Radiation dose calculation method and device based on magnetic resonance image, equipment and storage medium

Abstract

The invention provides a radiation dose calculation method, device and equipment based on a magnetic resonance image and a storage medium, a deep residual network architecture inspired by U-Net is adopted, an effective end-to-end network is developed as a reliable dose calculation engine specially customized for MRgART, dose calculation is carried out only by using the MR image, and the dose calculation efficiency is improved. And the calculation of the MR exclusive radiotherapy dose is promoted. The deep learning method eliminates the demand for deformation registration, and directly integrates MRI data into dose calculation in the learning stage of the network, which marks a significant progress compared with the traditional technology which generally depends on CT images to perform dose calculation, thereby improving the efficiency and potential precision of the MRgART treatment plan process. In the field of images, a direct relation between a distance correction cone (DCC) flux diagram and dose distribution is established, a complex and time-consuming ray tracing step generally needed in a deep learning (DL) dose calculation algorithm is avoided, and the method is particularly beneficial to an online self-adaptive working process and is suitable for rapid dose calculation and verification.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation dose calculation algorithms for magnetic resonance imaging, and in particular to a radiation dose calculation method, device, equipment and storage medium based on magnetic resonance imaging. Background Art

[0002] The Magnetic Resonance Imaging (MRI)-guided Adaptive Radiation Therapy (MRgART) system integrates an MRI scanner with a linear accelerator (LINAC). This integration enables doctors to obtain high-resolution MRI images of tumors and surrounding normal tissues before and during treatment, thereby improving the accuracy and effectiveness of cancer treatment. The MRgART system makes real-time adaptive adjustments to the patient's daily treatment plan based on the latest patient anatomy to ensure dose coverage of the tumor target and minimize irradiation of endangered organs. The seamless execution of this online adaptive replanning process relies on fast and accurate dose calculation algorithms. Therefore, accurate dose calculation algorithms play a vital role in the treatment planning process to ensure patient safety and maintain quality assurance and control in radiotherapy.

[0003] Kernel-based and Monte Carlo (MC)-based dose calculation algorithms are currently the most widely used methods in clinical practice. Kernel-based algorithms have relatively fast dose calculation speeds, but cannot accurately consider the effects of magnetic fields and have limited accuracy. MC-based dose calculation algorithms can more accurately capture the effects in magnetic fields, but they take too long. Although the implementation of MC dose calculation engines based on graphics processing units (GPUs) has significantly accelerated the speed of MC simulations, it is still difficult to calculate MC doses with a low statistical uncertainty level of 1% within a few seconds. For the online MRgART replanning process, highly accurate and almost real-time dose calculations are crucial and in high demand. New types of dose calculation methods need to be developed for use in the clinical practice of radiotherapy planning.

[0004] In recent years, in order to reduce the time and improve the accuracy of radiotherapy dose calculation, a number of studies on dose calculation algorithms based on deep learning have been conducted. However, these dose calculation algorithms based on deep learning have a common feature, that is, they need to input density information containing patient CT images. Traditionally, CT images are required for dose calculation, and these CT-based dose calculation algorithms may not be directly applied to MR images in traditional treatment planning because of differences in image acquisition, tissue contrast, and conversion of electron density (ED) to CT values. For example, in the MRgART process using the Unity system, the process of aligning the daily MRI images used for online planning with the pre-treatment CT images is a key step. An algorithm is needed to address the deformation registration requirements of MR images, directly integrate MRI data into dose calculations during the learning phase of the network, and focus on the magnetic field environment to avoid complex beam tracking process calculations. Summary of the invention

[0005] The purpose of the present invention is to provide a method, device, equipment and storage medium for radiation dose calculation based on magnetic resonance imaging, which solves the key needs for accurate and fast dose calculation in the MRgART workflow without relying on CT images or time-consuming beam tracking processes.

[0006] In order to achieve the above object, the present invention provides a method for calculating radiation dose based on magnetic resonance imaging, comprising the following steps:

[0007] S1, obtaining patient data and preprocessing, obtaining DICOM-RT data of prostate cancer patients receiving MRgART treatment plan, and obtaining radiation field direction view dose slices, radiation field direction view MRI volume, and distance-corrected cone flux map after analysis to form an independent data set;

[0008] S2, network construction, establishing a residual network based on radiation dose calculation in the field direction, wherein the residual network is used to establish the relationship between the MRI volume in the field direction, the distance-corrected cone flux map, and the dose slice in the field direction;

[0009] S3, network training and evaluation, the distance-corrected cone flux map and the field direction MRI volume of a single beam are input into the trained model to obtain the predicted field direction dose of the beam; the predicted field direction dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

[0010] Furthermore, the parsing process of the DICOM-RT data also includes:

[0011] S101, obtaining 3D MRI volume from radiological images;

[0012] S102, deriving the 3D dose of each beam from the radiotherapy plan as the gold standard;

[0013] S103, re-slicing the 3D MRI volume and 3D beam dose of the cross-sectional view into a field direction view corresponding to each beam;

[0014] S104, obtaining the isocenter flux map of each beam from the radiotherapy plan, adjusting the flux map to the distance between the specified slice and the radiation source in the field direction based on the inverse square law, deriving a distance-corrected cone flux map, and adjusting the cone flux map to maintain the same resolution as the MRI volume in the field direction.

[0015] Furthermore, in step S2, the MRI slice in the direction of the beam and the distance-corrected flux map are used as the first input of the model, and the MRI volume in the direction of the beam and the distance-corrected cone flux map are used as the second input of the model, wherein the MRI in the direction of the beam and the distance-corrected cone flux map include all 10 layers of BEV MRI images or distance-corrected flux maps in front of the specified slice and behind it.

[0016] Furthermore, in step S2, the residual network includes two input ends, each of which is connected to a six-layer network structure. The first and second layers are convolutional layers, the kernel sizes of the convolutional layers are 7×7 and 3×3, and the step sizes are 1×1 and 2×2, respectively. In subsequent hierarchical stages, residual convolution blocks are used, followed by a configurable number of residual identity blocks.

[0017] Furthermore, the output end of the six-layer network structure on the input side is connected to the two convolutional layers in the final stage to produce a single-channel output. Convolution and deconvolution processes are performed in each convolutional layer with a kernel size of 3×3 or 1×1. Batch normalization is performed after convolution and deconvolution, and then the final result is obtained through processing with a rectified linear unit (ReLU) activation function.

[0018] Furthermore, during the network training process, a data enhancement strategy is used to address the overfitting problem. The strategy includes the Adam optimizer, dynamic learning rate, and early stopping technology. The network architecture adopts the Keras DL library and is implemented with TensorFlow as the backend.

[0019] Furthermore, during the evaluation of the network, the mean absolute error (MAE) was used to evaluate the accuracy of the prediction results. The mean absolute error (MAE) is the pixel-level dose difference between the patient beam dose predicted by the deep learning algorithm and the patient's actual dose.

[0020]

[0021] Among them, ROI is the region of interest for evaluating pixel-level dose differences, and i represents the voxel point in the ROI, including the whole body, target area, and organs at risk.

[0022] On the other hand, the present invention also provides a radiation dose calculation device based on magnetic resonance imaging, comprising:

[0023] The patient data acquisition and preprocessing module acquires DICOM-RT data of prostate cancer patients receiving MRgART treatment plans, and after parsing, obtains radiation field direction view dose slices, radiation field direction view MRI volume, and distance-corrected cone flux map to form an independent data set;

[0024] A network construction module is used to establish a residual network based on radiation dose calculation in the field direction, wherein the residual network is used to establish a relationship between the MRI volume in the field direction, the distance-corrected cone flux map, and the dose slice in the field direction;

[0025] The network training and evaluation module inputs the distance-corrected cone flux map and the field direction MRI volume of a single beam into the trained model to obtain the predicted field direction dose of the beam; the predicted field direction dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

[0026] On the other hand, the present invention also provides a device, characterized in that it includes a memory and a processor, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the operations performed by the above-mentioned method for calculating radiation dose based on magnetic resonance imaging.

[0027] On the other hand, the present invention also provides a computer-readable storage medium, characterized in that the storage medium stores at least one instruction, and the instruction is loaded and executed by a processor to implement the operations performed by the above-mentioned method for calculating radiation dose based on magnetic resonance imaging.

[0028] The present invention provides a method, device, equipment and storage medium for calculating radiation dose based on magnetic resonance imaging, which adopts a deep residual network architecture inspired by U-Net as a reliable dose calculation engine tailored for MRgART. It develops an effective end-to-end network that uses only MR images for dose calculation, facilitating MR-exclusive radiotherapy dose calculation. The deep learning method eliminates the need for deformation registration and directly integrates MRI data into dose calculation during the learning phase of the network, marking a significant improvement over traditional techniques that typically rely on CT images for dose calculation, thereby improving the efficiency and potential accuracy of the MRgART treatment planning process. In the field of images, the present invention establishes a direct relationship between the distance-corrected cone (DCC) flux map and the dose distribution, bypassing the complex and time-consuming ray tracing steps typically required in deep learning (DL) dose calculation algorithms. The method is particularly advantageous for online adaptive workflows and is suitable for rapid dose calculation and verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 The present invention is a flowchart of a method for calculating radiation dose based on magnetic resonance imaging according to an embodiment of the present invention.

[0031] Figure 2 The system framework diagram of a radiation dose calculation system based on magnetic resonance imaging according to an embodiment of the present invention.

[0032] Figure 3 This is a network architecture diagram of a residual network according to an embodiment of the present invention.

[0033] Figure 4 The dose of a patient calculated based on MC (first column), the dose calculated based on DL (second column), and the pixel-level dose difference (third column) according to an embodiment of the present invention are shown.

[0034] Figure 5 It is a box plot of MAE of the whole body, target area and OAR according to an embodiment of the present invention.

[0035] Figure 6 Box plot of the gamma analysis pass rates of the whole body, target area, and OAR according to an embodiment of the present invention.

[0036] Figure 7This is a comparison diagram of the DVH curve of a patient in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0039] If similar descriptions of "first / second" appear in the application documents, the following instructions are added. In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0041] The present invention provides a method, device, equipment and storage medium for calculating radiation dose based on magnetic resonance imaging, which uses only magnetic resonance images for radiation therapy dose calculation. The prior art still relies on CT images, and needs to input data such as patient density information from CT images, and the method cannot be directly used for magnetic resonance images. For example, in the ATS workflow of the Unity system, daily CT images are not available. The ATS workflow needs to optimize the plan through daily MRI and adaptive contours, and the average electron density (ED) on the preoperative CT is assigned to the corresponding contour on the daily MRI to complete the dose calculation, and deformable registration of the daily MRI planned online with the preoperative CT is a necessary step. The deep learning algorithm proposed in the present invention eliminates the need for deformable registration, and integrates the conversion from MRI to ED into the network learning process for dose calculation. By directly using MR images for dose calculation, the DL method simplifies the input data requirements and simplifies the treatment planning process. This not only reduces the calculation time, but also reduces the dependence on CT images, making the process more efficient and potentially more accurate. In addition, most deep learning (DL) dose calculation algorithms involve a complex and time-consuming ray tracing process, which is not suitable for online adaptive workflows that require fast dose calculation. Our DL method establishes a direct relationship between DCC irradiation maps and beam field view (BEV) dose in the image domain, eliminating the need for a complex ray tracing process.

[0042] A method, apparatus, device and storage medium for calculating radiation dose based on magnetic resonance imaging according to an embodiment of the present invention will be described below with reference to the accompanying drawings. First, a method for calculating radiation dose based on magnetic resonance imaging according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0043] Figure 1 FIG. 1 is a flow chart of a method for calculating radiation dose based on magnetic resonance imaging according to an embodiment of the present invention. Figure 1 As shown, the calculation method includes the following steps:

[0044] S1, obtain patient data and preprocess, obtain DICOM-RT data of prostate cancer patients receiving MRgART treatment plan, and obtain dose slices in the direction of the field, MRI volume in the direction of the field, and cone flux map corrected by distance after analysis to form an independent data set.

[0045] In a specific embodiment of the present invention, data from 30 patients with prostate cancer were used, who received fixed field intensity modulated radiation therapy (IMRT) using Elekta Unity (Elekta AB, Stockholm, Sweden), which is a 7MV non-flattened filter (FFF) radiotherapy system combined with a 1.5 Tesla (T) Philips (Philips Healthcare, Best, The Netherlands) MRI system, with a source axial distance (SAD) of 143.5 cm, a maximum dimension of 57.4 cm x 22.0 cm, and 160 multi-leaf collimator leaves in the cross-sectional direction with a leaf width of 0.7175 cm. The prostate cancer patients received a clinical target volume (CTV) of 40 Gy and a planning target volume (PTV) of 36.25 Gy divided into 5 treatments.

[0046] For each treatment fraction, this embodiment adopts a treatment plan adaptation strategy of "shape adaptation" (ATS) to obtain an online MRgART treatment plan. The ATS process optimizes the plan through daily MRI and adaptive contours, realizing plan adjustment based on the patient's new anatomical structure. Specifically, the initial step is to align the online planning MRI with the pre-treatment CT using deformable registration. Then, the pre-treatment contour is automatically projected onto the online planning MRI. When it is deemed necessary, the radiation oncologist can adjust the contour. According to the average electron density (ED) of the contour on the pre-treatment CT, the ED is assigned to the corresponding contour on the MRI. The online MRgART treatment plan is then generated by the "Optimize Weights and Shapes from Flux" adaptive planning method in the Monaco 5.4 (Elekta AB, Stockholm, Sweden) treatment planning system (TPS). For each treatment fraction, the gantry angle is set to 200, 240, 280, 320, 0, 40, 80, 120 and 160 degrees.

[0047] A 3D T2-weighted spin echo cross-sectional MRI sequence was used for data acquisition, with a flip angle of 90 degrees, a repetition time (TR) of 1535 ms, an echo time (TE) of 277.818 ms, a slice thickness of 2 mm, an in-plane resolution of 0.833 mm × 0.833 mm, and a field of view (FOV) of 400 mm × 400 mm.

[0048] The Digital Imaging and Communications (DICOM)-RT data for each treatment fraction, including RT images, RT plans, RT structures, and RT doses, are then transferred from Monaco TPS to the computational server. The DICOM-RT data is parsed as follows: I) A 3D MRI volume is acquired from the RT image and resampled to an in-plane resolution of 512×512 with a resolution of 1.435mm x1.435mm. II) The 3D dose of each beam is derived from the RT dose as the gold standard. The gold standard dose is calculated using the dose engine of Monaco5.4TPS with a statistical uncertainty of 1% and a grid spacing of 3mm 3 . III) The 3D MRI volume and 3D beam dose of the cross-sectional view are re-sliced ​​into BEV views aligned with each beam. The generated 3D beam direction view (BEV) MRI and 3D beam direction view (BEV) dose are both 512x 512x 121 in size, with a resolution grid of 1.435mm x1.435mm x 5mm, ensuring that the overall image and dose distribution of each patient are within the grid. IV) The isocenter flux map of each beam is obtained from the RT plan, and the distance correction cone (DCC) flux map is derived by adjusting the flux map to the BEV distance between the specified slice and the radiation source based on the inverse square law. This adjustment ensures that the distance correction cone (DCC) flux map maintains the same resolution as the 3D beam direction view (BEV) MRI. It was determined that the combination of the specified beam direction view (BEV) dose slice, beam direction view MRI volume, and distance correction cone (DCC) flux map constitutes an independent dataset, which is then used as the main input and output of the network. A total of 120 online treatment programs were included in this study.

[0049] Preferably, for the included patient data, four fractionated online MRgART treatment plans for each patient are randomly selected from the database to increase the data sample size and maintain a balance between anatomical universality and diversity. A total of 120 online treatment plans are included in this embodiment. The data are evenly distributed between training, validation and test sets according to the treatment fractions. After final allocation, the data set is divided into 76 training plans, 20 validation plans and 24 test plans. S2, network construction, establish a residual network based on the field direction radiation dose calculation, the residual network is used to establish the relationship between the field direction MRI volume, the distance-corrected cone flux map and the field direction dose slice.

[0050] Specifically, the network constructed in this embodiment uses a deep learning algorithm to establish the correlation between the beam field directional view (BEV) dose distribution and the corresponding distance correction cone (DCC) flux map in the magnetic field. Figure 3As shown in FIG. 1 , the model uses the specified beam direction view MRI slice and the distance correction flux map as input, and the beam direction view MRI volume and the distance correction cone (DCC) flux map as another input. It should be noted that the beam direction view MRI volume or the distance correction cone (DCC) flux map refers to the beam direction view MRI image or the distance correction (DCC) flux map in front of the specified slice, and the beam direction view MRI image or the distance correction flux map of 10 layers behind the specified slice.

[0051] The model then generates the corresponding beam field view (BEV) dose output. The architecture includes two six-layer hierarchical structures on the left side of the network, which gradually reduces the feature map size from 512×512 to the final 16×16 while increasing the number of filters from 16 to 512. In the first two hierarchical stages, two convolutional layers are used. The kernel sizes of the convolutional layers are 7×7 and 3×3, respectively, and the strides are 1×1 and 2×2, respectively.

[0052] In the subsequent hierarchical stages, residual convolution blocks are used, followed by a configurable number of residual identity blocks. The convolution blocks downsample the feature maps using a stride of 2×2, while the identity blocks help preserve important features. On the right side of the network, two input feature maps are combined and expanded to 512×512 through five hierarchical structures. In each hierarchy, residual deconvolution blocks with a stride of 2×2 are used, followed by a varying number of residual identity blocks to further enhance the features. Skip connections are used at each layer to preserve the reusability of features from the previous layer. This stabilizes training and increases the convergence rate of the model by directly connecting features of the same dimension from the previous layer to the current layer. In the final stage on the right, a single-channel output is produced through two convolutional layers. In each layer of the network, convolution and deconvolution processes are performed with kernel sizes of 3×3 or 1×1. Zero padding is used to maintain the size of features during convolution or deconvolution. Batch normalization is usually applied after convolution and deconvolution layers, followed by a rectified linear unit (ReLU) activation function. This method can introduce nonlinearity and enhance feature expression capabilities, which helps improve network performance and stability during training and testing.

[0053] S3, network training and evaluation, the distance-corrected cone flux map and the field direction MRI volume of a single beam are input into the trained model to obtain the predicted field direction dose of the beam; the predicted field direction dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

[0054] Specifically, during the training process, in order to improve the generalization ability of the model, a variety of data enhancement strategies were used to deal with the overfitting problem. For example, the Adam optimizer was used, and its initial learning rate was set to 10 -4, to minimize the mean square error loss function, the batch size is 8, and dynamic learning rate and early stopping techniques can also be used to solve the overfitting problem and improve convergence efficiency. In addition, if the loss value does not improve within 20 cycles, the training is terminated. The model with the best performance on the validation sample is selected for subsequent testing.

[0055] Specifically, the network architecture uses the Keras DL library and is implemented with TensorFlow as the backend. The training process is carried out on an NVIDIA 2080Ti GPU (12GB memory). The entire training process lasts about four weeks, and finally a trained model is obtained.

[0056] During the testing phase, Figure 3 As shown in the figure, the distance-corrected cone (DCC) flux map and the BEV MRI volume of a single beam are input into the trained model to obtain the predicted beam directional view (BEV) dose of the beam. The predicted beam directional view (BEV) dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted 3D patient dose.

[0057] Specifically, in order to evaluate the performance of the model, in this embodiment, the mean absolute error (MAE) is used to evaluate the accuracy of the prediction result. The mean absolute error (MAE) is the pixel-level dose difference between the patient beam dose predicted by the deep learning algorithm (DL) used in this method and the patient's actual dose (MC calculation), and the calculation process is:

[0058]

[0059] Where ROI is the region of interest for evaluating pixel-level dose differences, and i represents the voxel point in the ROI, including the whole body, target area, and organs at risk. Figure 4 As shown, the MC-based dose calculation results (first column), DL-based dose calculation results (second column), and pixel-level patient dose differences (third column) are superimposed on MRI images of prostate patients acquired from different angles. The dose calculated based on DL showed good consistency with the dose calculated based on MC. Table 1 shows the median, range, and interquartile range (IQR) of the dose difference MAE for the whole body, target area, and OAR for all tested cases.

[0060] Table 1:

[0061]

[0062] To further evaluate the robustness of the model, this method calculated the 3D gamma pass rate using two standards: 3% / 3mm and 3% / 2mm, as shown in Table 2 and Figure 6 As shown, the algorithm performs well.

[0063] Table 2

[0064]

[0065] Figure 7 The DVH curves calculated based on DL are shown, which are highly similar to those calculated by MC. This indicates that the DL method has achieved a clinically acceptable level of accuracy. Table 4 shows the mean and standard deviation of clinical dosimetric indicators for all test cases based on MC and DL calculated dose distributions. Statistical analysis showed that there were no significant differences between most dosimetric indicators, but there was a significant difference in DVH between CTV and PTV. 95 In terms of the values, there are significant differences between the DL and MC calculation results. The target dose distribution based on DL shows higher dose inhomogeneity, with a relatively higher maximum dose and a lower minimum dose. Overall, the two dose distributions calculated by MC and DL techniques are clinically equivalent.

[0066] Table 4. Clinical dosimetric parameters for all tested patients (mean ± SD). aVm represents the absolute volume received in m Gy (where m = 36, 37, 38, ...).

[0067]

[0068] In a specific embodiment, in order to facilitate the use of medical staff at work, the present invention uses a mobile device or a computer as an electronic device for running the method of the present invention, and the electronic device includes: a processor and a memory for storing executable instructions of the processor. Among them, the executable instructions stored in the memory are configured to run a visualization program of the radiation dose calculation method based on magnetic resonance imaging of the present invention, and the visualization program can communicate data with other terminals, servers or other forms of devices to complete the background program calculation.

[0069] Figure 2 FIG. 1 is a system framework diagram of a radiation dose calculation device based on magnetic resonance imaging according to an embodiment of the present invention. Figure 2 As shown, a radiation dose calculation device based on magnetic resonance imaging of the present invention comprises:

[0070] The patient data acquisition and preprocessing module 100 acquires DICOM-TR data of a prostate cancer patient who receives an MRgART treatment plan, and obtains radiation field direction view dose slices, radiation field direction view MRI volumes, and distance-corrected cone flux maps after parsing to form an independent data set;

[0071] A network construction module 200 is used to establish a residual network based on radiation dose calculation in the field direction, wherein the residual network is used to establish a relationship between the MRI volume in the field direction, the distance-corrected cone flux map, and the dose slice in the field direction;

[0072] The network training and evaluation module 300 inputs the distance-corrected cone flux map and the field direction view MRI volume of a single beam into the trained model to obtain the predicted field direction view dose of the beam; the predicted field direction view dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

[0073] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. A person of ordinary skill in the art can change, modify, replace and modify the above embodiments within the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for calculating radiation dose based on magnetic resonance imaging, characterized in that: The following steps are involved: S1, obtaining patient data and preprocessing, obtaining DICOM-RT data of prostate cancer patients receiving MRgART treatment plan, and obtaining radiation field direction view dose slices, radiation field direction view MRI volume, and distance-corrected cone flux map after analysis to form an independent data set; S2, network construction, establishing a residual network based on radiation dose calculation in the field direction, wherein the residual network is used to establish the relationship between the MRI volume in the field direction, the distance-corrected cone flux map, and the dose slice in the field direction; S3, network training and evaluation, the distance-corrected cone flux map and the field direction MRI volume of a single beam are input into the trained model to obtain the predicted field direction dose of the beam; the predicted field direction dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

2. A method for calculating radiation dose based on magnetic resonance imaging as claimed in claim 1, characterized in that: In step S1, the parsing process of the DICOM-RT data further includes: S101, obtaining 3D MRI volume from radiological images; S102, deriving the 3D dose of each beam from the radiotherapy plan as the gold standard; S103, re-slicing the 3D MRI volume and 3D beam dose of the cross-sectional view into a field direction view corresponding to each beam; S104, obtaining the isocenter flux map of each beam from the radiotherapy plan, adjusting the flux map to the distance between the specified slice and the radiation source in the field direction based on the inverse square law, deriving a distance-corrected cone flux map, and adjusting the cone flux map to maintain the same resolution as the MRI volume in the field direction.

3. A method for calculating radiation dose based on magnetic resonance imaging as claimed in claim 2, characterized in that: In step S2, the MRI slice in the direction of the beam and the distance-corrected flux map are used as the first input of the model, and the MRI volume in the direction of the beam and the distance-corrected cone flux map are used as the second input of the model, wherein the MRI in the direction of the beam and the distance-corrected cone flux map include all 10 layers of BEV MRI images or distance-corrected flux maps in front of the specified slice and behind it.

4. A method for calculating radiation dose based on magnetic resonance imaging as claimed in claim 3, characterized in that: In step S2, the residual network includes two input terminals, each of which is connected to a six-layer network structure. The first and second layers are convolutional layers, and the kernel sizes of the convolutional layers are 7×7 and 3×3, respectively, and the step sizes are 1×1 and 2×2, respectively. In the subsequent stage, a residual convolution block is used, followed by a configurable number of residual identity blocks.

5. The method for calculating radiation dose based on magnetic resonance imaging according to claim 4, characterized in that: The output end of the six-layer network structure on the input side is connected to the two convolutional layers in the final stage to produce a single-channel output. Convolution and deconvolution processes are performed in each convolutional layer with a kernel size of 3×3 or 1×1. Batch normalization is performed after convolution and deconvolution, and then the final result is obtained through processing with a rectified linear unit (ReLU) activation function.

6. A method for calculating radiation dose based on magnetic resonance imaging as claimed in claim 5, characterized in that: During the network training process, data enhancement strategies are used to deal with the overfitting problem. The strategies include Adam optimizer, dynamic learning rate, and early stopping technology. The network architecture uses the Keras DL library and is implemented with TensorFlow as the backend.

7. The method for calculating radiation dose based on magnetic resonance imaging according to claim 6, characterized in that: In the evaluation process of the network, the mean absolute error (MAE) is used to evaluate the accuracy of the prediction results. The mean absolute error (MAE) is the pixel-level dose difference between the patient beam dose predicted by deep learning of this method and the patient's actual dose. The calculation process is: Among them, ROI is the region of interest for evaluating pixel-level dose differences, and i represents the voxel point in the ROI, including the whole body, target area, and organs at risk.

8. A radiation dose calculation device based on magnetic resonance imaging, characterized in that: include: The patient data acquisition and preprocessing module acquires DICOM-RT data of prostate cancer patients receiving MRgART treatment plans, and after parsing, obtains radiation field direction view dose slices, radiation field direction view MRI volume, and distance-corrected cone flux map to form an independent data set; A network construction module is used to establish a residual network based on radiation dose calculation in the field direction, wherein the residual network is used to establish a relationship between the MRI volume in the field direction, the distance-corrected cone flux map, and the dose slice in the field direction; The network training and evaluation module inputs the distance-corrected cone flux map and the field direction MRI volume of a single beam into the trained model to obtain the predicted field direction dose of the beam; the predicted field direction dose is then re-sliced ​​to obtain the cross-sectional beam dose, which is then synthesized to obtain the final predicted patient beam dose.

9. A device, characterized in that: It comprises a memory and a processor, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the operation performed by the radiation dose calculation method based on magnetic resonance imaging as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, which is loaded and executed by the processor to implement the operations performed by the method for calculating radiation dose based on magnetic resonance imaging as described in any one of claims 1 to 7.

Images