the output unit 24 may each be implemented in different information processing devices.

[0025] The following describes how this system is used. 1. Creation of Facility Relation Information [Aspect Using Multiple Standard Relation Information] Figure 2 shows an example of the processing flow executed by the deep learning device in this embodiment. The communication unit 14 receives explanatory affected area information and target dose distribution information, which serve as training data (step S101). The control unit 11 acquires relationship information through dose distribution learning based on the explanatory affected area information and target dose distribution information (step S102). Here, the training data for deep learning (target dose distribution information associated with explanatory affected area information) uses a dataset selected from dose distributions used to treat a specific tumor by multiple specialists at a single facility. Numerous radiation treatments have been performed on a specific tumor, and treatment plans recognized by the specialists as appropriate are selected as training data. The training results are then stored in the storage unit 13 as standard relationship information (step S103). As described below, because dose distributions vary significantly depending on the irradiation technique, multiple training datasets are prepared for each type, and the training results are stored as standard relationship information. The algorithm for the specified deep learning may be any deep learning algorithm or model that can learn the relationship between the position of the affected area, the cross-sectional contour, and the dose distribution, and any algorithm may be used, such as an autoencoder, a recurrent neural network, or a convolutional neural network.

[0026] The relationship information may be generated in any manner, but highly accurate relationship information can be obtained by a method using deep learning, as already disclosed by the present inventors in Patent Document 3. Specifically, the relationship information may be the learning results of deep learning, in which information indicating the position of the affected area and the outline of the cross section is used as an explanatory variable and information indicating the dose distribution of radiation irradiated to the affected area is used as a target variable. Furthermore, the dose distribution indicated by the target dose distribution information input to the deep learning device 1 may be a dose distribution normalized by the average value of the dose in the target affected area. Note that the mean-normalized dose distribution is a dose distribution normalized by the average value of the dose in the target affected area. Reference relationship information obtained from a dataset deemed appropriate according to the above criteria may be used as the standard relationship information.

[0027] Because dose distributions vary significantly depending on the irradiation technique, standard relationship information is created for each irradiation technique. Standard relationship information can also be created for protocols with significantly different dose distributions. The standard relationship information is the result of learning treatment data performed by doctors at a specific medical facility (hereinafter simply referred to as the facility) using a radiation irradiation device at the medical facility from which the training data was obtained. Therefore, although very high-quality relationship information has been accumulated, it may need to be modified when used at another facility due to differences in treatment devices, etc. Furthermore, treatment policies, such as prioritizing the planning target volume (PTV) or OAR, may differ depending on the facility. Therefore, the deep learning device 1 stores multiple standard relationship information obtained from multiple medical facilities with different irradiation techniques, etc., and prepares it so that it can be selected for each facility. After selecting an irradiation technique, etc., an institution using this system to perform radiation therapy can select the model of radiation device used at the medical facility and the treatment policy to modify the standard relationship information to suit the facility.

[0028] Each facility inputs the irradiation technique, the name of the treatment device to be used, and the facility's treatment policy as the facility treatment policy via the input unit 12 (step 104). Figure 3 shows an example of the facility treatment policy selection screen. Each facility can select and input the protocol, plan policy, treatment device, irradiation technique, etc. used. Here, a protocol refers to a standardized procedure or guideline followed when performing radiation therapy, including the dose administered to the target, the dose constraints for normal organs that serve as indicators for side effect management, treatment frequency, and treatment duration. A plan policy is a policy that determines the priority policy for each case, and refers to priority policies such as dose administration to the tumor and reduction of the dose administered to normal organs. The treatment device used at each facility can be selected from the "Treatment Device" menu. Examples of irradiation techniques include fixed multi-port IMRT, which irradiates radiation from multiple fixed angles, and intensity-modulated rotational radiotherapy (VMAT), which is a rotational IMRT that irradiates radiation continuously while rotating the device.

[0029] The obtained facility treatment policy is compared with multiple standard relationship information. Because the training data used to create the standard relationship information also contains protocols, plan policies, etc., the control unit 11 comprehensively evaluates this information and selects the most similar standard relationship information (step S105). The control unit 11 transmits the standard relationship information as facility relationship information to the dose distribution generation device 2 via the communication unit 14 (step S106).

[0030] In this aspect, since multiple pieces of training data are prepared, it is possible to determine, for example, how the setting of the plan policy affects the dose distribution based on the training data, and therefore it is possible to determine and select which standard relation information provides the dose distribution candidate that is closest to the setting of the facility treatment policy.

[0031] [Aspect Using Single Standard Relationship Information] Another aspect of generating facility relationship information will be described ( FIG. 4 ). Similarly to the above, the description of affected area information and target dose distribution information are acquired (step S111), and relationship information is acquired by deep learning (step S112). In this aspect, at least one learning result for each irradiation technique, etc., with significantly different dose distributions, is stored as standard relationship information (step S113). A facility treatment policy is similarly acquired (S114), and the standard relationship information is modified using the facility treatment policy (S115). The relationship information modified to suit the facility is stored as facility relationship information (S116), and the learning result is transmitted to the dose distribution generation device 2 (step 117).

[0032] An example of how to modify standard-related information in accordance with facility-related information plan policies is shown below. Figure 5 shows a schematic diagram of the dose distribution when administering radiation to a PTV with an adjacent OAR. The "basic model" shows a schematic diagram of the dose distribution based on standard-related information. The basic model administers a high radiation dose to the PTV, while also administering some radiation to the OAR. Although the diagram is a schematic diagram, both the PTV and the OAR have complex three-dimensional shapes, with intricate outer edges. While a high radiation dose can generally be administered to the PTV, it cannot be said that it is administered reliably. On the other hand, the "target-priority dose administration model" administers radiation slightly outside the target's contour, ensuring reliable administration of radiation to the PTV. According to this model, a high radiation dose is administered to the PTV, but radiation administration to the OAR is unavoidable. In contrast, the two models in the bottom row prioritize protection of the OAR, with model (2) administering radiation with greater consideration for the OAR.

[0033] The standard relationship information can be modified by adjusting the weighting of the amount of increase in the dose to the PTV from the basic model when administration to the PTV is prioritized in the method used for learning, such as a convolutional neural network, and by adjusting the weighting for the organs at risk or by modifying the input data for the intersection area between the PTV and OAR when administration to the PTV is prioritized (Figures 5 and 6). The models shown in Figure 5 show the results of learning in which the weighting for the organs at risk was modified in the method used for learning.

[0034] As shown in Figure 6, reducing the dose to the organ at risk can also be achieved by modifying the input data. As shown in the lower right of Figure 6, the dose to the OAR can be reduced by setting the actual PTV region smaller in the contact area with the OAR. For simplicity, the example described here is one in which there is only one OAR. However, if there are multiple OARs, the risk for each organ can be taken into account, and modifications can be made to take into account the dose for each OAR. Which method to adopt depends on the OAR and the target tumor, so it is best to allow medical professionals to set it appropriately.

[0035] Either the method of selecting facility-related information from a plurality of standard relationship information or the method of modifying a single standard relationship information is used, and the learning result adapted to the facility is transmitted as facility-related information to the dose distribution generation device 2. After selecting from a plurality of standard relationship information, further modification may be made according to the facility treatment policy.

[0036] 2. Creating Patient-Related Information Next, we will explain how to generate a dose distribution appropriate for a patient based on facility-related information. For example, depending on the patient's age and condition, OAR protection may be prioritized over treating the target tumor. In addition, when multiple OARs exist, it may be necessary to determine which OAR to prioritize and adjust the intensity of the OAR dose reduction. In addition, when radiation therapy and chemotherapy are used in combination, the intensity of the radiation therapy may be adjusted. Physicians must select the optimal treatment plan by taking into account all of these individual patient conditions.

[0037] FIG. 7 shows an example of the flow of processing executed by the dose distribution generation device 2 in this embodiment. The control unit 21 acquires facility-related information via the communication unit 20 (step S201). The communication unit 20 receives target area information of the patient to be treated (step S202). The patient's treatment plan is acquired as a patient treatment policy via the input unit 22 (step 203). The control unit 21 modifies the facility-related information stored in the storage unit 23 based on the acquired patient treatment policy (step 204). The facility-related information can be modified based on the patient treatment policy in accordance with the method for creating facility-related information. The modified information based on the patient treatment policy is stored in the storage unit 23 as patient-related information (step 205). The control unit 21 generates dose distribution candidates for each cross section of the target area indicated by the target area information based on the patient-related information and the target area information (step 206). Note that if the patient's treatment plan does not need to be changed or modified from the facility treatment policy, steps 203 to 205 can be skipped and the process proceeds to step 206. The communication unit 20 and the output unit 24 output the dose distribution candidate information generated in step 206 (step 207). The communication unit 20 outputs information indicating the dose distribution by transmitting the dose distribution candidate information to an external device via the network 9 or transmitting the dose distribution candidate information to an external storage device such as a USB, or both. When outputting, reference information, which will be described later, may be output together with the generated dose distribution information.

[0038] Reference information refers to information other than the dose distribution candidate information generated from the patient-related information. For example, it presents the type of dose distribution that would result from the dose distribution candidate information generated from the target affected area information without modifying the facility-related information or the standard-related information. Alternatively, it presents the type of dose distribution that would result if the standard-related information selected as the facility-related information were selected from other facilities. By presenting other dose distribution candidates as reference information, medical professionals can further consider whether the dose distribution candidate information is appropriate for the patient and make modifications if necessary.

[0039] FIG. 8 shows an example of a display in which dose distribution candidate information created based on target affected area information and patient-related information is output together with reference information. As described above, the reference information is dose distribution candidate information created from facility-related information or standard-related information, or information indicating a dose distribution candidate created based on a treatment policy of another facility. Here, an example is shown in which two pieces of reference information with different protection against OAR are presented for the dose distribution candidate information, but any number of pieces of reference information may be presented, or in some cases, none may be presented at all. By presenting visualized reference information, a physician can easily determine whether the created dose distribution candidate is appropriate.

[0040] 1... deep learning device, 2... dose distribution generation device, 11, 21... control unit, 12, 22... input unit, 13, 23... storage unit, 14, 20... communication unit, 24... output unit, 91, 93... processor, 92, 94... memory, 100... dose distribution generation system

Claims

1. A method for generating a radiation therapy dose distribution using relationship information created by deep learning, comprising: a facility treatment policy acquisition step of acquiring a facility treatment policy at a medical facility; and a facility relationship information creation step of correcting standard relationship information, which is the learning result obtained by deep learning from explanatory affected area information and target dose distribution information, based on the facility treatment policy, to create facility relationship information.

2. The dose distribution generation method according to claim 1, wherein the facility-related information is selected from a plurality of standard related information in accordance with the facility treatment policy.

3. A dose distribution generation method according to claim 1 or 2, characterized in that the facility-related information adjusts the priority for PTV or OAR based on the standard related information and the facility treatment policy.

4. A dose distribution generation method as claimed in claim 2 or 3, comprising a target affected area information acquisition step for acquiring target affected area information, a patient treatment policy acquisition step for acquiring a patient treatment policy, and a patient related information acquisition step for correcting facility related information based on the patient treatment policy and acquiring patient related information.

5. A dose distribution generating method according to claim 4, further comprising a dose distribution generating step of generating candidate dose distribution information for a target patient based on target affected area information and patient-related information.

6. A dose distribution generating method according to claim 5, further comprising an output step of outputting the dose distribution candidate information and the reference information from an output unit.

7. A deep learning device having a deep learning unit that performs deep learning using information indicating the position of the affected area and the outline of a cross section as explanatory variables and information indicating the dose distribution of radiation to be irradiated to the affected area as objective variables, the deep learning device having an input unit that acquires a facility treatment policy, and a control unit that generates standard relationship information that is the learning result learned using the explanatory variables that serve as teacher data and the objective variables, modifies the standard relationship information using the facility treatment policy, and generates facility relationship information.

8. A dose distribution generation system comprising the deep learning device of claim 7 and a dose distribution generation device, wherein the dose distribution generation device comprises: an acquisition unit that acquires target affected area information, which is information indicating the position and cross-sectional contour of a target affected area that is an affected area to be irradiated with radiation, and a patient treatment policy; and a control unit that generates patient-related information from the facility-related information and the patient treatment policy generated by the deep learning device, and generates candidate dose distribution information for radiation to be irradiated to the target affected area based on the generated patient-related information and the target affected area information.

9. A program for causing a computer to execute the method according to any one of claims 1 to 6.

10. A trained model of standard relational information, which is the result of dose distribution learning using deep learning that learns the relationship between the position of the affected area, the cross-sectional contour, and the dose distribution.

11. A trained model of facility-related information as described in claim 10, which is modified by facility treatment policies.

Images