Deep learning model for radiotherapy dose distribution prediction and radiotherapy plan generation method

Abstract

The invention discloses a deep learning model for radiotherapy dose distribution prediction and a radiotherapy plan generation method, and the deep learning model is constructed based on a UNet model, is obtained by inputting a training data set into the UNet model for training, and comprises a coding part which is composed of a mixing module and a down-sampling layer and is used for extracting data features, and the down-sampling layer uses maxpoo; the decoding part is composed of a decoding layer mixing module and an up-sampling layer and is used for gradually recovering the size of the original resolution, and the up-sampling layer adopts transposition convolution; the jump connection part is used for connecting the encoding part and the decoding part, and extracting features which are beneficial for the decoding part to recover to the original resolution from the encoding part; wherein the mixing module is composed of a plurality of branch paths, and outputs of the plurality of branch paths are added to serve as final output. The deep learning model is constructed by using a multi-branch path combination strategy, three-dimensional dose distribution can be quickly and accurately predicted, and plan generation in radiotherapy can be quickly completed.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy, and particularly to a deep learning model for radiotherapy dose distribution prediction, a radiotherapy plan generation method and a device. Background Art

[0002] Tumor radiation therapy is one of the three major means of tumor treatment. Its basic goal is to maximize the treatment gain ratio, that is, to ensure that the target area receives a certain coverage rate of the prescribed dose while protecting the surrounding normal tissues from unnecessary irradiation or reducing such irradiation. The radiotherapy dose is a very important parameter in radiotherapy treatment, which determines the intensity and duration of radiation. Therefore, solving the dose prediction problem in radiotherapy has a great impact on both the treatment effect and side effects.

[0003] Deep learning technology can learn a large amount of radiotherapy data, automatically extract features and build a model, so as to realize the prediction of radiotherapy dose. This method can not only improve the prediction accuracy, but also reduce the cost and time of manual intervention. At present, many studies have proved the effectiveness and feasibility of the radiotherapy dose prediction method based on deep learning, and it has been widely applied in clinical practice. For example, the existing Chinese invention patent application with the application number 202210701129.X and the name of "Automatic Contouring Method for Cervical Cancer Target Area and Automatic Generation Method for Adaptive Radiotherapy Plan" discloses an automatic contouring method for cervical cancer target area and an automatic generation method for adaptive radiotherapy plan, which includes: 1) image preprocessing; 2) target area segmentation of the input image based on the RefineNetPlus3D segmentation model, where the RefineNetPlus3D segmentation model includes: a left encoder, which is composed of a cascade of K-level residual networks and includes a plurality of convolutional blocks, and the resolution is successively reduced by downsampling between two adjacent convolutional blocks; a right decoder, which includes K cascaded decoding convolutional layers, and a 3DRefine Block is provided between the corresponding decoding convolutional layer and the convolutional block of the corresponding level, and the 3DRefine Block corresponding to the (K + 1)-th layer decoding convolutional layer is connected to the K-th layer decoding convolutional layer. The automatic segmentation and dose prediction model of this invention patent application can show a good automatic segmentation effect on the test data set while providing a reasonable dose distribution reference.

[0004] However, due to the design principle of the residual connection, the above-mentioned invention patent still has problems and disadvantages in that it cannot effectively reuse features.

[0005] In view of this, the present invention patent is specifically proposed. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a deep learning model for radiotherapy dose distribution prediction, a radiotherapy plan generation method and device, establishes a dose prediction model based on deep learning, and realizes fast and accurate radiotherapy plan generation. Specifically, the following technical solutions are adopted:

[0007] A deep learning model for radiotherapy dose distribution prediction, the deep learning model is constructed based on the UNet model and obtained by inputting a training data set into the UNet model for training. The deep learning model includes:

[0008] An encoding part, which is composed of a hybrid module and a downsampling layer, and is used to extract data features. The downsampling layer uses maxpool;

[0009] A decoding part, which is composed of a decoding layer hybrid module and an upsampling layer, and is used to gradually restore the original resolution size. The upsampling layer uses transposed convolution;

[0010] A skip connection part, which is used to connect the encoding part and the decoding part, and extract features from the encoding part that are helpful for the decoding part to restore to the original resolution;

[0011] Wherein, the hybrid module is composed of multiple branch paths, and the outputs of the multiple branch paths are added as the final output.

[0012] As an optional implementation manner of the present invention, the hybrid module includes:

[0013] A first branch path convolutional layer for dimension elevation;

[0014] A second branch path convolutional network, including a first convolutional layer for channel dimension elevation, a second convolutional layer for depthwise separation, and a third convolutional layer for channel dimension reduction and mixing of each channel's information. The number of convolutional kernels of the second convolutional layer is greater than that of the first convolutional layer and the third convolutional layer;

[0015] A third branch path convolutional layer, which is a DenseNet for feature reuse;

[0016] The first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer are arranged in parallel. The input data enters the first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer respectively for convolutional processing, and the output results of the first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer are added as the final output.

[0017] As an alternative embodiment of the present invention, the loss function of the deep learning model adopts the combination of the loss function for dose and the loss function for DVH; wherein, the loss function for dose measures the mean absolute error between the predicted dose and the actual dose of the pixel points within the contour, and no prediction is made outside the contour. The loss function for DVH includes the dose loss at the percentile position and the first derivative loss, and the loss function for DVH is

[0018] dvh loss(pred, gt) = L1(pred[sort indice], sort(gt)) + L1(pred′[sortindice], sort′(gt))

[0019] wherein, pred represents the predicted dose, gt represents the actual dose, L1 represents the mean absolute loss, and sort represents re - sorting the actual dose and using the original sub - scripts corresponding to its sorted elements for sorting the predicted dose.

[0020] As an alternative embodiment of the present invention, the training method of the deep learning model includes:

[0021] Collect the planned data of precise radiotherapy for tumors of historical patients, including CT data, organ delineation data, planned prescription data, and dose data;

[0022] Pre - process the collected planned data to make it suitable for reading by the UNet model;

[0023] Perform data augmentation on the pre - processed planned data to generate more data and obtain a training dataset;

[0024] Train the UNet model using the training dataset to obtain a deep learning model.

[0025] As an alternative embodiment of the present invention, the pre - processing of the collected planned data includes:

[0026] Perform resampling, matching, and cropping of the CT data, dose data, and organ delineation data to a preset size;

[0027] Perform scaling processing on the CT data: set the regions with HU values less than the first preset HU value to the first preset HU value, set the regions with HU values greater than the second preset HU value to the second preset HU value, and keep the regions with HU values between the first preset HU value and the second preset HU value as the original HU values, and output in the format for reading by the UNet model.

[0028] As an alternative embodiment of the present invention, the data augmentation of the pre - processed planned data includes:

[0029] Perform data augmentation operations on the preprocessed CT data, dose data, and organ delineation data synchronously;

[0030] The data augmentation operations include translation, flipping, and rotation. For translation, the maximum offset is set, and the data after translation is restored to the original data size by padding with 0. Flipping is randomly performed on the three axes, and rotation is randomly selected from fixed angle values.

[0031] The present invention also provides a radiotherapy plan generation method, including:

[0032] Obtain the user's CT data and organ delineation data, and input them into the deep learning model for predicting the radiotherapy dose distribution;

[0033] The deep learning model performs prediction and outputs the three-dimensional dose distribution.

[0034] As an optional implementation manner of the present invention, the radiotherapy plan generation method of the present invention includes calculating the DVH of each organ according to the three-dimensional dose distribution output by the deep learning model, and setting DVH constraints according to the DVH of each organ;

[0035] Use an optimization algorithm to optimize the three-dimensional dose distribution to obtain a radiotherapy plan.

[0036] The present invention also provides a radiotherapy plan generation device, including:

[0037] An acquisition module that acquires the user's CT data and organ delineation data;

[0038] A model prediction module that performs prediction on the acquired CT data and organ delineation data through the deep learning model for predicting the radiotherapy dose distribution according to any one of claims 1-6, and outputs the three-dimensional dose distribution.

[0039] As an optional implementation manner of the present invention, a radiotherapy plan generation device of the present invention includes:

[0040] A user data collection module that collects the planned data of precise radiotherapy for historical patients, including CT data, organ delineation data, planned prescription data, and dose data;

[0041] A preprocessing module that preprocesses the collected planned data to make it suitable for reading by the UNet model;

[0042] A data augmentation module that performs data augmentation processing on the preprocessed planned data to generate more data and obtain a training data set;

[0043] A model training module that trains the UNet model using the training data set to obtain a deep learning model.

[0044] Advantages of the present invention compared with the prior art:

[0045] The deep learning model for radiotherapy dose distribution prediction of the present invention is a hybrid UNet model constructed based on the UNet model. Hybrid modules are introduced into the encoding part and the decoding part respectively. The hybrid module has multiple branch paths. In this way, when using the deep learning model for radiotherapy dose distribution prediction, the multiple branch paths of the hybrid module can enhance the fitting ability of the convolution kernel, improve the expression ability of the network, perform different convolution operations on the input data to obtain different representations, and then add the convolution results of each branch path for output, realizing fast and accurate radiotherapy dose distribution prediction.

[0046] A radiotherapy plan generation method and device of the present invention input the user's CT data and organ delineation data into the deep learning model for radiotherapy dose distribution prediction. The deep learning model for radiotherapy dose distribution prediction is a hybrid UNet model constructed based on the UNet model. Hybrid modules are introduced into the encoding part and the decoding part respectively. The hybrid module has multiple branch paths. In this way, when using the deep learning model for radiotherapy dose distribution prediction, the multiple branch paths of the hybrid module can perform different convolution operations on the input data to obtain different representations, and then add the convolution results of each branch for output, realizing fast and accurate radiotherapy dose distribution prediction, and optimizing the radiotherapy dose distribution to generate a radiotherapy plan. Brief Description of the Drawings:

[0047] Figure 1 Principle framework diagram of the deep learning model for radiotherapy dose distribution prediction in an embodiment of the present invention;

[0048] Figure 2 Specific example diagram of the hybrid module in the deep learning model for radiotherapy dose distribution prediction in an embodiment of the present invention;

[0049] Figure 3 Modular schematic diagram of a radiotherapy plan generation device in an embodiment of the present invention;

[0050] Figure 4 Comparison of the dvh curves of the predicted dose and the actual dose of a radiotherapy plan generation method in an embodiment of the present invention. Detailed Embodiments

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0052] Accordingly, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents some embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0053] It should be noted that, without conflict, the embodiments in the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0054] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0055] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present invention is usually placed during use, or the orientation or positional relationship commonly understood by those skilled in the art. Such terms are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the present invention. In addition, the terms "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0056] See Figure 1 As shown, the deep learning model for predicting radiotherapy dose distribution in this embodiment is characterized in that the deep learning model is constructed based on the UNet model and obtained by inputting the training data set into the UNet model for training. The deep learning model includes:

[0057] An encoding part, composed of a hybrid module and a downsampling layer, for extracting data features, and the downsampling layer uses maxpool;

[0058] A decoding part, composed of a decoding layer hybrid module and an upsampling layer, for gradually restoring the original resolution size, and the upsampling layer uses transposed convolution;

[0059] A skip connection part, for connecting the encoding part and the decoding part, and extracting features from the encoding part that are helpful for the decoding part to restore to the original resolution;

[0060] Among them, the hybrid module is composed of multiple branch paths, and the outputs of the multiple branch paths are added as the final output.

[0061] The deep learning model for radiotherapy dose distribution prediction in this embodiment is a hybrid UNet model based on the UNet model. Hybrid modules are introduced into the encoding part and the decoding part respectively. The hybrid module has multiple branch paths. In this way, when using the deep learning model to predict the radiotherapy dose distribution, the multiple branch paths of the hybrid module can perform different convolutional operations on the input data to obtain different representations, and then add the convolutional results of each branch for output, realizing fast and accurate radiotherapy dose distribution prediction and generating a radiotherapy plan.

[0062] See Figure 2 As shown, in the deep learning model for radiotherapy dose distribution prediction in this embodiment, the hybrid module includes:

[0063] The first branch path convolutional layer L1, used to increase the dimension;

[0064] The second branch path convolutional network L2 includes a first convolutional layer for increasing the channel dimension, a second convolutional layer for depthwise separable convolution, and a third convolutional layer for reducing the channel dimension and mixing the information of each channel. The number of convolutional kernels of the second convolutional layer is greater than that of the first convolutional layer and the third convolutional layer;

[0065] The third branch path convolutional layer L3 is a DenseNet for feature reuse;

[0066] The first branch path convolutional layer L1, the second branch path convolutional network L2, and the third branch path convolutional layer L3 are arranged in parallel. The input data enters the first branch path convolutional layer L1, the second branch path convolutional network L2, and the third branch path convolutional layer L3 respectively for convolutional processing, and the output results of the first branch path convolutional layer L1, the second branch path convolutional network L2, and the third branch path convolutional layer L3 are added as the final output.

[0067] See Figure 2 As shown in the example in, the first branch path convolutional layer L1 on the leftmost side of this embodiment is a 1×1×1 convolutional layer, whose function is to increase the dimension. The middle second branch path convolutional network L2 is a grouped convolutional network with an inverted bottleneck layer. It starts with a 1×1×1 first convolutional layer to increase the channel dimension, followed by a 3×3×3 second convolutional layer, and finally a 1×1×1 third convolutional layer to reduce the channel dimension and mix the information of each channel. The rightmost is the DenseNet for feature reuse, and the outputs of the three branches are added as the final output. Further, the loss function of the deep learning model in this embodiment adopts the combination of the loss function for dose and the loss function for dvh.

[0068] Among them, the loss function for the dose measures the mean absolute error between the predicted dose and the actual dose of the pixels within the contour. No prediction is made outside the contour. Since the actual dose outside the contour is 0, there is no prediction, and at the same time, it can prevent the overall predicted dose from being too low due to excessive 0 doses.

[0069] The loss function for the DVH in this embodiment includes the dose loss at the percentile position and the first derivative loss. The loss function for the DVH is

[0070] dvh_loss(pred, gt) = L1(pred[sort_indice], sort(gt)) + L1(pred'[sort_indice], sort'(gt))

[0071] Among them, pred represents the predicted dose, gt represents the actual dose, L1 represents the mean absolute loss, and sort represents re - sorting the actual dose and using the original subscripts corresponding to its sorted elements to sort the predicted dose.

[0072] As an alternative implementation of this embodiment, the training method of the deep - learning model for radiotherapy dose distribution prediction in this embodiment includes:

[0073] Collect the planned data of precise radiotherapy for tumors of historical patients, including CT data, organ delineation data, planned prescription data, and dose data;

[0074] Pre - process the collected planned data to make it suitable for reading by the UNet model;

[0075] Perform data augmentation on the pre - processed planned data to generate more data and obtain a training dataset;

[0076] Use the training dataset to train the UNet model to obtain a deep - learning model.

[0077] The pre - processing of the collected planned data in this embodiment includes:

[0078] Resample, match, and crop the CT data, dose data, and organ delineation data to a preset size;

[0079] Perform scaling processing on the CT data: Set the regions with HU values less than the first preset HU value to the first preset HU value, set the regions with HU values greater than the second preset HU value to the second preset HU value, and keep the regions with HU values between the first preset HU value and the second preset HU value as the original HU values, and output in the format for reading by the UNet model.

[0080] The data augmentation processing of the pre - processed planned data in this embodiment includes:

[0081] Perform data augmentation operations on the preprocessed CT data, dose data, and organ delineation data synchronously;

[0082] The data augmentation operations include translation, flipping, and rotation. For translation, a maximum offset is set, and the data after translation is restored to the original data size by padding with 0. Flipping is randomly performed on the three axes, and rotation is randomly selecting a rotation angle from fixed angle values.

[0083] This embodiment also provides a radiotherapy plan generation method, including:

[0084] Obtain the user's CT data and organ delineation data, and input them into the deep learning model for radiotherapy dose distribution prediction;

[0085] The deep learning model makes predictions and outputs the three-dimensional dose distribution.

[0086] In the radiotherapy plan generation method of this embodiment, the user's CT data and organ delineation data are input into the deep learning model for radiotherapy dose distribution prediction. The deep learning model for radiotherapy dose distribution prediction is a hybrid UNet model based on the UNet model. The encoding part and the decoding part respectively introduce hybrid modules. The hybrid module has multiple branch paths. In this way, when using the deep learning model to predict the radiotherapy dose distribution, the multiple branch paths of the hybrid module can perform different convolution operations on the input data to obtain different representations, and then add the convolution results of each branch and output them to achieve fast and accurate radiotherapy dose distribution prediction and generate a radiotherapy plan.

[0087] Furthermore, in the radiotherapy plan generation method of this embodiment, the three-dimensional dose distribution predicted is guided for optimization to optimize the final plan and complete plan generation; specifically including: calculating the DVH of each organ for the three-dimensional dose distribution output by the deep learning model, setting DVH constraints according to the DVH of each organ; using an optimization algorithm to optimize the three-dimensional dose distribution to obtain a radiotherapy plan.

[0088] See Figure 3 As shown, this embodiment also provides a radiotherapy plan generation device, including:

[0089] An acquisition module that acquires the user's CT data and organ delineation data;

[0090] A model prediction module that makes predictions on the acquired CT data and organ delineation data through the deep learning model for radiotherapy dose distribution prediction and outputs the three-dimensional dose distribution.

[0091] A radiotherapy plan generation device according to this embodiment. The acquisition module inputs the acquired CT data and organ delineation data of the user into the deep learning model of the model prediction module. The deep learning model for radiotherapy dose distribution prediction is a hybrid UNet model constructed based on the UNet model. Hybrid modules are introduced into the encoding part and the decoding part respectively. The hybrid module has multiple branch paths. In this way, when using the deep learning model to predict the radiotherapy dose distribution, the multiple branch paths of the hybrid module can perform different convolution operations on the input data to obtain different representations, and then add the convolution results of each branch and output them, realizing fast and accurate radiotherapy dose distribution prediction and generating a radiotherapy plan.

[0092] Further, a radiotherapy plan generation device according to this embodiment includes:

[0093] A user data collection module that collects the planned data for the precise radiotherapy of tumors of historical patients, including CT data, organ delineation data, planned prescription data, and dose data;

[0094] A preprocessing module that preprocesses the collected planned data to make it suitable for reading by the UNet model;

[0095] A data augmentation module that performs data augmentation processing on the preprocessed planned data to generate more data and obtain a training data set;

[0096] A model training module that trains the UNet model using the training data set to obtain a deep learning model.

[0097] A radiotherapy plan generation device according to this embodiment realizes generating a training data set for training a deep learning model according to the planned data for the precise radiotherapy of tumors of historical patients through the user data collection module, the preprocessing module, and the data augmentation module, and obtains the deep learning model of this embodiment through training by the model training module.

[0098] A radiotherapy plan generation device according to this embodiment further includes a model optimization module that guides and optimizes the predicted three-dimensional dose distribution to optimize the final plan and complete the radiotherapy plan generation.

[0099] This embodiment provides a specific example of a radiotherapy plan generation method. The specific implementation steps include:

[0100] 1) Collect the planned data for the precise radiotherapy of tumors of historical patients, including CT data, organ delineation data, planned prescription data, and dose data;

[0101] 2) First, resample, match, and crop the CT data, organ contour data, and dose data to 128×128×128. Then, perform scaling on the CT data: set all HU values less than -1024 to -1024, set all HU values greater than 1500 to 1500, and keep the rest as the original values. Finally, output in the.npy format for later model reading.

[0102] 3) Perform data augmentation on the training data, and synchronously perform data augmentation operations on the CT data, organ contour data, and dose data. The specific data augmentation operations include translation, flipping, and rotation. Among them, the maximum offset for translation is set to 20, and the data after translation is restored to the original data size by padding 0. Flipping randomly flips along the three axes. The rotation operation introduces a lot of interpolation and causes deviation in the final prediction, so the rotation angle is randomly selected from fixed angle values.

[0103] 4) Build a deep learning model based on the UNet model. This model includes:

[0104] 4.1. The encoding part is composed of hybrid modules (as shown in Figure 2 ) and downsampling layers, including five hybrid modules and four downsampling layers, which are used to extract data features. The hybrid model consists of multi-branched paths. The leftmost is a 1×1×1 convolutional layer, whose function is to increase the dimension. In the middle is a grouped convolutional network with an inverted bottleneck layer. At the beginning, a 1×1×1 convolution is used to increase the channel dimension, then a 3×3×3 depthwise separable convolutional layer, and finally a 1×1×1 convolutional layer is used to reduce the channel dimension and mix the information of each channel. The rightmost is DenseNet, whose function is to reuse features. The outputs of the three branches are added as the final output. The downsampling layer uses maxpool. Each time of downsampling, the number of channels doubles and the feature map is halved, that is, the number of channels: 16→32→64→128→256, and the feature map: 128×128×128→64×64×64→32×32×32→16×16×16→8×8×8.

[0105] 4.2. The decoding part is composed of hybrid modules and upsampling layers, including four hybrid modules and four upsampling layers, which are used to gradually restore the original resolution size. The upsampling layer uses transposed convolution, so that the number of channels: 256→128→64→

[0106] 32→16, and the feature map: 8×8×8→16×16×16→32×32×32→64×64×64→128×128×128.

[0107] 4.3. Skip connection part, which is a bridge connecting the encoding part and the decoding part, consists of a gated mechanism with attention. Its function is to extract the features from the encoding part that are most helpful for the decoding part to restore to the original resolution.

[0108] 5) Design the loss function. The loss function adopts a combination of two loss functions, namely the loss for dose and the loss for DVH. Among them, the loss function for dose measures the mean absolute error between the predicted dose and the actual dose of the pixels within the contour. The actual dose outside the contour is 0, so it is not predicted, which can also prevent the overall predicted dose from being too low due to excessive 0 doses. The loss function for DVH is as follows, including the dose loss at the percentile position and the first derivative loss.

[0109] dvh loss(pred, gt) = L1(pred[sort indice], sort(gt)) + L1(pred′[sort indice], sort′(gt))

[0110] 6) Train the model. The optimizer uses the Adam optimizer, the learning rate is set to 3e-4, the weight decay is 0, β1 and β2 are 0.9 and 0.999 respectively, and the learning rate scheduling algorithm uses OneCycleLR with learning rate warm-up and learning rate decay. The model is trained for 80 epochs, and the model parameters at the time when the loss on the validation set reaches the minimum are saved. As Figure 4 shown, the prediction result of a prostate patient. The dotted line represents the DVH calculated from the predicted dose, and the solid line represents the DVH calculated from the actual dose. The average absolute error of the prediction is about 2% of the prescription dose.

[0111] This embodiment also provides a computer-readable storage medium storing a computer-executable program, which, when executed, implements the radiotherapy plan generation method as described above.

[0112] The computer-readable storage medium in this embodiment may include a data signal propagated in the baseband or as part of a carrier wave, which carries the readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable storage medium can also be any readable medium other than the readable storage medium, and this computer-readable medium can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium can be transmitted by any appropriate medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination of the above.

[0113] This embodiment also provides an electronic device, including a processor and a memory. The memory is used to store computer-executable programs. When the computer program is executed by the processor, the processor executes the radiotherapy plan generation method.

[0114] The electronic device is presented in the form of a general computing device. The processor can be one or multiple and work collaboratively. The present invention does not exclude distributed processing, that is, the processors can be dispersed in different physical devices. The electronic device of the present invention is not limited to a single entity, but can also be the sum of multiple physical devices.

[0115] The memory stores computer-executable programs, usually machine-readable codes. The computer-readable program can be executed by the processor so that the electronic device can execute the method of the present invention or at least some of the steps in the method.

[0116] The memory includes volatile memory, such as a random access storage unit (RAM) and / or a cache storage unit, and can also be non-volatile memory, such as a read-only storage unit (ROM).

[0117] It should be understood that the electronic device of the present invention may also include elements or components not shown in the above examples. For example, some electronic devices also include a display unit such as a display screen, and some electronic devices also include human-computer interaction elements, such as buttons, keyboards, etc. As long as the electronic device can execute the computer-readable program in the memory to implement the method of the present invention or at least some of the steps of the method, it can be considered as the electronic device covered by the present invention.

[0118] From the above description of the embodiments, those skilled in the art can easily understand that the present invention can be implemented by hardware capable of executing specific computer programs, such as the system of the present invention, and the electronic processing units, servers, clients, mobile phones, control units, processors, etc. included in the system. The present invention can also be implemented by computer software that executes the method of the present invention, such as control software executed by a microprocessor, an electronic control unit, a client, a server, etc. However, it should be noted that the computer software that executes the method of the present invention is not limited to being executed in one or specific hardware entities. It can also be implemented in a distributed manner by unspecified specific hardware. For computer software, the software product can be stored in a computer-readable storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), or can be distributed and stored on the network, as long as it can enable the electronic device to execute the method according to the present invention.

[0119] The above embodiments are only used to illustrate the present invention and do not limit the technical solutions described in the present invention. Although the present specification has described the present invention in detail with reference to the above respective embodiments, the present invention is not limited to the above specific embodiments. Therefore, any modification or equivalent replacement of the present invention; and all technical solutions and their improvements that do not depart from the spirit and scope of the invention are covered by the scope of the claims of the present invention.

Claims

1. A deep learning model for predicting radiotherapy dose distribution, characterized in that, the deep learning model is constructed based on the UNet model and obtained by inputting a training data set into the UNet model for training. The deep learning model includes: An encoding part, composed of a hybrid module and a downsampling layer, is used to extract data features. The downsampling layer uses maxpool; A decoding part, composed of a decoding layer hybrid module and an upsampling layer, is used to gradually restore the original resolution size. The upsampling layer uses transposed convolution; A skip connection part, used to connect the encoding part and the decoding part, extracts features from the encoding part that help the decoding part restore to the original resolution; Among them, the hybrid module is composed of multiple branch paths, and the outputs of the multiple branch paths are added as the final output.

2. The deep learning model for predicting radiotherapy dose distribution according to claim 1, characterized in that, the hybrid module includes: A first branch path convolutional layer for dimension elevation; A second branch path convolutional network, including a first convolutional layer for channel dimension elevation, a second convolutional layer for depthwise separation, and a third convolutional layer for reducing channel dimension and mixing channel information. The number of convolutional kernels of the second convolutional layer is greater than that of the first convolutional layer and the third convolutional layer; A third branch path convolutional layer, which is a DenseNet for feature reuse; The first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer are arranged in parallel. Input data enters the first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer respectively for convolutional processing, and the output results of the first branch path convolutional layer, the second branch path convolutional network, and the third branch path convolutional layer are added as the final output.

3. The deep learning model for predicting radiotherapy dose distribution according to claim 1 or 2, characterized in that, the loss function of the deep learning model adopts the combination of a loss function for dose and a loss function for dvh; among them, the loss function for dose measures the mean absolute error between the predicted dose and the actual dose of the pixel points within the contour, and the loss function for dvh includes the dose loss at the percentile position and the first derivative loss. The loss function for dvh is dvh loss(pred, gt) = L1(pred[sort indice], sort(gt)) + L1(pred'[sort indice], sort'(gt)) where pred represents the predicted dose, gt represents the actual dose, L1 represents the mean absolute loss, and sort represents reordering the actual dose and using the original subscripts corresponding to its sorted elements to sort the predicted dose.

4. The deep learning model for predicting radiotherapy dose distribution according to claim 1, characterized in that, the training method of the deep learning model includes: Collecting the planned data of precise radiotherapy for tumors of historical patients, including CT data, organ delineation data, planned prescription data, and dose data; Preprocess the collected planning data to make it suitable for reading by the UNet model; Perform data augmentation on the preprocessed planning data to generate more data and obtain a training dataset; Train the UNet model using the training dataset to obtain a deep learning model.

5. The deep learning model for radiotherapy dose distribution prediction according to claim 4, wherein, The preprocessing of the collected planning data includes: Resample, match, and crop CT data, dose data, and organ delineation data to a preset size; Perform scaling processing on CT data: Set the regions with HU values less than the first preset HU value to the first preset HU value, set the regions with HU values greater than the second preset HU value to the second preset HU value, and keep the regions with HU values between the first preset HU value and the second preset HU value as the original HU values, and output in a format suitable for reading by the UNet model.

6. The deep learning model for radiotherapy dose distribution prediction according to claim 4, wherein, The data augmentation processing of the preprocessed planning data includes: Perform data augmentation operations on the preprocessed CT data, dose data, and organ delineation data synchronously; The data augmentation operations include translation, flipping, and rotation. Among them, the maximum offset is set for translation, and the original data size is restored by filling 0 for the translated data. Flipping is randomly performed on three axes, and the rotation angle is randomly selected from fixed angle values.

7. A radiotherapy plan generation method, wherein, includes: Obtain the user's CT data and organ delineation data, and input them into the deep learning model for radiotherapy dose distribution prediction according to any one of claims 1-6; The deep learning model makes a prediction and outputs a three-dimensional dose distribution.

8. The radiotherapy plan generation method according to claim 7, wherein, includes calculating the DVH of each organ based on the three-dimensional dose distribution output by the deep learning model, and setting DVH constraints according to the DVH of each organ; Use an optimization algorithm to optimize the three-dimensional dose distribution to obtain a radiotherapy plan.

9. A radiotherapy plan generation device, wherein, includes: An acquisition module that acquires the user's CT data and organ delineation data; A model prediction module that makes a prediction on the acquired CT data and organ delineation data through the deep learning model for radiotherapy dose distribution prediction according to any one of claims 1-6, and outputs a three-dimensional dose distribution.

10. The radiotherapy plan generation device according to claim 9, wherein, includes: A user data collection module that collects the planning data of precise radiotherapy for historical patients, including CT data, organ delineation data, planned prescription data, and dose data; A preprocessing module that preprocesses the collected planning data to make it suitable for reading by the UNet model; A data augmentation module that performs data augmentation processing on the preprocessed planning data to generate more data and obtain a training dataset; A model training module that trains the UNet model using the training dataset to obtain a deep learning model.

Images