Radiotherapy dose prediction method, system and equipment based on cue word guidance and medium

Abstract

The invention discloses a cue word guidance-based radiotherapy dose prediction method, system and device, and a medium, belongs to the technical field of artificial intelligence medicine, and aims to solve the technical problem of low radiotherapy dose prediction accuracy caused by single and fixed radiation field angle in the prior art. The method comprises the steps that a radiotherapy dose prediction model is constructed, the radiotherapy dose prediction model comprises a U-Net network and a text cue word encoder, a CT image, a target area image and an organ-at-risk image are input into the U-Net network for feature extraction, and a radiation field angle is input into the text cue word encoder; fusing the output of the down-sampling in the U-Net network and the output of the text cue word encoder, and then carrying out convolution and up-sampling; the trained radiotherapy dose prediction model is used for radiotherapy dose prediction. A text cue word encoder is added in a radiotherapy dose prediction model, discretized radiation field angle information is embedded into a high-dimensional feature space through the text cue word encoder, and image data and the angle information are fused, so that the dose prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medical technology and relates to a radiotherapy dose prediction method, and in particular to a radiotherapy dose prediction method, system, device and medium based on prompt word guidance. Background Art

[0002] Radiation therapy (RT), a core approach to cancer treatment, uses high-energy radiation to precisely damage the DNA structure of tumor cells. While killing cancer cells, it also maximizes the protection of surrounding normal tissues through advanced techniques such as intensity-modulated radiation therapy (IMRT) and stereotactic body body therapy (SBRT). Approximately 60%–70% of patients with malignant tumors require RT at various stages of treatment.

[0003] During radiotherapy, it is necessary to adjust the intensity of the radiation within the irradiation field in various directions to create a non-uniform irradiation field, achieving a high-dose three-dimensional conformal distribution for the tumor and a low-dose distribution for organs at risk. This increases the irradiation dose to the tumor, minimizes the dose to organs at risk and normal tissue, and ultimately improves the local tumor control rate. Therefore, dose prediction is a key technology in radiotherapy.

[0004] Traditionally, radiotherapy plans have been manually developed based on imaging data from CT and MRI. This process takes 4-8 hours, which is not only time-consuming and labor-intensive, but also very inefficient. However, with the development of artificial intelligence technology in recent years, deep learning models have been increasingly used to simulate radiotherapy dose distribution. While ensuring that the tumor target receives a radical dose, the exposure to organs at risk is precisely controlled within a safe threshold, effectively reducing the incidence of complications such as radiation-induced xerostomia. Compared to traditional manual radiotherapy planning, the use of deep models for radiotherapy dose prediction can compress this process to approximately 10 minutes, significantly improving work efficiency and providing timely support for clinical decision-making.

[0005] Using deep learning models to simulate / predict radiotherapy doses is a core component of precision radiotherapy planning. This approach primarily uses a deep learning model to generate a three-dimensional dose distribution map using voxel-level regression, taking the patient's CT images, tumor target volume, and organ-at-risk (OAR) contours as input. This technology has been widely applied to a variety of common tumor types, including head and neck cancer, lung cancer, esophageal cancer, rectal cancer, and cervical cancer. For example, Wang et al. proposed an error correction network (ECN) for head and neck gliomas. By dynamically adjusting the target boundary distance map, it significantly improved the spinal cord dose compliance rate to 98%. Zhang et al. designed a variant model based on a 3D U-Net architecture, achieving high-precision prediction of millimeter-level dose gradients in gamma knife therapy. Chen et al. introduced a graph convolutional network (GCN) combined with a self-attention mechanism to construct a hybrid model to capture the dose distribution characteristics of the complex anatomical structure of esophageal cancer. Furthermore, the introduction of generative adversarial networks (GANs) has further advanced the development of dose prediction technology. For example, Zhan et al. constructed an adversarial architecture with an improved U-Net as the generator and a convolutional neural network as the discriminator, and controlled the gradient error of dose prediction in rectal cancer to within 3%.

[0006] The invention patent application with application number 202011489700.3 discloses a method for generating a dose prediction model for nasopharyngeal carcinoma, a dose prediction method and a device, which include: obtaining the disease type information of the nasopharyngeal carcinoma patient based on the acquired medical images of the nasopharyngeal carcinoma patient; outlining the medical images to obtain the geometric anatomical structure of the nasopharyngeal carcinoma patient; obtaining prescription information based on the disease type information and the geometric anatomical structure; determining the number of fields of the radiotherapy plan and the field angle of each field, and the number of fields is within a preset range, and the preset range is [7, 12]; inputting the number of fields, the field angle of each field, the disease type information, the prescription information and the geometric anatomical structure into the dose prediction model generated by the above-mentioned method for generating a dose prediction model for nasopharyngeal carcinoma to obtain a dose distribution map of the radiotherapy plan.

[0007] Numerous advances have been made in radiotherapy dose prediction methods, such as using error correction networks to dynamically optimize target boundaries, combining graph convolution with self-attention mechanisms, and constructing hybrid models to characterize dose distributions in the complex structure of esophageal cancer. However, most methods still focus on a single disease type and a fixed perspective, lacking comprehensive consideration of multiple angles and disease types, and thus failing to meet physicians' diverse needs for radiotherapy planning. Furthermore, finding more accurate dose predictions remains a persistent goal for researchers. Summary of the Invention

[0008] The purpose of the present invention is to provide a radiotherapy dose prediction method, system, device and medium based on prompt word guidance in order to solve the technical problem of low accuracy of radiotherapy dose prediction in the prior art due to the single and fixed field angle.

[0009] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions: A radiotherapy dose prediction method based on prompt word guidance includes the following steps: Step S1, obtaining sample data; Acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples, and use the radiotherapy dose samples as label data; Step S2, constructing a radiotherapy dose prediction model; A radiotherapy dose prediction model was constructed, which included a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk were input into the U-Net network for feature extraction. The radiation field angle was input into the text prompt word encoder. The features output by the last downsampling in the U-Net network were convolved and fused with the features output by the text prompt word encoder. The fused features were then convolved and upsampled. Step S3, training a radiotherapy dose prediction model; The radiotherapy dose prediction model constructed in step S2 is trained using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and label data in step S1; Step S4, radiotherapy dose prediction; The radiotherapy CT image to be tested is obtained, and the target area and organs at risk of the radiotherapy CT image are outlined to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose at each field angle.

[0010] Furthermore, in step S2, the U-Net network includes a wavelet convolution layer, a downsampling module, an upsampling module and a 3*3 convolution layer; The input of the U-Net network is first subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the first feature; the first feature is downsampled for the first time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the second feature; the second feature is downsampled for the second time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the third feature; the third feature is downsampled for the third time by the downsampling module, and the output feature is fused with the feature output by the text prompt word encoder through the first layer of wavelet convolution layer, and the fused feature is upsampled by the first layer of wavelet convolution layer; the feature output after the first upsampling by the upsampling module is skip-linked with the third feature and then convolved through two layers of 3*3 convolution layers to obtain the fourth feature; the fourth feature is skip-linked with the second feature and then convolved through two layers of 3*3 convolution layers to obtain the fifth feature; the fifth feature is skip-linked with the first feature and then convolved through two layers of 3*3 convolution layers to obtain the sixth feature.

[0011] Furthermore, in step S2, when wavelet convolution is performed using the wavelet convolution layer, wavelet decomposition and wavelet reconstruction are included; The expression of wavelet decomposition is: ; The expression of wavelet reconstruction is: ; in, represents the input image, represents the wavelet filter, represents the convolution operation, represents the downsampling operation, represents transposed convolution, represents the reconstruction filter, Represents the feature map of each sub-band.

[0012] Furthermore, in step S2, the text prompt word encoder includes an Embedding module and three fully connected layers arranged in sequence; The field angle is mapped to a d-dimensional feature vector through the Embedding module. The d-dimensional feature vector is sequentially passed through three fully connected layers to obtain a higher-order angle feature. This angle feature is fused with the feature output after the last downsampling and convolution in the U-Net network.

[0013] Furthermore, when the Embedding module performs mapping, the calculation formula is: ; The fully connected layer extracts high-order angle features through nonlinear activation functions to obtain angle features , The calculation formula is: ; in, is the field angle, represents the embedding layer, represents the dimension of the embedding vector, Represents the feature vector obtained by mapping; 、 、 Both represent trainable weight matrices, Represents the ReLU activation function. Finally, a 1024-dimensional angle feature is obtained , used for subsequent feature fusion.

[0014] Furthermore, in step S2, when the features output by the U-Net network after the last downsampling and convolution are fused with the features output by the text prompt word encoder, the specific steps are as follows: Step S2-1: Use the KQV attention mechanism for interaction, specifically: The feature output after the last downsampling and convolution in the U-Net network is taken as X, the feature output by the text prompt word encoder is taken as Z, and the feature Z is used as the key and value in the calculation; the specific calculation method is: ; ; ; ; in, 、 、 Both represent trainable parameters, Indicates the scaling factor of the Key dimension; 、 、 Represents the query vector (query representation after image feature conversion), the key vector (key representation after angle feature conversion), and the value vector (value representation after angle feature conversion). represents transpose; Step S2-2 uses a multi-head attention mechanism to independently calculate attention in multiple subspaces and concatenate the final results to obtain the fused features; specifically:

[0015] in, represents the h-th attention head, Represents the output projection matrix (trainable parameter).

[0016] Furthermore, in step S3, when training the radiotherapy dose prediction model, the Adam optimizer is used, the initial learning rate is set to 1e-4, the batch size is considered to be 16, and the total number of training rounds is set to 250; when the loss does not decrease significantly within several consecutive rounds, the learning rate is automatically reduced, and the minimum learning rate is set to 1e-6; During training, the loss function is: ; in, Indicates the true value of radiotherapy dose, represents the expected value of radiotherapy dose, Represents the total number of voxels.

[0017] A radiotherapy dose prediction system based on prompt word guidance, comprising: A sample data acquisition module is used to acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples. The radiotherapy dose samples are used as label data. The radiotherapy dose prediction model construction module is used to build a radiotherapy dose prediction model. The radiotherapy dose prediction model includes a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk are input into the U-Net network for feature extraction. The radiation field angle is input into the text prompt word encoder. The features output by the last downsampling in the U-Net network are convolved and fused with the features output by the text prompt word encoder. The fused features are then convolved and upsampled. a radiotherapy dose prediction model training module, for training the radiotherapy dose prediction model constructed by the radiotherapy dose prediction model construction module using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and label data in the sample data acquisition module; The radiotherapy dose prediction module is used to obtain the radiotherapy CT image to be tested, and to outline the target area and organs at risk of the radiotherapy CT image to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose under each field angle.

[0018] The beneficial effects of the present invention are as follows: In the present invention, a text prompt word encoder is added to the radiotherapy dose prediction model. The discretized field angle information is embedded into the high-dimensional feature space through the text prompt word encoder, and the image data and angle information are fused to improve the accuracy of dose prediction. The U-Net network is used as the backbone network. It extracts the image features of CT, planned target area and organs at risk in the encoding stage, and combines the field angle information with the image features at the high-level features, so that the image of the field angle on the dose distribution can be more comprehensively modeled. By discretizing the seven field angles and encoding them into prompt word vectors, the model is guided to pay attention to the correspondence between the angle information and the image data, thereby realizing the dose prediction capability of any angle. This mechanism not only enhances the angle generalization capability of the model, but also improves the flexibility of the clinical field deployment plan, providing a more efficient and controllable dose calculation tool for individualized precision treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow diagram of the present invention; Figure 2 It is a schematic diagram of the structure of the radiotherapy dose prediction model in the present invention; Figure 3 Schematic diagram of the structure of the text prompt word encoder in the present invention; Figure 4 It is the Loss curve of the traditional U-Net network; Figure 5 is the Loss curve of the present invention; Figure 6 It is a comparison diagram of the difference in prediction effects of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, 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 in conjunction with the drawings 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.

[0021] Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides a radiotherapy dose prediction method based on prompt word guidance, which is used to predict the radiotherapy dose of CT images of patients with rectal cancer. Figure 1 As shown, the specific steps include: Step S1, obtaining sample data; Acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples, and use the radiotherapy dose samples as label data.

[0023] The radiotherapy sample data came from the radiotherapy planning data of 90 anonymous rectal cancer patients admitted to West China Hospital of Sichuan University. These data included each patient's CT image (3 mm slice thickness, 512 × 512 pixels), the structural outline files (DICOM RT Structure format) of the planned target volume image and organ-at-risk (bladder, small intestine) images, as well as each radiation field angle (0°, 52°, 104°, 156°, 204°, 256°, 308°) and the corresponding radiation dose.

[0024] The above-mentioned CT images are obtained by the physicist using a CT machine to scan the patient's tumor area. One scan will generate multiple continuous CT images. The target area image and the organ at risk image are obtained by the physicist outlining the target area and organs at risk based on the pathological characteristics of the case. The radiotherapy dose is the dose data obtained by the physicist simulating the radiotherapy process of the prepared radiotherapy plan.

[0025] Standardized data preprocessing methods are used for preprocessing. All data are normalized, resized, and standardized to ensure the consistency of model input and output and training stability.

[0026] The dataset formed by all samples was divided into training set, validation set and test set in a ratio of 8:1:1. The training set was used for model parameter optimization, the validation set was used to evaluate and adjust hyperparameters, and the test set was used for the final model evaluation to verify its performance in subsequent actual clinical applications.

[0027] Step S2, constructing a radiotherapy dose prediction model; A radiotherapy dose prediction model was constructed. The radiotherapy dose prediction model includes a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk were input into the U-Net network for feature extraction. The radiation field angle was input into the text prompt word encoder. The features output by the last downsampling in the U-Net network were convolved and fused with the features output by the text prompt word encoder. The fused features were convolved and upsampled, as shown in the following example. Figure 2 shown.

[0028] Wavelet convolution is introduced to improve the structure of the U-Net network. Wavelet convolution has the ability to decompose the frequency domain at multiple scales and can extract finer dose gradients and structural detail features in the joint space-frequency domain. It is especially suitable for modeling complex intensity variation areas in the modulated dose map. Figure 2As shown in the figure, the U-Net network includes a wavelet convolution layer, a downsampling module, an upsampling module and a 3*3 convolution layer; the input of the U-Net network is first subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the first feature; the first feature is downsampled for the first time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the second feature; the second feature is downsampled for the second time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the third feature; the third feature is downsampled for the third time by the downsampling module. The features output after sampling are fused with the features output by the text prompt word encoder through the first wavelet convolution layer, and the fused features are upsampled by the first wavelet convolution layer; the features output after the first upsampling by the upsampling module are jump-linked with the third feature and then convolved through two layers of 3*3 convolution layers to obtain the fourth feature; the fourth feature is jump-linked with the second feature and then convolved through two layers of 3*3 convolution layers to obtain the fifth feature; the fifth feature is jump-linked with the first feature and then convolved through two layers of 3*3 convolution layers to obtain the sixth feature.

[0029] When using the wavelet convolution layer to perform wavelet convolution, it includes wavelet decomposition and wavelet reconstruction; The expression of wavelet decomposition is: ; The expression of wavelet reconstruction is: ; in, represents the input image, represents the wavelet filter, represents the convolution operation, represents the downsampling operation, represents transposed convolution, represents the reconstruction filter, Represents the feature map of each sub-band.

[0030] In order to ensure that angle information effectively guides dose prediction, an "angle-image" pairing structure will be constructed, and angle encoding information will be used as auxiliary input for the model. In this embodiment, the seven field angles corresponding to the case are discretized and encoded into prompt word vectors, and the attention mechanism is used to guide the model to focus on the correspondence between angle information and image structure, thereby achieving the dose prediction capability for any angle. Figure 3 As shown, the text prompt word encoder includes an Embedding module and three fully connected layers set in sequence; the field angle The embedding module maps this information to a d-dimensional feature vector, which then passes through three fully connected layers. The nonlinear activation functions of these layers extract and generate higher-order angular features (1024-dimensional angular features). These angular features are then fused with the features output after the final downsampling and convolution in the U-Net network. This approach enables the interaction of beam field angle information with imaging features in a high-dimensional space, improving the model's ability to model dose distribution.

[0031] When the Embedding module performs mapping, the calculation formula is: ; The fully connected layer extracts high-order angle features through nonlinear activation functions to obtain angle features , The calculation formula is: ; in, is the field angle, represents the embedding layer, represents the dimension of the embedding vector, Represents the feature vector obtained by mapping; 、 、 Both represent trainable weight matrices, Represents the ReLU activation function. Finally, a 1024-dimensional angle feature is obtained , used for subsequent feature fusion.

[0032] In the task of radiotherapy dose prediction, the dose distribution depends not only on CT images, planned target volumes, and organs at risk, but is also significantly affected by the field angle. Traditional methods based on convolutional neural networks mainly rely on local convolution operations, which have difficulty capturing long-range dependencies, resulting in limited utilization of field angle information. However, this embodiment introduces the Multi-HeadAttention mechanism in the Transformer structure to the high-level features of the U-Net network encoder, enabling the model to model the impact of field angle on dose distribution through the correlation between global features. The specific fusion method is: The KQV (Key-Query-Value) attention mechanism is used to interact between image features and field of view angle features. Specifically: The feature output after the last downsampling and convolution in the U-Net network is taken as X, the feature output by the text prompt word encoder is taken as Z, and the feature Z is used as the key and value in the calculation; the specific calculation method is: ; ; ; ; in, 、 、 Both represent trainable parameters, Represents the scaling factor of the Key dimension to prevent gradient explosion; 、 、 Represents the Query vector (query representation after image feature conversion), the Key vector (key representation after angle feature conversion), and the Value vector (value representation after angle feature conversion). represents transpose; By calculating the degree of match between image features and field angle features, the model can adapt to the dose distribution under a specific field angle.

[0033] To enhance the feature expression capability, a multi-head attention mechanism is further adopted to independently calculate attention in multiple subspaces and concatenate the final results to obtain the fused features. Specifically: ; in, represents the h-th attention head, Represents the output projection matrix (trainable parameter).

[0034] This approach allows the model to learn different attention patterns from multiple perspectives, improving its ability to model field-of-view angle information. Subsequently, the fused features are added to the original image features via a residual connection to preserve the original information, and are further transformed via batch normalization and ReLU activation to enhance model stability.

[0035] Step S3, training a radiotherapy dose prediction model; The radiotherapy dose prediction model constructed in step S2 is trained using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples and label data in step S1.

[0036] The data in the training set is input into the radiotherapy dose prediction model to perform forward calculation and back propagation in sequence, and repeated multiple times until the model converges. When the output on the validation set is the minimum value in multiple trainings, the training can be stopped to obtain the optimal model after training.

[0037] When training the radiotherapy dose prediction model, the Adam optimizer was used, the initial learning rate was set to 1e-4, the batch size was considered to be 16, and the total number of training rounds was set to 250. The learning rate was automatically reduced when the loss did not decrease significantly within several consecutive rounds, and the minimum learning rate was set to 1e-6.

[0038] During training, the model predicts the radiotherapy dose at each radiation field angle based on a sample CT image, its corresponding target area image, and the image of the organ at risk. The loss is then calculated using the predicted radiotherapy dose and the actual radiotherapy dose (i.e., the labeled data). The specific loss function is: ; in, Indicates the true value of radiotherapy dose, represents the expected value of radiotherapy dose, Represents the total number of voxels.

[0039] The model's plan generation effect on the test set can also be quantitatively evaluated. Common evaluation indicators include MAE, SSIM, PSNR, and gamma pass rate, which are defined as follows: ; ; ; in, and represent the true value and the predicted value respectively, is the total number of voxels, is the maximum possible value of the image, MSE is the mean square error between the predicted image and the true image, 、 is the mean, 、 is the variance, is the covariance 、 is a stability constant.

[0040] The Gamma Passing Rate (GPR) is the clinical gold standard for assessing dose distribution consistency in radiotherapy. It comprehensively considers both dose difference and distance-to-agreement (DTA). At each voxel, if both the dose difference is less than a certain threshold (e.g., 3%) and the spatial offset is less than a certain distance (e.g., 3mm), the point is considered "passing." The GPR represents the proportion of passing points: ; In this study, the commonly used clinical threshold of 3% / 3mm (i.e. =3%, =3mm) to calculate GPR. Compared with traditional pixel-level indicators, GPR is more consistent with clinical quality control standards and can effectively capture space-dose coordination errors.

[0041] Step S4, radiotherapy dose prediction; The radiotherapy CT image to be tested is obtained, and the target area and organs at risk of the radiotherapy CT image are outlined to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose at each field angle.

[0042] Example 2 A radiotherapy dose prediction system based on prompt word guidance, comprising: A sample data acquisition module is used to acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples. The radiotherapy dose samples are used as label data. The radiotherapy dose prediction model construction module is used to build a radiotherapy dose prediction model. The radiotherapy dose prediction model includes a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk are input into the U-Net network for feature extraction. The radiation field angle is input into the text prompt word encoder. The features output by the last downsampling in the U-Net network are convolved and fused with the features output by the text prompt word encoder. The fused features are then convolved and upsampled. a radiotherapy dose prediction model training module, for training the radiotherapy dose prediction model constructed by the radiotherapy dose prediction model construction module using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and label data in the sample data acquisition module; The radiotherapy dose prediction module is used to obtain the radiotherapy CT image to be tested, and to outline the target area and organs at risk of the radiotherapy CT image to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose under each field angle.

[0043] Example 3 A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a radiotherapy dose prediction method based on prompt word guidance.

[0044] The computer device may be a desktop computer, a notebook computer, a PDA, a cloud server, etc. The computer device may interact with the user via a keyboard, a mouse, a remote control, a touchpad, or a voice control device.

[0045] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk equipped with the computer device, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Of course, the memory can also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the radiotherapy dose prediction method based on prompt word guidance. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0046] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to execute program code stored in the memory or process data, such as executing the program code of the radiotherapy dose prediction method based on prompt word guidance.

[0047] Example 4 A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor is caused to perform the steps of a radiotherapy dose prediction method based on prompt word guidance.

[0048] The computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor to enable the at least one processor to perform the steps of the radiotherapy dose prediction method based on prompt word guidance as described above.

[0049] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment method can be implemented by means of software plus the necessary general hardware platform, or of course by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for enabling a terminal device (which can be a mobile phone, computer, server or network device, etc.) to execute the radiotherapy dose prediction method based on prompt word guidance described in the embodiment of the present application.

Claims

1. A radiotherapy dose prediction method based on prompt word guidance, characterized in that: The following steps are involved: Step S1, obtaining sample data; Acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples, and use the radiotherapy dose samples as label data; Step S2, constructing a radiotherapy dose prediction model; A radiotherapy dose prediction model was constructed, which included a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk were input into the U-Net network for feature extraction. The radiation field angle was input into the text prompt word encoder. The features output by the last downsampling in the U-Net network were convolved and fused with the features output by the text prompt word encoder. The fused features were then convolved and upsampled. Step S3, training a radiotherapy dose prediction model; The radiotherapy dose prediction model constructed in step S2 is trained using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and label data in step S1; Step S4, radiotherapy dose prediction; The radiotherapy CT image to be tested is obtained, and the target area and organs at risk of the radiotherapy CT image are outlined to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose at each field angle.

2. The radiotherapy dose prediction method based on prompt word guidance according to claim 1, characterized in that: In step S2, the U-Net network includes a wavelet convolution layer, a downsampling module, an upsampling module and a 3*3 convolution layer; The input of the U-Net network is first subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the first feature; the first feature is downsampled for the first time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the second feature; the second feature is downsampled for the second time by the downsampling module, and the output feature is then subjected to wavelet convolution by two layers of wavelet convolution layers to obtain the third feature; the third feature is downsampled for the third time by the downsampling module, and the output feature is fused with the feature output by the text prompt word encoder through the first layer of wavelet convolution layer, and the fused feature is upsampled by the first layer of wavelet convolution layer; the feature output by the upsampling module after the first upsampling is skip-linked with the third feature, and then convolved through two layers of 3*3 convolution layers to obtain the fourth feature; the fourth feature is skip-linked with the second feature, and then convolved through two layers of 3*3 convolution layers to obtain the fifth feature; The fifth feature is skip-linked with the first feature and then convolved through two 3*3 convolutional layers to obtain the sixth feature.

3. The radiotherapy dose prediction method based on prompt word guidance according to claim 2, characterized in that: In step S2, when wavelet convolution is performed using the wavelet convolution layer, wavelet decomposition and wavelet reconstruction are included; The expression of wavelet decomposition is: ； The expression of wavelet reconstruction is: ； in, represents the input image, represents the wavelet filter, represents the convolution operation, represents the downsampling operation, represents transposed convolution, represents the reconstruction filter, Represents the feature map of each sub-band.

4. The radiotherapy dose prediction method based on prompt word guidance according to claim 1, characterized in that: In step S2, the text prompt word encoder includes an Embedding module and three fully connected layers arranged in sequence; The field angle is mapped to a d-dimensional feature vector through the Embedding module. The d-dimensional feature vector is sequentially passed through three fully connected layers to obtain a higher-order angle feature. This angle feature is fused with the feature output after the last downsampling and convolution in the U-Net network.

5. The radiotherapy dose prediction method based on prompt word guidance according to claim 4, characterized in that: When the Embedding module performs mapping, the calculation formula is: ； The fully connected layer extracts high-order angle features through nonlinear activation functions to obtain angle features , The calculation formula is: ； in, is the field angle, represents the embedding layer, represents the dimension of the embedding vector, Represents the feature vector obtained by mapping; 、 、 Both represent trainable weight matrices, Represents the ReLU activation function.

6. The radiotherapy dose prediction method based on prompt word guidance according to claim 1, characterized in that: In step S2, when the features output by the U-Net network after the last downsampling and convolution are fused with the features output by the text prompt word encoder, the specific steps are as follows: Step S2-1: Use the KQV attention mechanism for interaction, specifically: The feature output after the last downsampling and convolution in the U-Net network is taken as X, the feature output by the text prompt word encoder is taken as Z, and the feature Z is used as the key and value in the calculation; the specific calculation method is: ； ； ； ； in, 、 、 Both represent trainable parameters, Indicates the scaling factor of the Key dimension; 、 、 Represents Query vector, Key vector, and Alue vector; represents transpose; Step S2-2 uses a multi-head attention mechanism to independently calculate attention in multiple subspaces and concatenate the final results to obtain the fused features; specifically: ； in, represents the h-th attention head, Represents the output projection matrix.

7. The radiotherapy dose prediction method based on prompt word guidance according to claim 1, characterized in that: In step S3, the Adam optimizer is used to train the radiotherapy dose prediction model, with an initial learning rate of 1e-4, a batch size of 16, and a total number of training rounds of 250. The learning rate is automatically reduced when the loss does not decrease significantly over several consecutive rounds, and the minimum learning rate is set to 1e-6. During training, the loss function is: ； in, Indicates the true value of radiotherapy dose, represents the expected value of radiotherapy dose, Represents the total number of voxels.

8. A radiotherapy dose prediction system based on prompt word guidance, characterized in that: include: A sample data acquisition module is used to acquire radiotherapy sample data, which includes radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and radiotherapy dose samples. The radiotherapy dose samples are used as label data. The radiotherapy dose prediction model construction module is used to build a radiotherapy dose prediction model. The radiotherapy dose prediction model includes a U-Net network and a text prompt word encoder. CT images, target images, and images of organs at risk are input into the U-Net network for feature extraction. The radiation field angle is input into the text prompt word encoder. The features output by the last downsampling in the U-Net network are convolved and fused with the features output by the text prompt word encoder. The fused features are then convolved and upsampled. a radiotherapy dose prediction model training module, for training the radiotherapy dose prediction model constructed by the radiotherapy dose prediction model construction module using the radiotherapy CT sample images, target area sample images, organ at risk sample images, field angle samples, and label data in the sample data acquisition module; The radiotherapy dose prediction module is used to obtain the radiotherapy CT image to be tested, and to outline the target area and organs at risk of the radiotherapy CT image to obtain the target area image and the organ at risk image; the radiotherapy CT image, the target area image, and the organ at risk image are input into the trained radiotherapy dose prediction model, and the radiotherapy dose prediction model outputs the predicted radiotherapy dose under each field angle.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images