Intelligent offline adaptive radiotherapy method and system based on time series prediction model

Abstract

The application discloses an intelligent offline self-adaptive radiotherapy method and system based on a time series prediction model, and comprises the following steps: S1, collecting radiotherapy patient data; S2, performing image quality improvement on CBCT, and performing delineation of a tumor target area and surrounding normal tissue organs on the CBCT after image quality improvement; S3, accurately calculating a clinically required radiotherapy plan dosimetry evaluation index on the CBCT after image quality improvement; S4, judging whether the radiotherapy plan needs to be adjusted according to the calculated dosimetry evaluation index; and S5, designing and training a time series prediction model for the radiotherapy plan dosimetry evaluation index to predict a time node at which the patient needs to perform offline self-adaptive radiotherapy. The technical scheme greatly improves the shortcomings of a conventional offline self-adaptive scheme, such as complicated process, time and labor consumption, and inconvenient clinical application, and reduces the inevitable errors introduced in the conventional offline self-adaptive scheme. In addition, the homogeneous effect is good, and the application is easy to popularize in clinical application.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy, and more particularly to an intelligent offline adaptive radiotherapy method based on a time series prediction model. Background Technology

[0002] Numerous studies both domestically and internationally have confirmed that offline adaptive radiotherapy can provide more precise radiation therapy for cancer patients, killing tumor cells to a great extent while protecting surrounding normal tissues and organs from extremely low doses of radiation. However, due to the many limitations and problems of conventional offline adaptive radiotherapy protocols, it is difficult to truly promote its use in clinical practice, and it remains a hot topic and challenge in current research.

[0003] Currently, the standard offline adaptive radiotherapy protocol in clinical practice mainly involves acquiring cone-beam computed tomography (CBCT) images of patients undergoing radiotherapy on the same day. Using specialized radiotherapy software (such as Varian's medical software Velocity), a synthetic CT scan is generated based on the acquired CBCT images through registration and other methods. The original radiotherapy plan is then recalculated on the synthetic CT scan, and the dose to the tumor target area and surrounding normal tissues and organs is assessed to determine if the plan needs to be readjusted.

[0004] The main problems with this solution are as follows:

[0005] (1) The process is cumbersome, time-consuming, labor-intensive, and requires strict hardware and software conditions, resulting in low practicality. First, currently, apart from Varian's medical linear accelerators such as Halcyon, most medical linear accelerators do not require patients to undergo CBCT scans for localization before each radiotherapy session. Therefore, most radiotherapy departments do not collect complete CBCT datasets of radiotherapy patients (typically, radiotherapy patients undergo 25-30 radiotherapy sessions, and most medical linear accelerators and radiotherapy departments only scan CBCT during the first 5 radiotherapy sessions, and only scan once every 5 subsequent sessions for the purpose of confirming tumor size, location, and morphology). Incomplete CBCT datasets will result in the inability to reconstruct the most accurate CT images. Second, converting CBCT into synthetic CT requires special radiotherapy software (such as Varian's medical software Velocity), or the development of algorithms to implement this step. For many primary-level hospital radiotherapy departments, it is impossible to achieve homogenized image processing, or even to implement this step. Finally, to ensure the accuracy of this plan, multiple tasks need to be completed for each CBCT scan of each radiotherapy patient, including synthetic CT reconstruction, dose calculation, and dose assessment. Assuming that a radiotherapy patient receives 25-30 radiotherapy sessions, the implementation of this plan is time-consuming and labor-intensive, making it extremely inconvenient for clinical application.

[0006] (2) Inevitable Error Introduction. Currently, there is no reliable and easily implemented technique to guarantee high accuracy in converting CBCT scans of radiotherapy patients into synthetic CT images. Due to the anatomical heterogeneity and varying tumor sensitivity to radiation among radiotherapy patients, the algorithm is generally unable to accurately identify and judge the differences in the nature of the mass within the tumor boundary when converting CBCT scans of radiotherapy patients into synthetic CT images. Therefore, converting CBCT scans from each radiotherapy scan for each patient into synthetic CT images may introduce unavoidable, and even clinically unacceptable, errors.

[0007] (3) Verifying the accuracy of synthetic CT is difficult to implement. Synthetic CT is an approximate CT image generated based on CBCT. Therefore, for the most accurate verification of synthetic CT, it is necessary to perform a routine localization CT scan on the radiotherapy patient and compare the actual localization CT image with the synthetic CT image to verify the accuracy of the synthetic CT. However, CT scans also cause radiation to radiotherapy patients. If the sole purpose is to verify the accuracy of synthetic CT, multiple CT scans will have a certain impact on the prognosis of radiotherapy patients and cannot be used as a routine verification method. Summary of the Invention

[0008] To overcome the aforementioned technical deficiencies, the present invention aims to provide an intelligent offline adaptive radiotherapy method based on a time series prediction model, comprising the following steps:

[0009] Step S1: Collect CBCT, EPDI (Electronic Field Imaging Device) projection image data, radiotherapy plan, localization CT, and delineation data of tumor target area and surrounding normal tissues and organs on localization CT for radiotherapy patients.

[0010] Step S2: Using the first neural network algorithm, the image quality of CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality.

[0011] Step S3: Using the second neural network algorithm, with the improved image quality of the CBCT, the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, and EPD projection image data as input, the dosimetric evaluation indicators of the radiotherapy plan required in clinical practice are accurately calculated on the CBCT.

[0012] Step S4: Determine whether the radiotherapy plan needs to be adjusted based on the calculated radiotherapy dosimetric evaluation index. If so, design a new plan or adjust the plan. The new radiotherapy plan will be implemented after verification.

[0013] Step S5: Design a time series prediction model for radiotherapy planning dosimetric assessment indicators. Use radiotherapy patient data from steps S1 to S4 to train the model and predict the time points when radiotherapy patients need to perform offline adaptive radiotherapy.

[0014] Furthermore, in step S2, the image quality of the CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality. Both of these can be output by the same neural network algorithm in multiple tasks.

[0015] Furthermore, the algorithm used in step S2 to delineate the tumor target area and surrounding normal tissues and organs on the CBCT after image quality improvement is an algorithm for precise deformation registration and / or automatic delineation.

[0016] Furthermore, step S4 is performed manually.

[0017] Furthermore, the input data in step S5 includes the patient's EPID, radiotherapy plan, CBCT, localization CT, localization CT delineation data, and radiotherapy plan dosimetric evaluation indicators.

[0018] Another aspect of the present invention provides an intelligent offline adaptive radiotherapy system based on a time series prediction model, comprising an input module, a first neural network module, a second neural network module, an adjustment module, and a prediction model module, wherein:

[0019] The input module is used to collect CBCT and EPID projection image data, radiotherapy plans, localization CT, and delineation data of tumor target area and surrounding normal tissues and organs on localization CT for radiotherapy patients.

[0020] The first neural network module uses the first neural network algorithm to improve the image quality of CBCT and delineate the tumor target area and surrounding normal tissues and organs on the CBCT with improved image quality.

[0021] The second neural network module uses the second neural network algorithm to accurately calculate the radiotherapy planning dosimetric assessment indicators required in clinical practice on the CBCT, taking the improved image quality of the CBCT, the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, and EPD projection image data as input.

[0022] The adjustment module determines whether the radiotherapy plan needs to be adjusted based on the calculated dosimetric evaluation indicators. If so, it designs a completely new plan or an adjusted plan. The new radiotherapy plan is implemented only after it has been validated.

[0023] The prediction model module utilizes a third neural network algorithm to design a time series prediction model for radiotherapy planning dosimetric assessment indicators. The model is trained using radiotherapy patient data from steps S1 to S4 to predict the time points when radiotherapy patients need to undergo offline adaptive radiotherapy.

[0024] Furthermore, in the first neural network module, the image quality of the CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality. These two tasks can be output by the same neural network algorithm.

[0025] Furthermore, in the first neural network module, the algorithm used in step S2 to delineate the tumor target area and surrounding normal tissues and organs on the CBCT after image quality improvement is an algorithm for precise deformation registration and / or automatic delineation.

[0026] Furthermore, the adjustment and verification plan in the adjustment module is completed manually.

[0027] Furthermore, the input data in the prediction model module includes EPID, radiotherapy plan, CBCT, localization CT, localization CT delineation data, and radiotherapy plan dosimetric evaluation indicators.

[0028] Compared with existing technologies, the above technical solution has the following advantages:

[0029] This intelligent offline adaptive radiotherapy method based on time series prediction models is a novel offline adaptive radiotherapy approach in this field, and it has the following main advantages:

[0030] (1) The entire process input uses the actual images and dose data of the radiotherapy patient, such as dose data acquired by CBCT and EPD. The processing of the actual patient images is only to improve the quality of CBCT images. Other than that, no transformation is performed on the actual patient images and doses, thus avoiding the errors that occur in the conventional offline adaptive radiotherapy process to the greatest extent.

[0031] (2) Steps S2, S3 and S5 are all implemented intelligently and automatically through the developed algorithm model. Only human evaluation of the final result is required, which greatly reduces the disadvantages of the conventional offline adaptive solution being cumbersome, time-consuming and labor-intensive.

[0032] (3) Good homogenization effect and easy clinical promotion. Most medical linear accelerators will collect the first five CBCT images. Through the time series prediction model developed in step S5, it is possible to predict the time node that offline adaptive radiotherapy needs to be performed in the future based on a small amount of CBCT data. The requirements for the amount of CBCT data and the hardware and software configuration of the department are low, and it can be applied in most radiotherapy departments and medical linear accelerators in a homogenized manner. Attached Figure Description

[0033] Figure 1 This is a flowchart of an intelligent offline adaptive radiotherapy method based on a time series prediction model.

[0034] Figure 2 A roadmap for adversarial generative networks (GDN) technology.

[0035] Figure 3 This is a diagram of the discriminator network structure.

[0036] Figure 4 This is a diagram of the generator structure.

[0037] Figure 5 This is the original CBCT image.

[0038] Figure 6 This is a CBCT image after image quality improvement.

[0039] Figure 7 This is a diagram of the UNET structure.

[0040] Figure 8 This is a roadmap for the second neural network algorithm.

[0041] Figure 9 This is a diagram of the UNET structure after adjusting the number of network layers.

[0042] Figure 10 This is a network structure diagram of dosimetric assessment indicators for radiotherapy planning.

[0043] Figure 11 This is a flowchart of time series prediction based on neural networks.

[0044] Figure 12 This is a flowchart of a graph-based prediction method. Detailed Implementation

[0045] The advantages of the present invention are further illustrated below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following detailed description is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0046] To improve the feasibility of offline adaptive radiotherapy in clinical practice, we designed a novel, intelligent, rapid, and reliable offline adaptive radiotherapy technique: an intelligent offline adaptive radiotherapy method based on a time-series prediction model. This method automates the entire offline adaptive radiotherapy process to the greatest extent possible, not only overcoming hardware and software limitations and achieving homogenization, but also simplifying the adaptive process significantly. For example... Figure 1 As shown, the implementation steps of this intelligent offline adaptive radiotherapy technology based on a time series prediction model are as follows:

[0047] Step S1: Collect the CBCT, EPD transmission imaging data of the radiotherapy patient on the day of treatment, the radiotherapy plan used for the treatment on the day, the localization CT, and the delineation data of the tumor target area and surrounding normal tissues and organs on the localization CT.

[0048] Step S2: Utilizing the first neural network algorithm to achieve multi-task output: Improving the image quality of the CBCT scan, and delineating the tumor target area and surrounding normal tissues and organs on the improved CBCT. Using a neural network deep learning method, taking the localization CT and CBCT as input, an improved CBCT is generated. Simultaneously, using the localization CT and the tumor target area and surrounding normal tissues and organs delineated by the physician on the localization CT, an algorithm of precise deformation registration and / or automatic delineation is employed to delineate the target area and surrounding normal tissues and organs on the improved CBCT.

[0049] The first neural network algorithm improves CBCT image quality mainly through GAN (Generative Adversarial Network) (GAN includes, but is not limited to, standard GAN, DCGAN, LSGAN, WGAN, CycleGAN, etc., or you can modify the network structure of GAN yourself. The specific algorithm needs to be determined in conjunction with the characteristics of clinical data in subsequent evaluation).

[0050] The input data for GAN is CBCT and localization CT. For example... Figure 2 To construct an adversarial generative network (GDN) technology roadmap, CBCT and localization CT first need to undergo data preprocessing, which includes: (1) transforming the HU values ​​in CBCT and localization CT to the same range; (2) normalizing the CBCT and localization CT after HU value transformation; and (3) performing deformation registration on CBCT and localization CT. During registration, CBCT is used as the reference sequence, and localization CT is used as the secondary sequence. Figure 3 and Figure 4 As shown, in GAN, CBCT is used as the initial data for the generator, and the registered localized CT is used as the ground truth to train the discriminator. After the network training reaches equilibrium, it is used to generate CBCT with image quality close to that of the localized CT. Figure 5 and6 .

[0051] like Figure 7 As shown, the first neural network algorithm delineates the tumor target area and surrounding normal tissues and organs on CBCT primarily using UNET or a similarly structured convolutional network. After registration between the localization CT and CBCT, two or more senior radiation oncologists review and correct the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, thus forming the delineation dataset on the CBCT. This dataset is then appropriately segmented for training, testing, and calibration of the convolutional network.

[0052] Step S3. Using the improved CBCT image quality, the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, and EPD projection image data as input, the second neural network algorithm is used to accurately calculate the clinically required radiotherapy plan dosimetric assessment indicators on the improved CBCT image quality. Taking lung cancer as an example, the clinically required radiotherapy plan dosimetric assessment indicators include the maximum dose to the tumor target area, the dose covered by 95% of the tumor volume, the lung volume covered by a 20Gy dose to surrounding normal tissues and organs, the average dose to the heart, the maximum dose to the spinal cord, and the average dose to the contralateral breast, etc. Based on the calculated dosimetric assessment indicators, it is determined whether the radiotherapy plan needs to be adjusted.

[0053] like Figure 8 The diagram shows the technical roadmap for the second neural network algorithm. The second neural network algorithm mainly consists of two parts: the first part is a neural network model for accurate dose calculation, and the second part is a neural network model for accurate output of radiotherapy planning dosimetric evaluation indicators.

[0054] The neural network model in the first part is mainly implemented using UNET or its variants. For example... Figure 9 As shown, the number of network layers can be adjusted based on the feature data. The network input consists of the equivalent water dose for each field in the radiotherapy plan performed on the radiotherapy patient that day (including information such as gantry angle, collimator angle, field shape, and flux, enabling accurate conversion to the actual radiation dose) and the CBCT image quality improved by the first neural network algorithm. The patient's radiotherapy dose transmission EPID data collected on the same day is used for dose calculation verification, or it can be used to construct the loss function of this part of the network. The network output is the radiotherapy patient dose on the CBCT image quality improved. Data augmentation methods such as cyclic strategies and image matrix transformations (rotation, distortion, etc.) are adopted according to specific circumstances.

[0055] The neural network model in the second part mainly consists of convolutional networks and linear networks. For example... Figure 10As shown, the number of network layers can be adjusted based on the feature data. The input to this part of the network is the radiotherapy patient dose output by the neural network model of the first part of the second neural network algorithm and the delineation data output by the first neural network algorithm. The output is the radiotherapy planning dosimetry evaluation index.

[0056] Step S4. If the radiotherapy plan needs adjustment, the plan is adjusted on a CBCT scan with improved image quality or a new plan is designed using a re-localization CT scan. After plan validation, it is used for the subsequent clinical treatment of the patient. If no adjustment to the radiotherapy plan is required, the original radiotherapy plan is used for clinical treatment. This step is performed manually.

[0057] Step S5. Design a time-series prediction model for radiotherapy planning dosimetric assessment indicators, and train the model using a large amount of radiotherapy patient data that has undergone offline adaptive adjustments in steps S1 to S4 until it is stable and accurate. After a patient has received a few radiotherapy sessions (3-5 sessions), use this model to predict the time point at which the patient will need to undergo offline adaptive radiotherapy in the future, providing reference guidance for the clinical use of offline adaptive radiotherapy.

[0058] The initial time series prediction model primarily consisted of a convolutional network and a linear network (its network structure included, but was not limited to, convolutional networks such as LSTM (Long Short-Term Memory) and Transformer). Transfer learning was used to train the time series prediction network, such as... Figure 11 As shown, a time-series prediction network is first trained using a large dataset. Then, dosimetric assessment indicators after a few (3 to 5) radiotherapy sessions for a specific subset of validation patients are used as input to the network, outputting dosimetric assessment indicators for each subsequent radiotherapy session for these patients. By analyzing these predicted dosimetric assessment indicators, the timing of future offline adaptive radiotherapy interventions can be predicted.

[0059] To improve the accuracy of time series prediction networks, an algorithm similar to graph matching is proposed to predict dosimetric assessment indicators for patients after each radiotherapy session. For example... Figure 12As shown, a radiation therapy patient database is first constructed. In addition to the data mentioned above, it should also include data such as the patient's age, gender, height, weight, cancer type, stage, specific target area, and radiation therapy prescription (total dose and fractions) (and other clinically relevant data, not limited to these). In clinical radiation therapy, for each patient, the database is matched with the closest historical radiation therapy patient data. Based on the planned dosimetric assessment indicators after the first few radiation therapy sessions for each new patient, the historical radiation therapy patient data in the database is fitted and adjusted to more accurately predict the dosimetric assessment indicators after each subsequent radiation therapy session for the new patient. The initial time series prediction network is combined with a graph-based prediction algorithm to form a complete time series prediction model.

[0060] This intelligent offline adaptive radiotherapy technology based on time series prediction models is a novel offline adaptive radiotherapy solution in this field, and it has the following main advantages:

[0061] (1) The entire process input uses the actual images and dose data of the radiotherapy patient, such as dose data acquired by CBCT and EPD. The processing of the actual patient images is only to improve the quality of CBCT images. Other than that, no transformation is performed on the actual patient images and doses, thus avoiding the errors that occur in the conventional offline adaptive radiotherapy process to the greatest extent.

[0062] (2) Steps S2, S3 and S5 are implemented intelligently and automatically through the developed algorithm model. Only human evaluation of the final result is required, which greatly reduces the disadvantages of the conventional offline adaptive solution being cumbersome, time-consuming and labor-intensive.

[0063] (3) Good homogenization effect and easy clinical promotion. Most medical linear accelerators will collect the first five CBCT images. Through the time series prediction model developed in step S5, it is possible to predict the time node that offline adaptive radiotherapy needs to be performed in the future based on a small amount of CBCT data. The requirements for the amount of CBCT data and the hardware and software configuration of the department are low, and it can be applied in most radiotherapy departments and medical linear accelerators in a homogenized manner.

[0064] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. An intelligent offline adaptive radiotherapy method based on a time series prediction model, characterized in that, Includes the following steps: Step S1: Collect CBCT, EPD projection image data, radiotherapy plan, localization CT, and delineation data of tumor target area and surrounding normal tissues and organs on localization CT for radiotherapy patients. Step S2: Using the first neural network algorithm, the image quality of CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality. Step S3: Using the second neural network algorithm, with the improved image quality of the CBCT, the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, and EPD projection image data as input, the dosimetric evaluation indicators of the radiotherapy plan required in clinical practice are accurately calculated on the CBCT. Step S4: Determine whether the radiotherapy plan needs to be adjusted based on the calculated radiotherapy dosimetric evaluation index. If so, design a new plan or adjust the plan. The new radiotherapy plan will be implemented after verification. Step S5: Design a time series prediction model for dosimetric assessment indicators of radiotherapy plans. Use radiotherapy patient data from steps S1 to S4 to train the model and predict the time nodes when radiotherapy patients need to perform offline adaptive radiotherapy. The main structure of the initial time series prediction model is a time series prediction network composed of a convolutional network and a linear network. To improve the accuracy of the time series prediction network, an algorithm similar to graph matching is proposed to predict the dosimetric assessment indicators of patients after each radiotherapy. The initial time series prediction network is combined with the graph-based prediction algorithm to form a complete time series prediction model.

2. The intelligent offline adaptive radiotherapy method based on a time series prediction model according to claim 1, characterized in that, In step S2, the image quality of the CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality. Both of these can be output by the same neural network algorithm in multiple tasks.

3. The intelligent offline adaptive radiotherapy method based on a time series prediction model according to claim 1 or 2, characterized in that, The algorithm used in step S2 to delineate the tumor target area and surrounding normal tissues and organs on the CBCT after image quality improvement is a precise deformation registration and / or automatic delineation algorithm.

4. The intelligent offline adaptive radiotherapy method based on a time series prediction model according to claim 1, characterized in that, Step S4 is performed manually.

5. The intelligent offline adaptive radiotherapy method based on a time series prediction model according to claim 1, characterized in that, The input data in step S5 includes the patient's EPID, radiotherapy plan, CBCT, localization CT, localization CT delineation data, and radiotherapy plan dosimetric evaluation indicators.

6. A smart offline adaptive radiotherapy system based on a time-series prediction model, comprising the smart offline adaptive radiotherapy method based on a time-series prediction model according to any one of claims 1 to 5, characterized in that, It includes an input module, a first neural network module, a second neural network module, an adjustment module, and a prediction model module, wherein: The input module is used to collect CBCT and EPID projection image data, radiotherapy plans, localization CT, and delineation data of tumor target area and surrounding normal tissues and organs on localization CT for radiotherapy patients. The first neural network module uses the first neural network algorithm to improve the image quality of CBCT and delineate the tumor target area and surrounding normal tissues and organs on the CBCT with improved image quality. The second neural network module uses the second neural network algorithm to accurately calculate the radiotherapy planning dosimetric assessment indicators required in clinical practice on the CBCT, taking the improved image quality of the CBCT, the delineation of the tumor target area and surrounding normal tissues and organs on the CBCT, and EPD projection image data as input. The adjustment module determines whether the radiotherapy plan needs to be adjusted based on the calculated dosimetric evaluation indicators. If so, it designs a completely new plan or an adjusted plan. The new radiotherapy plan is implemented only after it has been validated. The prediction model module utilizes a third neural network algorithm to design a time series prediction model for dosimetric assessment indicators of radiotherapy plans. The model is trained using radiotherapy patient data from steps S1 to S4 to predict the time points when radiotherapy patients need to undergo offline adaptive radiotherapy. The initial version of the time series prediction model is primarily a time series prediction network composed of a convolutional network and a linear network. To improve the accuracy of the time series prediction network, an algorithm similar to graph matching is proposed to predict the dosimetric assessment indicators after each radiotherapy session. The initial version of the time series prediction network is combined with the graph-based prediction algorithm to form a complete time series prediction model.

7. The intelligent offline adaptive radiotherapy system based on a time series prediction model according to claim 6, characterized in that, In the first neural network module, the image quality of the CBCT is improved, and the tumor target area and surrounding normal tissues and organs are delineated on the CBCT with improved image quality. Both of these can be output by the same neural network algorithm in multiple tasks.

8. The intelligent offline adaptive radiotherapy system based on a time series prediction model according to claim 6 or 7, characterized in that, In the first neural network module, the algorithm used in step S2 to delineate the tumor target area and surrounding normal tissues and organs on the CBCT after image quality improvement is a precise deformation registration and / or automatic delineation algorithm.

9. The intelligent offline adaptive radiotherapy system based on a time series prediction model according to claim 6, characterized in that, The adjustment and verification plan in the adjustment module is done manually.

10. The intelligent offline adaptive radiotherapy system based on a time series prediction model according to claim 6, characterized in that, The input data for the prediction model module includes EPID, radiotherapy plan, CBCT, localization CT, localization CT delineation data, and radiotherapy plan dosimetric evaluation indicators.

Images