System, method, program, and recording medium

Abstract

To provide a technology to assist determination on the necessity of an absorptive tissue spacer indwelling operation for radiotherapy of prostate cancer.SOLUTION: A system for assisting determination on the necessity of an absorptive tissue spacer indwelling operation for radiotherapy of prostate cancer includes: an acquisition unit for acquiring subject information including at least one of a CT image, an MRI image, and organ contour information on a subject; a generation unit for generating a predicted dose distribution in a case where radiotherapy is performed for the subject using a learning model constructed by performing machine learning using a combination of at least one of the CT image and the MRI image, the organ contour information, and a dose distribution assumed in a case where radiotherapy is performed for the subject in the past subject as teacher data on the basis of the subject information; and an output unit for outputting prediction information regarding occurrence of an adverse event accompanying the radiotherapy for the subject, using the generated predicted dose distribution and a normal tissue disorder occurrence probability model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a system. [Background technology]

[0002] In radiotherapy for prostate cancer, the proximity of the prostate and rectum means that the rectum is also exposed to a radiation dose, which can result in rectal adverse events. For example, as described in Non-Patent Document 1, before the start of radiotherapy, a procedure is sometimes performed in which an absorbent tissue spacer for radiotherapy is placed in the fascia between the prostate and rectum. The absorbent tissue spacer for radiotherapy can physically separate the prostate and rectum, which is expected to reduce high-dose rectal exposure. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Mariados N, et al., Int J Radiat Oncol Biol Phys 2015; 92: 971-977. Summary of the Invention [Problem to be solved by the invention]

[0004] Absorbable tissue spacer placement for radiotherapy involves invasive procedures, such as placing a hydrogel spacer by puncturing the perineum under rectal ultrasound imaging guidance. Furthermore, absorbable tissue spacer placement for radiotherapy may result in complications such as infection, hematuria, and renal damage. Furthermore, Non-Patent Document 1 reports that the incidence of macroscopic bloody stools due to radiation proctitis is 2.0% with hydrogel spacer placement and 7.0% without. Therefore, performing absorbable tissue spacer placement for radiotherapy in all cases is inappropriate from the perspectives of invasiveness, potential complications, and efficacy, and may lead to overtreatment. Therefore, it is desirable to determine whether or not to perform absorbable tissue spacer placement for radiotherapy on a patient-by-patient basis. Therefore, a technology that can assist in determining whether or not absorbable tissue spacer placement for radiotherapy is necessary for prostate cancer is needed. [Means for solving the problem]

[0005] The present invention can be realized as the following aspects.

[0006] (1) According to one aspect of the present disclosure, there is provided a system for assisting in determining whether or not to perform an absorbent tissue spacer placement procedure for radiotherapy of prostate cancer. The system includes an acquisition unit that acquires subject information, including at least one of CT images, MRI images, and organ contour information of a subject; a generation unit that generates a predicted dose distribution for radiotherapy of the subject based on the subject information using a learning model constructed by performing machine learning using a combination of at least one of CT images and MRI images of past subjects, the organ contour information, and an expected dose distribution for radiotherapy of the subject as training data; and an output unit that outputs predictive information regarding the occurrence of adverse events associated with radiotherapy of the subject using the generated predicted dose distribution and a normal tissue damage probability model. The system of this aspect outputs predictive information regarding the occurrence of adverse events associated with radiotherapy of the subject using the generated predicted dose distribution and the normal tissue damage probability model, thereby assisting in determining whether or not to perform an absorbent tissue spacer placement procedure for radiotherapy of prostate cancer.

[0007] (2) In the system described in (1) above, the prediction information may include a probability of occurrence of the adverse event, and the output unit may output information regarding the necessity of radiation therapy absorbent tissue spacer placement using the probability of occurrence and a predetermined threshold. According to this form of the system, information regarding the necessity of radiation therapy absorbent tissue spacer placement is output using the probability of occurrence of the adverse event and a predetermined threshold, making it easier to determine the necessity of radiation therapy absorbent tissue spacer placement.

[0008] (3) In the system described in (1) or (2), the subject information may include a CT image. According to this system, the subject information includes a CT image, which can improve the accuracy of the predicted dose distribution, thereby improving the accuracy of the prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject.

[0009] (4) In the system according to any one of (1) to (3), the subject information may include a planned target volume as the organ contour information. According to this system, the subject information includes a planned target volume as the organ contour information, which can improve the prediction accuracy of the predicted dose distribution, thereby improving the accuracy of prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject.

[0010] (5) In the system according to any one of (1) to (4), the adverse event may include late rectal bleeding. This system outputs predictive information regarding the occurrence of late rectal bleeding associated with radiation therapy for a subject, thereby assisting in determining whether or not an absorbable tissue spacer placement procedure for radiation therapy of prostate cancer is required.

[0011] (6) According to another aspect of the present disclosure, there is provided a method executed by a system for assisting in determining whether or not to perform an absorbent tissue spacer placement procedure for radiotherapy of prostate cancer. The method includes the steps of: acquiring subject information about a subject, the subject including at least one of CT images, MRI images, and organ contour information; generating a predicted dose distribution for radiotherapy of the subject based on the subject information using a learning model constructed by performing machine learning using a combination of at least one of CT images and MRI images of past subjects, the organ contour information, and an expected dose distribution for radiotherapy of the subject as training data; and outputting prediction information regarding the occurrence of adverse events associated with radiotherapy of the subject using the generated predicted dose distribution and a normal tissue damage probability model. According to this aspect of the method, the generated predicted dose distribution and the normal tissue damage probability model are used to output prediction information regarding the occurrence of adverse events associated with radiotherapy of the subject, thereby assisting in determining whether or not to perform an absorbent tissue spacer placement procedure for radiotherapy of prostate cancer.

[0012] (7) According to another aspect of the present disclosure, there is provided a program for causing a computer to execute the method described in (6) above. This program uses the generated predicted dose distribution and a normal tissue damage probability model to output predictive information regarding the occurrence of adverse events associated with radiation therapy in a subject, thereby assisting in determining whether or not absorbable tissue spacer placement for radiation therapy of prostate cancer is necessary.

[0013] (8) According to another aspect of the present disclosure, there is provided a computer-readable storage medium storing the program described in (7) above. This storage medium uses the generated predicted dose distribution and a normal tissue damage probability model to output predictive information regarding the occurrence of adverse events associated with radiation therapy in a subject, thereby assisting in determining whether or not absorbable tissue spacer placement for radiation therapy of prostate cancer is necessary.

[0014] The present disclosure can be realized in various forms, such as a device for assisting in determining whether or not absorbent tissue spacer placement is necessary for radiotherapy of prostate cancer, a system for determining whether or not absorbent tissue spacer placement is necessary for radiotherapy of prostate cancer, a system for predicting the occurrence of adverse events associated with radiotherapy of prostate cancer, etc. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a block diagram illustrating a schematic configuration of a system according to an embodiment of the present disclosure. [Figure 2] 10 is a flowchart illustrating an example of processing executed by the system. [Figure 3] FIG. 1 is an explanatory diagram showing an outline of the architecture of a 2D U-net model for dose prediction. [Figure 4] FIG. 10 is an explanatory diagram showing a typical example of dose prediction. [Figure 5] FIG. 10 is an explanatory diagram showing a typical example of dose prediction. [Figure 6] FIG. 10 is an explanatory diagram showing a typical example of dose prediction. [Figure 7]FIG. 10 is an explanatory diagram showing plots of iDSC for cross-validation and test cases. [Figure 8] FIG. 1 is an explanatory diagram showing the results of a linear regression analysis regarding the goodness of fit between theoretical values ​​and predicted values ​​of the probability of occurrence of adverse events. DETAILED DESCRIPTION OF THE INVENTION

[0016] FIG. 1 is a block diagram illustrating a schematic configuration of a system 100 according to an embodiment of the present disclosure. The system 100 of the present disclosure is a system 100 for assisting in determining whether or not an absorbable tissue spacer placement procedure for radiotherapy for prostate cancer is required. In the present disclosure, "absorbable tissue spacer placement procedure for radiotherapy" refers to a procedure in which an absorbable tissue spacer for radiotherapy is placed in the gap between the prostate and the rectum prior to radiotherapy. By using an absorbable tissue spacer placement procedure for radiotherapy, the rectum, a normal organ, can be isolated from the prostate, which is the target of irradiation in radiotherapy for prostate cancer, thereby reducing the absorbed dose in the rectum during radiotherapy. Examples of absorbable tissue spacers for radiotherapy include, but are not limited to, hydrogel spacers such as SpaceOAR (registered trademark) (manufactured by Boston Scientific) and Neskeep (registered trademark) (manufactured by Alfresa Pharma).

[0017] Generally, in radiation therapy for prostate cancer, a CT scan of the patient is taken after a radiation therapy absorbent tissue spacer placement procedure is performed, and a doctor creates a radiation therapy plan based on the CT scan. The present inventors envisioned a scenario in which a CT scan of the patient is performed, a radiation therapy plan is created based on the CT scan, and the doctor performs the radiation therapy absorbent tissue spacer placement procedure if the risk of adverse events is deemed high, and omits the radiation therapy absorbent tissue spacer placement procedure if the risk of adverse events is deemed low. The radiation therapy absorbent tissue spacer placement procedure changes the position and shape of the organ. Therefore, to ensure irradiation accuracy in radiation therapy, it is assumed that a CT scan is taken again after the radiation therapy absorbent tissue spacer placement procedure, and a radiation therapy plan is created before the radiation therapy is performed. Generally, in radiation therapy planning, beam shapes and other parameters are determined based on multifaceted information, including the characteristics of the irradiation device and the patient's specific conditions (body size, organ location, tumor progression, etc.). Each time, a dose distribution simulation is performed, and the dose distribution is evaluated and modified. Furthermore, because there is a complex trade-off between the effectiveness of radiation therapy and the occurrence of adverse events in normal organs, finding the optimal solution in radiation therapy planning is difficult. For this reason, radiation therapy planning requires the doctor's skill and experience, as well as a long period of trial and error. Therefore, performing an additional radiation therapy plan before radiation therapy absorbent tissue spacer placement for the purpose of evaluating adverse events not only places a burden on the doctor, but also leads to delays in prostate cancer treatment, which is disadvantageous to the patient, making it difficult to implement in reality.

[0018] The system 100 of the present disclosure can assist in determining whether or not to perform an absorbable tissue spacer placement procedure for radiotherapy based on prediction information regarding the occurrence of adverse events. Examples of adverse events include, but are not limited to, radiation proctitis. Radiation proctitis is a serious side effect that occurs in radiation therapy for prostate cancer, and generally manifests as symptoms such as bleeding, diarrhea, and frequent bowel movements. While acute or late reactions may be anticipated as adverse events, it is preferable to anticipate late reactions from the perspective of severity. Therefore, it is preferable to anticipate late rectal bleeding as an adverse event, and it is more preferable to anticipate late rectal bleeding of grade 2 or higher. According to this embodiment, it is possible to assist in determining whether or not to perform an absorbable tissue spacer placement procedure for radiotherapy based on prediction information regarding the occurrence of late rectal bleeding, which is a serious adverse event.

[0019] The system 100 of this embodiment is configured to include a device such as a computer on which the program of the present disclosure is installed. The device is not particularly limited, but examples thereof include a radiation therapy planning device and a radiation therapy device. The system 100 may be realized by a single device, or may be realized by multiple devices that can exchange data with each other via a network or the like. The system 100 includes an acquisition unit 10, a generation unit 20, and an output unit 30. The system 100 of this embodiment is realized by a CPU (Central Processing Unit), which executes the program of the present disclosure to function as the acquisition unit 10, the generation unit 20, and the output unit 30, respectively.

[0020] The acquiring unit 10 acquires subject information including at least one of a CT (Computed Tomography) image, an MRI (Magnetic Resonance Imaging) image, and organ contour information of the subject. The acquiring unit 10 may acquire the subject information from an external server via a network or the like, or may acquire the subject information stored in a storage device of a computer included in the system 100.

[0021] Subjects in the present disclosure include prostate cancer patients, who may also have other cancers, and who have either undergone endocrine therapy or endocrine therapy, who have either undergone chemotherapy or chemotherapy, or who have either undergone immunotherapy or immunotherapy.

[0022] Both CT images and MRI images are three-dimensional medical images. CT images and MRI images are three-dimensional reconstructions of multiple consecutively acquired two-dimensional slice images. CT images and MRI images are obtained by imaging the area containing the subject's prostate using a CT scanner and an MRI scanner, respectively.

[0023] In the present disclosure, "organ contour information" refers to information regarding the contours of a subject's organs and is generated based on three-dimensional medical images. Examples of organ contour information include, but are not limited to, the prostate contour, rectum contour, bladder contour, body contour, left femoral head, and right femoral head. The organ contour information in the present disclosure also includes various volumetric information such as gross tumor volume (GTV), clinical target volume (CTV), internal target volume (ITV), and planning target volume (PTV). The GTV is defined as the tumor contour identified using three-dimensional medical images such as MRI images, biopsy, or visual and tactile examination. The CTV corresponds to an area expanded from the GTV, taking into account microscopic infiltration around the tumor. In prostate cancer radiotherapy, the CTV includes, for example, a portion of the prostate and seminal vesicles. The ITV corresponds to an area expanded from the CTV, taking into account errors due to physiological organ movement. The PTV is equivalent to the CTV or ITV expanded to allow for patient positioning errors. For example, the PTV is equivalent to the CTV expanded approximately 4 to 8 mm ventrally, craniocaudal, dorsal, and lateral. Irradiation is usually performed so that a specified radiation dose is administered to the PTV, but multiple GTVs, CTVs, ITVs, and PTVs may be set up to administer different doses to different regions.

[0024] The organ contour information may be one type of information as described above, but from the viewpoint of improving the prediction accuracy of dose distribution described below, it is preferable that the organ contour information be two or more types, more preferably three or more types, even more preferably four or more types, and even more preferably five or more types. From the viewpoint of improving the prediction accuracy of dose distribution, the organ contour information preferably includes a planning target volume, more preferably includes a planning target volume or prostate contour, even more preferably includes a planning target volume or prostate contour and a rectum contour, even more preferably includes a planning target volume or prostate contour, a rectum contour, and a bladder contour, and even more preferably includes a planning target volume or prostate contour, a rectum contour, a bladder contour, and a body contour. It is more preferable that the organ contour information be two or more types including the planning target volume, even more preferably includes two or more types including the planning target volume or prostate contour, and even more preferably includes three or more types including the planning target volume or prostate contour.

[0025] From the viewpoint of improving the prediction accuracy of the dose distribution, the subject information preferably includes a CT image, more preferably includes a CT image and organ contour information, and even more preferably includes a CT image and a planned target volume as organ contour information. Note that, as a result of being able to improve the prediction accuracy of the dose distribution, the accuracy of the prediction information described below can be improved.

[0026] The generating unit 20 generates a predicted dose distribution when radiation therapy is performed on the subject using a learning model based on the subject information acquired by the acquiring unit 10. The generating unit 20, for example, automatically extracts spatial features from the subject information and predicts and generates, using the learning model, the dose distribution that is most likely to be created by a radiation therapist or medical physicist in a radiation therapy plan for the subject. The generating unit 20 may use, as the learning model, a learning model on an external server via a network or the like, or a learning model stored in a storage device of a computer included in the system 100. The generating unit 20 may include a learning device for constructing the learning model.

[0027] The learning model is constructed by performing machine learning using a combination of at least one of CT images and MRI images of past subjects, organ contour information, and an estimated dose distribution when radiotherapy is performed on the subject as training data. The machine learning is not particularly limited, but from the viewpoint of appropriate learning, learning using a neural network (NN) is preferable, and learning using a convolutional neural network (CNN) is more preferable.

[0028] The previous CT images and MRI images of the subject used in constructing the learning model may be of the same type as the CT images and MRI images of the subject, or may be of different types. It is preferable that the CT images of the subject are used in constructing the learning model. The organ contour information of the subject used in constructing the learning model may be of the same type as the organ contour information of the subject, or may be of different types. The organ contour information of the subject used in constructing the learning model may be one of the above-mentioned organ contour information, but from the viewpoint of appropriate learning, it is preferable that it be two or more types, more preferably three or more types, even more preferably four or more types, and even more preferably five or more types. From the viewpoint of appropriate learning, the organ contour information used in constructing the learning model preferably includes a planning target volume, more preferably includes a planning target volume or a prostate contour, more preferably includes a planning target volume or a prostate contour and a rectum contour, even more preferably includes a planning target volume or a prostate contour, a rectum contour, and a bladder contour, and even more preferably includes a planning target volume or a prostate contour, a rectum contour, a bladder contour, and a body contour. It is more preferable that the organ contour information used in constructing the learning model is two or more types including the planned target volume, even more preferable that it is two or more types including the planned target volume or the contour of the prostate, and even more preferable that it is three or more types including the planned target volume or the contour of the prostate.

[0029] The radiation therapy for the subject is not particularly limited, but is preferably intensity-modulated radiation therapy (IMRT) from the viewpoint of improving therapeutic effects while suppressing the occurrence of adverse events. The dose distribution is a visualized distribution of radiation dose, and for example, a larger radiation dose is shown in red, and a smaller radiation dose is shown in blue. The dose distribution expected when radiation therapy is performed on the subject used in building the learning model is not particularly limited, but examples include a dose distribution created by a radiation therapist or medical physicist in a radiation therapy plan for the subject.

[0030] The output unit 30 uses the predicted dose distribution generated by the generation unit 20 and a normal tissue complication probability (NTCP) model to output prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject. In the present disclosure, the NTCP model refers to a mathematical model that predicts the probability of damage, which is the response probability of normal tissue, from the dose distribution. The NTCP model is not particularly limited, but it is preferable to use a model with default values ​​for parameters verified in past clinical trials, and it is more preferable to use the Lyman-Kutcher-Burman (LKB) model. Note that the output unit 30 may extract a dose volume histogram (DVH) curve for the rectum from the predicted dose distribution generated by the generation unit 20, and output prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject using the extracted rectal DVH curve and the NTCP model.

[0031] The predictive information regarding the occurrence of an adverse event is not particularly limited, but preferably includes the probability (%) of the occurrence of the adverse event. The predictive information is not limited to the probability (%) of the occurrence of the adverse event, but may also include the presence or absence of the adverse event, or may include a binary or multi-value classification representing the nature of the adverse event, such as the type or grade of the adverse event. More specific predictive information may include, for example, a 15% or higher probability of late rectal bleeding of grade 2 or higher. The output unit 30 may also output information regarding the necessity of absorbable tissue spacer placement for radiotherapy using the probability of the occurrence of the adverse event and a predetermined threshold. More specifically, the output unit 30 may output information indicating that an absorbable tissue spacer placement for radiotherapy is likely to be necessary if the probability of the adverse event is equal to or greater than a predetermined threshold, or may output information indicating that an absorbable tissue spacer placement for radiotherapy is likely not necessary if the probability of the adverse event is less than the predetermined threshold. The predetermined threshold is not particularly limited, but is preferably set to a value between 5% and 25%, and more preferably between 10% and 20%. Alternatively, the predetermined threshold may be set to, for example, the median NTCP value in the training data.

[0032] FIG. 2 is a flowchart showing an example of processing executed by the system 100 of this embodiment. First, subject information including at least one of CT images, MRI images, and organ contour information of a subject is acquired (step S10). Next, a predicted dose distribution when radiation therapy is performed on the subject is generated based on the subject information using a learning model constructed by machine learning using a combination of at least one of CT images and MRI images of past subjects, the organ contour information, and the dose distribution expected when radiation therapy is performed on the subject as training data (step S20). Next, using the generated predicted dose distribution and a normal tissue damage occurrence probability model, prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject is output (step S30). This completes the processing.

[0033] In the system 100 according to an embodiment of the present disclosure, the output unit 30 uses the predicted dose distribution generated by the generator 20 using subject information and a learning model, and the NTCP model to output prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject. Therefore, a physician can determine whether or not to perform an absorbable tissue spacer placement procedure for radiotherapy for prostate cancer based on the output prediction information. Therefore, the system 100 of the present disclosure can assist in determining whether or not to perform an absorbable tissue spacer placement procedure for radiotherapy for prostate cancer. As a result, it is possible to determine whether or not to perform an absorbable tissue spacer placement procedure for radiotherapy for each patient, thereby preventing excessive performance of an absorbable tissue spacer placement procedure for radiotherapy. Therefore, for patients predicted to have a low probability of adverse events, radiation therapy absorbable tissue spacer placement can be omitted, thereby reducing the invasiveness and complications associated with radiation therapy absorbable tissue spacer placement and preventing a prolonged treatment period. Furthermore, in patients predicted to have a low probability of adverse events, the placement of an absorbent tissue spacer for radiotherapy can be omitted, thereby reducing excessive medical treatment.

[0034] Furthermore, the system 100 of the present disclosure can assist in determining whether or not absorbent tissue spacer placement for radiotherapy is necessary, thereby eliminating the need for radiation therapy planning prior to the procedure. As a result, the personnel and time costs required for such radiation therapy planning can be reduced. Furthermore, the generation unit 20 generates a predicted dose distribution for a subject when radiotherapy is performed based on subject information using a learning model constructed by performing machine learning using a combination of at least one of past CT images and MRI images of the subject, organ contour information of the subject, and a dose distribution expected if radiotherapy is performed on the subject as training data. This prevents a decrease in the prediction accuracy of the subject's predicted dose distribution. Furthermore, by having the output unit 30 output information regarding the need for absorbent tissue spacer placement for radiotherapy, the need for radiation therapy can be more easily determined.

[0035] In the system 100 of the present disclosure described above, any part or all of the functional units may be realized by a program. Such a program may be stored in a computer-readable storage medium and distributed, may be distributed via a communication line such as the Internet, or may be distributed in a state where it is installed on any terminal. [Example]

[0036] The present invention will be explained in more detail below with reference to examples, but the present invention is not limited to the following examples.

[0037] 1. Subjects and Methods (1) Target patients The study involved 75 patients who underwent intensity-modulated radiotherapy for prostate cancer at Nagoya City University Hospital between 2015 and 2018. Of the 75 treatment plan datasets, 60 were used as pre-training data, and the remaining 15 were used in experiments to predict the probability of rectal adverse events and to evaluate their performance.

[0038] (2) Treatment procedure Treatment planning CT scans were performed using a GE Optima CT580W (GE Healthcare Technologies, Inc.). Radiation treatment plans were manually planned by physicians using RayStation version 4.5 (RaySearch Laboratories). Dose calculations were performed using a collapsed cone convolution algorithm with a resolution of 2.0 × 2.0 × 2.0 mm. All treatments were performed using volumetric modulated arc therapy (VMAT), a type of IMRT. VMAT was performed using a TrueBeam (Varian Medical Systems, Inc.) photon beam with a 10 MV beam, with one or two coplanar full arcs. The clinical target volume (CTV) and planning target volume (PTV) were defined according to the method described by Kita N., et al., J Radiat Res 2022; 63: 666-674. Regions of interest (ROIs) for all organ contours were contoured within the craniocaudal region of the PTV ± 1 cm. 50% Using the formula, a total dose of 74.8 Gy in 34 fractions was prescribed to the PTV.

[0039] (3) Pretreatment DICOM (Digital Imaging and Communications in Medicine) data for treatment planning CT scans, organ contour ROIs, and dose distributions for 75 patients were acquired from RayStation using pydicom version 2.3.1. For each horizontal slice, the DICOM data specifications were a matrix size of 512 × 512, a maximum number of slices of 81–156, a pixel spacing of 0.98 mm, and a slice thickness of 2.5 mm. The dose distribution matrix size was resized to 512 × 512 pixels by linear interpolation to match the 3D coordinates on the CT scan. To prevent overflow of GPU memory, the matrix size for each CT scan, organ contour ROI, and dose distribution was cropped to 400 × 400 pixels, resulting in a final size of 3.06 × 3.06 mm. 2 The images were resized to 128 × 128 pixels with a resolution of 128 × 128 pixels. All dose values ​​were rescaled so that 95% of the PTV matched 95% of the prescribed dose. Dose values ​​in each patient's dose distribution were normalized to the maximum dose value before training.

[0040] (4) Model training The correspondence between CT images and multiple organ contours and dose distributions for 60 cases was previously trained using a convolutional neural network. The details are described below.

[0041] Figure 3 shows an overview of the architecture of the 2D U-net model for dose prediction. As shown in Figure 3, the 2D U-net model consists of an input, an encoding block, a decoding block, skip connections, and an output. The input is a 7-channel 128 × 128 matrix containing CT scan data and six organ contours (body contour, PTV, rectum, bladder, left femoral head, and right femoral head). To combine the local feature maps and global feature maps, the outputs of each encoding block and decoding block are connected with skip connections. To enhance model training, residual connections were added between the corresponding convolutional layers of the encoding block and decoding block. To avoid overfitting, a dropout layer (with a dropout rate of 0.1) was applied after two batch normalization layers. The output is a predicted dose distribution in a single channel of a 128 × 128 matrix. Note that in Figure 3, the number above each box indicates the number of extracted feature maps, and the number to the left of each box indicates the matrix size of the feature maps.

[0042] For model training, 75 patients were randomly divided into a 60-patient subset and a 15-patient subset for training and testing. The number of input images was 4,784 for training and 1,200 for testing. The 60-patient training dataset was randomly divided into five validation subsets, each consisting of 12 patients, and five-fold cross-validation was performed. The final model was constructed by averaging the five U-net models obtained from each cross-validation. The predictive performance of the final model was evaluated using 15 test patients that were not used for model training. The loss function was defined as the mean squared error between the clinical dose distribution and the predicted dose distribution. Here, the clinical dose distribution refers to the dose distribution actually administered to the patient and was treated as the ground truth dose distribution in model training. The optimization algorithm for model training was Adam (adaptive moment estimation), with a learning rate of 1.0 × 10 -3We set β1=0.9, β2=0.999, weight decay=0, and mini-batch size 15. We trained the model for 250 epochs for each cross-validation run using an NVIDIA Tesla V100 GPU with 16GB of memory.

[0043] (5) Model performance and DVH analysis In all analyses, the dose distributions were scaled so that 95% of the PTV corresponded to 95% of the prescribed dose. To quantify the similarity between the clinical dose distribution obtained as three-dimensional data and the predicted dose distribution, the isodose volume dice similarity coefficient (iDSC) was calculated using the following equation (1): In equation (1), V represents the isodose volume to which the dose d (Gy) is irradiated.

[0044]

number

[0045] The mean iDSC was defined as the average of the iDSC values ​​for each d. Additionally, the following DVH indices were calculated: D of PTV 98% ,D 95% ,D 50% ,D 2% ,D mean , homogeneity index (HI(homogeneity index)=100·[D 2% -D 98% ] / [D 50% ]), R50%=100 V 50%isodose / V PTV ;V for rectum and bladder 70Gy ,V 60Gy ,V 50Gy ,V 40Gy ,D 2% ,D mean ;D of the left and right femoral heads 2% ,D mean .D x% is the minimum dose (Gy) that can be delivered to x% of the volume of an organ. V xGyis the percentage (%) of the volume of a given organ irradiated with more than x Gy. To evaluate the prediction accuracy of the dose volume histogram (DVH) and NTCP value between the clinical dose distribution and the predicted dose distribution, the percent mean absolute error (%MAE) was calculated. %MAE was defined as the ratio (%) of the mean absolute error (MAE) to the prescribed dose for endpoints in Gy, and as the MAE for endpoints in %.

[0046] (6)NTCP analysis The normal tissue complication probability (NTCP) model was used to convert the predicted dose distribution into the probability (%) of late rectal bleeding (LRB). The NTCP values ​​(%) for grade 1 or higher LRB (hereinafter also referred to as "G1-LRB-NTCP") and grade 2 or higher LRB (hereinafter also referred to as "G2-LRB-NTCP") were calculated based on the clinical dose distribution and the predicted dose distribution by CNN, respectively. These NTCP values ​​were considered as surrogate endpoints for LRB after prostate IMRT. The G2-LRB-NTCP values ​​were calculated using the Lyman-Kutcher-Burman (LKB) model and the Relative Seriality (RS) model. The LKB model was defined by the following equations (2) and (3). In equation (3), TD 50 represents the tolerable dose at which an adverse event occurs with a 50% probability, and m represents the steepness of the dose-response curve. eff represents the effective dose and is expressed by the following formula (4).

[0047]

number

[0048]

number

[0049]

number

[0050] In equation (4), n represents the volume effect of the organ, and the closer it is to 0, the higher the organ's serialization and the higher its reactivity to the high-dose band. i represents the relative volume of the organ in dose bin i. M represents the total number of dose bins. EQD2 i represents the 2 Gy equivalent dose in dose bin i. EQD2 i The parameters were calculated using a linear-quadratic model with α / β = 3.0 Gy, according to the method described in Fowler JF., Br J Radiol. 2010 Jul;83(991):554-68. The parameters for each endpoint of rectal adverse events were: G1-LRB-NTCP, n = 0.23, m = 0.37, TD 50 = 57.3 Gy; for G2-LRB-NTCP, n = 0.19, m = 0.32, TD 50 = 75.8 Gy; for G1 stool frequency, n = 0.27, m = 0.56, TD 50 = 55.7; for G2 stool frequency, n = 0.31, m = 0.36, TD 50 = 75.8 Gy f; G1 bowel pain (n = 0.17, m = 0.49, TD 50 = 142.6 Gy; for G1 sphincter control, n = 0.24, m = 0.32, TD 50 = 79.1 Gy; for G1 rectal stenosis or ulcer (G1 stricture / ulcer), n = 0.32, m = 0.25, TD 50 These parameters were verified by adverse event analysis of previous large-scale clinical trial data (Brand DH, et al., Int J Radiat Oncol Biol Phys 2021;110:596-608.), and each parameter (n, m, TD 50) was used as a maximum likelihood estimate. Grade 1 indicates the severity of symptoms requiring medical treatment among those with bloody stools, and grade 2 indicates the severity of symptoms requiring some kind of medical treatment among those with bloody stools.

[0051] The RS model is defined as follows: where s represents the seriality parameter, and v i represents the relative volume of the organ in dose bin i, M represents the total number of dose bins, and D i represents EQD2 in dose bin i. Also, P(D i ) represents the Poisson dose-response relationship and is given by the following equation (6): D 50 represents the tolerable dose at which an adverse event occurs with a 50% probability, and γ is D 50 represents the slope of the response curve at

[0052]

number

[0053]

number

[0054] In the RS model, the calculation of G2-LRB-NTCP requires D 50 The parameters used were γ = 83.6 Gy, γ = 1.42, s = 0.50, and α / β = 3.0 Gy. In the RS model, LRB grade was assessed using a modified RTOG protocol described by Rancati T. et al., Radiother Oncol 2004;73:21-32.

[0055] (7) Statistical analysis The primary endpoint was the predictive ability of the LKB-NTCP model to predict the incidence of grade 2 or higher LRB (G2-LRB-LKB-NTCP). The goodness of fit of the predictive ability was evaluated using the coefficient of determination (R) ranging from 0 to 1. 2) and a larger value indicates a better fit. For sensitivity analysis, the fit was also evaluated for G2-LRB-NTCP using the RS model (hereinafter also referred to as "G2-LRB-RS-NTCP"). P values ​​were calculated using the Wald test with the t-distribution of the test statistic, with the null hypothesis that the regression coefficient is equal to 0. With a sample size of 15, this study was performed using the R 2 The analysis had 80% power to detect a p < 0.35 at a significance level of α = 0.05. Goodness of fit was also assessed for other DVH indices. All statistical analyses were performed using SciPy version 1.7.3. P < 0.05 was considered statistically significant.

[0056] 2.Results Figures 4 to 6 show representative examples of dose prediction. Figures 4 to 6 show images of the best predicted case (patient 13, mean iDSC = 0.91) and the worst predicted case (patient 4, mean iDSC = 0.82). Figure 4 shows organ contours, the correct clinical dose distribution (Clinical), and the predicted dose distribution (Predicted). The dose distribution shows the absorbed dose (Gy) at a specific location on the CT image. Figure 5 shows a mapping of the dose difference (Predicted dose distribution - Clinical dose distribution), and Figure 6 shows a comparison of the dose-volume histogram between the clinical dose distribution (solid line) and the predicted dose distribution (dashed line). As shown in Figure 4, the predicted dose distribution exhibited a typical concave dose drop extending from the prostate toward the rectum, with an intermediate dose (30–50 Gy) extending toward the bladder, similar to the correct clinical dose distribution. Therefore, it was confirmed that the dose distribution and DVH predicted by CNN were generally consistent with the correct clinical dose distribution and DVH.

[0057] Figure 7 shows plots of iDSC for cross-validation and test cases. Figure 7(A) shows plots of iDSC for five-fold cross-validation (n = 12 each), and Figure 7(B) shows plots of iDSC for test (n = 15). Data represent the mean iDSC ± 1 standard deviation. The iDSC for the test cases ranged from 0.80 to 0.92. The mean iDSC was 0.87 ± 0.01 for five-fold cross-validation and 0.87 for the test cases, indicating high similarity between the clinical and predicted dose distributions.

[0058] The characteristics of the 15 test patients are shown in Table 1 below. The percent mean absolute error (%MAE) ± 1 standard deviation (%) for each evaluation index between the clinical dose distribution and the predicted dose distribution for the 15 test cases is shown in Table 2. The %MAE ± 1 standard deviation (%) was 2.22 ± 2.15% for G1-LRB-LKB-NTCP, 1.24 ± 1.42% for G2-LRB-LKB-NTCP (primary endpoint), and 0.23 ± 0.15% for G2-LRB-RS-NTCP. Thus, the %MAE was within 2.3%, demonstrating that late rectal bleeding of grade 1 or higher and grade 2 or higher can be predicted with a prediction error small enough for practical use. The range of percent absolute error was 0.01-6.96% for G1-LRB-LKB-NTCP and 0.02-5.31% for G2-LRB-LKB-NTCP. %MAE is PTV D 98% ,D 95% ,D5 0% ,D 2% Within 5.00%, rectal V 50Gy -V 70Gy Within 2.00%, bladder V 50Gy -V 70Gy was within 6.00%.

[0059] [Table 1]

[0060] [Table 2]

[0061] Figure 8 is an explanatory diagram showing the results of linear regression analysis regarding the goodness of fit between the theoretical and predicted values ​​of the probability of occurrence of adverse events. Figure 8(A) shows the results for late rectal bleeding of grade 1 or higher, and Figure 8(B) shows the results for late rectal bleeding of grade 2 or higher. In Figure 8, the vertical axis shows the theoretical value of the probability of occurrence of adverse events, the horizontal axis shows the predicted value of the probability of occurrence of adverse events, and the black line shows the regression line. The linear regression analysis showed a significant and strong correlation between the theoretical value and the predicted value of the probability of occurrence of adverse events. More specifically, for G1-LRB-LKB-NTCP, R 2 = 0.85, P < 0.001, and for the primary endpoint, G2-LRB-LKB-NTCP, R 2 = 0.80, P < 0.001. In other words, it was shown that rectal adverse events can be predicted with sufficient accuracy for practical use. Furthermore, as shown in Table 2, in G2-LRB-RS-NTCP, which was evaluated as a sensitivity analysis, R 2 = 0.93, P < 0.001, which was a significant and strong correlation, indicating that sufficient predictive ability for rectal adverse events was maintained even when the RS model was used. In addition, for other rectal adverse events, the R 2 =0.86, P<0.001, and for G2 stool frequency, R 2 =0.87, P<0.001, and in G1 bowel pain, R 2 =0.78, P<0.001, and in the G1 sphincter control, R 2 =0.85, P<0.001, and for G1 stricture / ulcer, R 2= 0.89, P < 0.001, demonstrating a consistent, significant and strong correlation. Therefore, sufficient predictive ability was maintained for rectal adverse events in general. The total time required from inputting medical images and organ contours to predicting dose distributions for all 15 cases was approximately 20 seconds, and the total time required from the predicted dose distributions to outputting predicted values ​​for the occurrence probabilities of all the above adverse events was approximately 97 seconds. In other words, the total time required to predict the occurrence probability of adverse events from medical images and organ contours for one case was approximately 7.8 seconds. Therefore, it can be said that the incidence rate of adverse events due to radiation therapy could be predicted automatically and quickly using medical images taken before radiation therapy. This suggests that the technology could be used to assist in determining the need for radiation therapy-related adverse events, such as selectively placing absorbable tissue spacers for radiation therapy in cases predicted to have a high incidence of adverse events.

[0062] The present invention is not limited to the above-described embodiments and can be realized in various configurations without departing from the spirit of the present invention. For example, the technical features in the embodiments and examples corresponding to the technical features in each aspect described in the Summary of the Invention section can be appropriately replaced or combined to solve some or all of the above-described problems or achieve some or all of the above-described effects. Furthermore, if a technical feature is not described as essential in this specification, it can be deleted as appropriate. [Explanation of symbols]

[0063] 10...acquisition unit, 20...generation unit, 30...output unit, 100...system

Claims

1. A system for assisting in determining whether or not an absorbable tissue spacer placement procedure for radiotherapy of prostate cancer is necessary, comprising: an acquisition unit that acquires subject information including at least one of a CT image, an MRI image, and organ contour information of the subject; a generating unit that generates a predicted dose distribution when radiation therapy is performed on the subject based on the subject information, using a learning model constructed by performing machine learning using a combination of at least one of CT images and MRI images of past subjects, organ contour information, and a dose distribution expected when radiation therapy is performed on the subject as training data; an output unit that outputs prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject using the generated predicted dose distribution and a normal tissue damage occurrence probability model; A system comprising:

2. 10. The system of claim 1, the prediction information includes a probability of occurrence of the adverse event; the output unit outputs information regarding the necessity of a radiotherapy absorbent tissue spacer placement procedure using the occurrence probability and a predetermined threshold value. system.

3. In the system according to claim 1 or claim 2, The subject information includes a CT image. system.

4. In the system according to claim 1 or claim 2, The subject information includes a planning target volume as the organ contour information. system.

5. In the system according to claim 1 or claim 2, The adverse events include late rectal bleeding. system.

6. A method carried out in a system for assisting in determining whether or not an absorbable tissue spacer placement procedure for radiotherapy of prostate cancer is necessary, comprising: acquiring subject information including at least one of a CT image, an MRI image, and organ contour information of the subject; generating a predicted dose distribution when radiation therapy is performed on the subject based on the subject information, using a learning model constructed by performing machine learning using a combination of at least one of CT images and MRI images of past subjects, organ contour information, and a dose distribution expected when radiation therapy is performed on the subject as training data; outputting prediction information regarding the occurrence of adverse events associated with radiation therapy for the subject using the generated predicted dose distribution and a normal tissue damage occurrence probability model; A method comprising:

7. A program for causing a computer to execute the method according to claim 6.

8. A computer-readable storage medium storing the program according to claim 7.

Images