Multi-disease radiotherapy dose prediction method, system, equipment and medium

Abstract

The invention discloses a multi-disease radiotherapy dose prediction method, system and device and a medium, belongs to radiotherapy dose prediction, and aims to solve the technical problem that an existing deep neural network lacks cross-disease universality and generalization ability. In a dose prediction model constructed by the method, image features obtained after a CT image, a critical organ image and a target region image are input into an encoder are input into a global average pooling module, image features are obtained, and target region features are obtained after the target region image is input into a target region encoder; after image features and target region features are input into a cross attention sub-module, obtained attention guiding features are fused with the image features and then input into a decoder to obtain image features, and language embedding features output by a text encoder are projected by a linear mapping layer to obtain task vectors. And splicing the task vector and the fusion feature, inputting the spliced task vector and fusion feature into a multi-layer perceptron, inputting the obtained dynamic convolution kernel and bias and image features into a dynamic solution head, and obtaining a dose prediction value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence medical technology, and relates to radiotherapy dose prediction, in particular to a multi-disease radiotherapy dose prediction method, system, device and medium. BACKGROUND

[0002] Radiotherapy as one of the core means of cancer treatment aims to kill tumor cells while protecting the surrounding normal tissues as much as possible. In recent years, intensity-modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) have become the mainstream radiotherapy technology in clinical practice due to their excellent dose modulation capability. In this treatment process, the patient is usually irradiated by multiple high-energy rays from different angles. These rays accurately deliver the prescribed dose to the planning target volume (PTV) through a complex optimization process and form a complex dose distribution around the target volume. In order to achieve the ideal radiotherapy effect, the planner needs to balance between “sufficient PTV coverage” and “OARs protection”, which often requires multiple iterations and manual parameter adjustment. However, this process not only consumes time and effort, but also highly depends on the experience level of the physicist, which can easily lead to unstable plan quality and even treatment delay. Therefore, an efficient and accurate automatic dose distribution prediction method has gradually become an important research direction to improve the efficiency and consistency of radiotherapy plan design.

[0003] With the rapid development of artificial intelligence and neural network technology, dose distribution automatic prediction based on deep neural network has made significant progress. The model usually takes the CT image of the patient, the target volume and the organs at risk as input to predict the dose value of each voxel for a specific disease. This paradigm has achieved good results in dose prediction tasks for multiple diseases. For example, Wang et al. proposed an error correction network for glioma dose prediction in the head and neck, Chen et al. proposed a dose prediction model combining graph convolution and self-attention modules to predict esophageal cancer dose distribution, and Wen et al. proposed a dose prediction method based on transfer learning strategy to migrate the dose prediction model from rectal cancer to cervical cancer, which belongs to the same pelvic position.

[0004] The application patent application with the application number 202311069534.5 also discloses a radiotherapy dose prediction system and method based on artificial intelligence, which comprises a 3D-Unet deep learning model construction module, a first model training module and a second model training module. The 3D-Unet deep learning model construction module is used to construct a 3D-Unet deep learning model. The first model training module is used to train a first model and extract an intermediate dose. The second model training module is used to train a second model and output a final dose distribution. The 3D-Unet deep learning model is a deep neural network structure, and the input end and the output end are linked. The input end inputs the critical organ, the target area and the distance from the target area. The output end outputs the real three-dimensional dose distribution map. The specific steps of the first model training module to train the first model and extract the intermediate dose are as follows: through learning rate gradient descent, the best matching relationship between the input end and the output end is found by iterative steps, and the best matching relationship is the first model. After the first model training is completed, the critical organ, the target area and the distance from the target area of a new patient are input into the first model, and the first model outputs a predicted dose distribution map, which is the intermediate dose. The specific steps of the second model training module to train the second model and output the final dose distribution are as follows: the predicted dose distribution map is input into the input end of the 3D-Unet deep learning model, and the second model is trained to output the final dose distribution.

[0005] As in the above-mentioned application patent application, most of the existing dose prediction models have a key limitation: they are usually designed for a single disease, i.e. each model can only predict the dose distribution of a specific cancer, lacking the cross-disease generality and generalization ability. When it is necessary to extend it to a new disease in clinic, it is often necessary to rebuild the model and retrain the model, which not only increases the time cost, but also limits the clinical practicability of the model. Therefore, it is necessary to provide a radiotherapy dose prediction method applicable to multiple diseases SUMMARY The purpose of the present application is to solve the technical problem that the existing radiotherapy dose prediction deep neural network lacks cross-disease generality and generalization ability, and to provide a multi-disease radiotherapy dose prediction method, system, device and medium, which can be applied to radiotherapy dose prediction of multiple cancers at the same time.

[0006] In order to achieve the above-mentioned purpose, the present application specifically adopts the following technical scheme: A multi-disease radiotherapy dose prediction method, comprising the following steps: Step S1, acquiring sample data; acquiring CT sample images and corresponding critical organ sample images, target area sample images, real dose distribution sample images and disease labels; Step S2, constructing a dose prediction model; The dose prediction model comprises a U-Net network, a global average pooling module, a target region guided attention module, a text encoder pre-training module and a dynamic decoding head; the U-Net network comprises an encoder and a decoder, the target region guided attention module comprises a target region encoder, a cross-attention submodule, and the text encoder pre-training module comprises a text encoder, a linear mapping layer and a multi-layer perceptron; The CT image, the critical organ image and the target region image are input into the encoder of the U-Net network to obtain image features The image features are input into the global average pooling module to obtain image features The target region image is input into the target region encoder to obtain target region features The image features and the target region features are input into the cross-attention submodule to obtain attention guided features The attention guided features are fused with the image features to obtain fused features The fused features are input into the decoder to obtain image features The disease label is input into the text encoder pre-training module to obtain prompt words and then input into the text encoder, and the text encoder outputs language embedding features The language embedding features are projected by the linear mapping layer to obtain task vectors The task vectors are spliced with the fused features to input into the multi-layer perceptron to obtain corresponding dynamic convolution kernels and biases The dynamic convolution kernels and biases are input into the dynamic decoding head to obtain dose prediction values Step S3, training the dose prediction model; The sample data obtained in step S1 is used to train the dose prediction model constructed in step S2; Step S4, real-time prediction; The CT image to be measured and the corresponding critical organ image, target region image and disease label are obtained, and the trained dose prediction model is input, and the dose prediction model outputs the dose prediction value.

[0007] Further, in step S2, when the image features and the target region features are input into the cross-attention submodule, the target region features are used as key vectors and value vectors, and the image features ​As a query vector, the feature alignment and fusion are performed through the cross-attention sub-module to obtain the fusion feature .

[0008] Further, in step S2, the dynamic convolution kernel and the bias are updated according to the image feature . The dynamic convolution kernel and the bias , , , , , are divided into three parts

[0009] Further, after the image feature is input into the dynamic decoding head, the three convolution layers sequentially convert the image feature into a continuous dose value : ; wherein, represents a nonlinear activation function, represents a 3D convolution operation.

[0010] Further, in step S3, when training the dose prediction model, the loss function is: ; wherein, represents the predicted dose of the i-th sample, represents the true dose of the i-th sample, represents the total number of samples. Further, for a data set composed of multiple disease groups,

[0011] the training target of each sub-data set is to minimize the average loss of all samples, which is represented as: ; wherein, represents the number of disease groups, represents the number of samples included in the i-th disease group, represents the CT image of the j-th sample under the i-th disease group, represents the structure mask of the j-th sample under the i-th disease group, represents the task prompt vector, represents the dose prediction model with parameters , represents the loss function, ​represents the true dose distribution of the jth sample under the ith disease.

[0012] A multi-disease radiotherapy dose prediction system comprises: A sample data acquisition module is configured to acquire CT sample images, corresponding organ-at-risk sample images, target region sample images, true dose distribution sample images, and disease labels; A dose prediction model construction module is configured to construct a dose prediction model, wherein the dose prediction model comprises a U-Net network, a global average pooling module, a target region guided attention module, a text encoder pre-training module, and a dynamic decoding head; the U-Net network comprises an encoder and a decoder; the target region guided attention module comprises a target region encoder and a cross-attention submodule; the text encoder pre-training module comprises a text encoder, a linear mapping layer, and a multi-layer perceptron; The CT images, the organ-at-risk images, and the target region images are input into the encoder of the U-Net network to obtain image features The image features are input into the global average pooling module to obtain image features The target region images are input into the target region encoder to obtain target region features The image features and the target region features are input into the cross-attention submodule to obtain attention guided features The attention guided features are fused with the image features to obtain fused features The fused features are input into the decoder to obtain image features The disease labels are input into the text encoder pre-training module to obtain prompt words and then input into the text encoder; the text encoder outputs language embedding features The language embedding features are projected by the linear mapping layer to obtain task vectors The task vectors are concatenated with the fused features to be input into the multi-layer perceptron to obtain corresponding dynamic convolution kernels and biases The dynamic convolution kernels and biases and the image features are input into the dynamic decoding head to obtain dose prediction values. A dose prediction model training module is configured to train the dose prediction model constructed by the dose prediction model construction module using the sample data acquired by the sample data acquisition module. A real-time prediction module is configured to acquire CT images, corresponding organ-at-risk images, target region images, and disease labels to be measured, and input the trained dose prediction model; the dose prediction model outputs dose prediction values.

[0013] A computer device comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0014] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0015] The beneficial effects of the present application are as follows: In the present application, the dose prediction model comprises a U-Net network, a global average pooling module, a target region guided attention module, a text encoder pre-training module and a dynamic decoding head; the global average pooling module can compress the features output by the encoder of the U-Net network and obtain image-level features, which contain global image semantic information and are the input feature source of the subsequent attention module; by inserting the target region guided attention module at the bottleneck layer of the encoder and the decoder of the U-Net network, explicit alignment and information interaction of the target region features and the global image features can be realized; by introducing the text encoder pre-training module, the semantic encoding of different task categories is performed to distinguish the semantic features of different diseases and retain the potential semantic correlation between diseases; a class-specific decoding head can generate a continuous dose distribution according to the convolution kernel parameters of each category. The dose prediction model introduces a semantic-driven task prompt mechanism and a target region coding guide mechanism, the model can learn the semantic relationship between different diseases, realize semantic distinction and correlation-aware feature fusion in multi-task dose prediction, guide the model to focus on the target region, effectively improve the spatial accuracy of the dose distribution and improve the accuracy of the model in predicting the dose of radiotherapy for multiple diseases. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the present application; Figure 2 is a structural diagram of the dose prediction model in the present application; Figure 3 is a connection diagram of the dynamic convolution kernel and bias and the dynamic decoding head in the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments of the present application.

[0018] Therefore, all other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative labor fall within the scope of protection of the present application.

[0019] Embodiment 1 The present embodiment provides a multi-disease radiotherapy dose prediction method, which can simultaneously process radiotherapy dose prediction tasks of multiple diseases such as rectal cancer, cervical cancer and gastric cancer in the same network model. The method learns the semantic relationship between different diseases through language embedding, thereby realizing semantic distinction and feature fusion of correlation perception in dose prediction; at the same time, the target mask serves as a spatial constraint input to guide the model to focus on the target area, effectively improving the spatial accuracy of the dose distribution. As shown in Figure 1 The method comprises the following steps: Step S1, obtaining sample data; Obtaining CT sample images and corresponding organ at risk sample images, target area sample images, real dose distribution sample images and disease labels.

[0020] The sample data set is a radiotherapy data set collection composed of multiple diseases , wherein the zth sub-data set is defined as , which represents that the zth disease contains samples, and j represents the sample number under a certain disease. For each sample, represents the CT image of the jth sample under the ith disease, represents the structure mask (including the target area (PTV) and the organ at risk (OAR)) of the jth sample under the ith disease, represents the real dose distribution of the jth sample under the ith disease. At the same time, each data set corresponds to a task category (task category, i.e. disease, used for text encoder pre-training module output prompt word), wherein represents the full set of all task categories (i.e. diseases).

[0021] Step S2, constructing a dose prediction model; The dose prediction model comprises a U-Net network, a global average pooling module, a target area guided attention module, a text encoder pre-training module and a dynamic decoding head, as shown in Figure 2 .

[0022] The U-Net network comprises an encoder and a decoder, the encoder is provided with 5 layers, the decoder is provided with 4 layers, and the U-Net network is designed symmetrically to capture multi-scale semantic features and spatial details. In the encoder part, each stage is composed of a 3*3*3 three-dimensional convolution, a ReLU activation and a 2*2*2 maximum pooling, and with down-sampling, the spatial resolution of the feature map is halved layer by layer, and the number of channels is gradually increased (64→1024) to obtain deep image features . Image features are fused with the features output by the cross-attention submodule to obtain fused features , which are input into the decoder, the decoder is up-sampled through transposed convolution, and the jump connection features of the corresponding encoder layer are fused to further restore the spatial resolution and refine the structural information. In addition, the remaining structure of the U-Net network can adopt the prior art, and those skilled in the art do not need to pay creative labor.

[0023] In clinical radiotherapy, the target region is the core area of dose design and is crucial to the treatment effect. In order to strengthen the model's attention to the target region, a target region guided attention module is inserted at the bottleneck layer of the encoder and the decoder to realize the explicit alignment and information interaction of the target region features and the global image features. The target region guided attention module comprises a target region encoder and a cross-attention submodule. The target region mask is first input into the target region encoder, local structural features are extracted through two layers of 3D convolution, and then the local structural features are compressed into a global representation with a spatial dimension of 1*1*1 through adaptive average pooling, and finally the target region features with a channel dimension of 256 are output through a 1*1*1 convolution ; the target region features are taken as a key vector and a value vector (i.e. Key and Value), and the image features output by the encoder are taken as a query vector (i.e. Query), and the cross-attention submodule is used for feature alignment and fusion to obtain fused features .

[0024] In order to enhance the generalization ability of the model among multiple disease tasks, a text encoder pre-training module is introduced to encode the semantics of different task categories, so as to distinguish the semantic features of different diseases and retain the potential semantic correlation between diseases. The text encoder pre-training module comprises a text encoder, a linear mapping layer and a multi-layer perceptron; for each category, the module first adopts a medical prompt word template "A photo of a [cls]", wherein [cls] is the name of the disease corresponding to the category cls; then a 512-dimensional language embedding feature is generated through the text encoder, and then the language embedding feature is projected into an image semantic space through the linear mapping layer (512→256) to generate a task vector ; finally, the task vector is fused with the fused features The concatenated features are input into a separate multi-layer perceptron (MLP) to generate dynamic convolution kernels and biases for each class .

[0025] Due to the significant difference in dose distribution patterns of different diseases, directly using a unified decoder can easily lead to feature confusion and degradation of prediction performance. Therefore, a class-specific dynamic decoding head is designed. The dynamic decoding head can generate continuous dose distribution maps according to the convolution kernel parameters of each class, thereby realizing personalized dose distribution modeling. In the dynamic decoding head, conditional convolution is used, which includes three convolution layers with 1*1*1 kernels arranged in sequence and a nonlinear function. The first two convolution layers each contain 8 channels, and the last convolution layer has 1 channel. To start the process, the dynamic convolution kernels and biases of each class are divided into three parts , 、 、 , 、 、 respectively corresponding to the three convolution layers input into the dynamic decoding head, as shown in Figure 3 . Figure 3 In each box, the three circles represent the three groups of parameters that the disease needs to input into the dynamic decoding head, that is, a group of parameters is assigned to each convolution layer.

[0026] Image features After being input into the dynamic decoding head, the three convolution layers convert the image features into continuous dose values : ; wherein represents a nonlinear activation function, represents a 3D convolution operation.

[0027] The CT image, the critical organ image, and the target image are input into the encoder of the U-Net network to obtain image features , the image features are input into the global average pooling module to obtain image features , the target image is input into the target encoder to obtain target features , the image features and the target features are input into the cross-attention submodule to obtain attention-guided features , the attention-guided features are fused with the image features to obtain fused features , and the fused features are input into the decoder to obtain image features After the disease label is input into the pre-trained module of the text encoder, prompt words are obtained and then input into the text encoder. The text encoder outputs language embedding features. Language embedding features The task vector is obtained after projection through the linear mapping layer. Task vector With fusion features After concatenation, the data is input into a multilayer perceptron to obtain the corresponding dynamic convolution kernel and bias. Dynamic convolution kernels and biases Image features Input the dynamic decoding head and obtain the dose prediction value.

[0028] Step S3: Train the dose prediction model; The dose prediction model constructed in step S2 is trained using the sample data obtained in step S1.

[0029] To effectively train the dose prediction model, the Adam optimizer, an adaptive optimization algorithm widely used in deep learning, is employed. Adam combines the advantages of momentum and RMSProp, automatically adjusting the learning rate for different parameters based on dynamic estimates of the first and second gradients, thereby accelerating model convergence and enhancing training stability.

[0030] During training, the mean squared error loss function (MSE) is used as the optimization objective to measure the difference between the predicted dose distribution and the true dose distribution. The overall fitting accuracy is measured by calculating the squared mean of the voxel-level measurement differences using MSE. The specific loss function is as follows: ; in, Indicates the first Predicted dose for each sample Indicates the first The actual dose of each sample This represents the total number of samples.

[0031] For datasets composed of multiple diseases Each subset The training objective is to minimize the average loss across all samples, expressed as: ; in, Indicates the number of diseases. This represents the number of samples included in the i-th disease category. This represents the CT image of the j-th sample under the i-th disease. This represents the structural mask of the j-th sample under the i-th disease. This represents the task prompt vector. denotes a dose prediction model with parameters , denotes a loss function, denotes the true dose distribution of the jth sample under the ith disease.

[0032] Step S4, real-time prediction; Obtain the CT image to be tested, the corresponding critical organ image, the target area image and the disease label, and input the trained dose prediction model. The dose prediction model outputs the dose prediction value.

[0033] Example 2 The embodiment provides a multi-disease radiotherapy dose prediction system, comprising the following steps: A sample data acquisition module is used to acquire CT sample images and corresponding critical organ sample images, target area sample images, true dose distribution sample images and disease labels.

[0034] The sample data set is a radiotherapy data set collection composed of multiple diseases , wherein the zth sub-data set is defined as , which represents that the zth disease contains samples, and j represents the sample number under a certain disease. For each sample, denotes the CT image of the jth sample under the ith disease, denotes the structure mask (including the target area (PTV) and the critical organ (OAR)) of the jth sample under the ith disease, denotes the true dose distribution of the jth sample under the ith disease. At the same time, each data set corresponds to a task category (the task category, i.e. the disease, is used for the output prompt word of the text encoder pre-training module), wherein denotes the universal set of all task categories (i.e. diseases).

[0035] A dose prediction model construction module is used to construct a dose prediction model. The dose prediction model includes a U-Net network, a global average pooling module, a target area guided attention module, a text encoder pre-training module and a dynamic decoding head, as shown in Figure 2 .

[0036] The U-Net network includes an encoder and a decoder. The encoder is set to 5 layers, and the decoder is set to 4 layers, which are symmetrically designed to capture multi-scale semantic features and spatial details. In the encoder part, each stage is composed of 3*3*3 three-dimensional convolution, ReLU activation and 2*2*2 maximum pooling. With down-sampling, the spatial resolution of the feature map is halved layer by layer, and the number of channels is gradually increased (64→1024) to obtain deep image features . Image features fuse the features output by the cross-attention sub-module to obtain fused features , as an input of the decoder, the decoder is upsampled by transposed convolution and fuses the skip connection features corresponding to the encoder layers to further restore the spatial resolution and refine the structural information. In addition, the remaining structure of the U-Net network can adopt the prior art, and those skilled in the art do not need to pay creative labor.

[0037] In clinical radiotherapy, the target region is the core area of dose design and is crucial to the treatment effect. To strengthen the model's attention to the target region, a target region guided attention module is inserted at the bottleneck layer of the encoder and the decoder to realize the explicit alignment and information interaction of the target region features and the global image features. The target region guided attention module includes a target region encoder and a cross-attention sub-module. The target region mask is first input into the target region encoder, and the local structural features are extracted through two layers of 3D convolution, and then compressed into a global representation with a spatial dimension of 1*1*1 through adaptive average pooling, and finally output through a 1*1*1 convolution to obtain a target region feature with a channel dimension of 256 . The target region feature is taken as the key vector and the value vector (i.e., Key and Value), and the image feature output by the encoder is taken as the query vector (i.e., Query), and the cross-attention sub-module is used for feature alignment and fusion to obtain the fused feature .

[0038] To enhance the generalization ability of the model among multiple disease tasks, a text encoder pre-training module is introduced to encode the semantics of different task categories, so as to distinguish the semantic features of different diseases and preserve the potential semantic correlation between diseases. The text encoder pre-training module includes a text encoder, a linear mapping layer and a multi-layer perceptron; for each category, the module first uses the medical prompt word template "A photo of a [cls]", where [cls] is the name of the disease corresponding to the category cls; then generates a 512-dimensional language embedding feature through the text encoder, and then projects it to the image semantic space through the linear mapping layer (512→256) to generate a task vector ; finally, the task vector is concatenated with the fused feature and input into an independent multi-layer perceptron (MLP), and the multi-layer perceptron is used to generate dynamic convolution kernels and biases for each category .

[0039] Due to the significant difference in dose distribution regularity of different diseases, if a unified decoder is directly used, it is easy to cause feature confusion and prediction performance degradation. Therefore, a class-specific dynamic decoding head is designed. The dynamic decoding head can generate a continuous dose distribution map according to the convolution kernel parameters of each class, thereby realizing personalized dose distribution modeling. In the dynamic decoding head, conditional convolution is used, which includes three sequentially arranged convolution layers with 1*1*1 kernel and a nonlinear function. The first two convolution layers each contain 8 channels, and the last convolution layer has 1 channel. To start the process, the dynamic convolution kernel and bias of each class are divided into three parts are divided into three parts 、 、 , 、 、 respectively correspond to the three convolution layers input into the dynamic decoding head.

[0040] Image features After inputting the dynamic decoding head, the three convolution layers sequentially convert the image features into continuous dose values : ; wherein represents a nonlinear activation function, represents a 3D convolution operation.

[0041] The CT image, the critical organ image, and the target image are input into the encoder of the U-Net network to obtain image features , the image features are input into the global average pooling module to obtain image features , the target image is input into the target encoder to obtain target features , the image features and the target features are input into the cross-attention submodule to obtain attention-guided features , the attention-guided features are fused with the image features to obtain fused features , the fused features are input into the decoder to obtain image features , the disease label is input into the text encoder pre-training module to obtain a prompt word and then input into the text encoder, the text encoder outputs language embedding features , the language embedding features are projected through a linear mapping layer to obtain a task vector , the task vector is fused with the fused features The multi-layer perceptron is inputted after splicing to obtain a corresponding dynamic convolution kernel and bias , the dynamic convolution kernel and bias and image features The dynamic decoding head is inputted to obtain a dose prediction value.

[0042] The dose prediction model training module is configured to train the dose prediction model constructed by the dose prediction model construction module using the sample data obtained by the sample data acquisition module.

[0043] To effectively train the dose prediction model, an Adam optimizer is used, which is a widely used adaptive optimization algorithm in deep learning. Adam combines the advantages of momentum method and RMSProp, and can automatically adjust the learning rate for different parameters according to the dynamic estimates of first and second order gradients, thereby accelerating model convergence and enhancing training stability.

[0044] In the training process, the mean square error loss function (MSE) is used as the optimization target to measure the difference between the predicted dose distribution and the true dose distribution. The overall fitting accuracy is measured by calculating the square average of the voxel-level measurement difference using MSE. The specific loss function is: ; Wherein, represents the predicted dose of the i-th sample, represents the true dose of the i-th sample, represents the total number of samples. For a data set composed of multiple disease groups , the training target of each sub-data set

[0045] is to minimize the average loss of all samples, which is represented as: ; ; Wherein, represents the number of disease groups, represents the number of samples included in the i-th disease group, represents the CT image of the j-th sample under the i-th disease group, represents the structure mask of the j-th sample under the i-th disease group, represents the task prompt vector, represents the dose prediction model with parameters , and represents the loss function, represents the true dose distribution of the j-th sample under the i-th disease group.

[0046] ​A real-time prediction module is configured to acquire a CT image to be tested, a corresponding critical organ image, a target region image, and a disease label, and input the trained dose prediction model, and the dose prediction model outputs a dose prediction value.

[0047] Embodiment 3 A computer device includes a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the multi-disease radiotherapy dose prediction method.

[0048] The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server, or the like. The computer device can interact with a user through a keyboard, a mouse, a remote controller, a touchpad, a voice control device, or the like.

[0049] The memory includes at least one type of readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or D interface display memory, or the like), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, or the like. In some embodiments, the memory can be an internal storage unit of the computer device, such as a hard disk or a 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, a smart media card (SMC), a secure digital (SD) card, a flash card, or the like. Of course, the memory can include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is usually used to store an operating system and various application software installed in the computer device, such as program codes of the multi-disease radiotherapy dose prediction method, or the like. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0050] The processor can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor is usually used to control the overall operation of the computer device. In this embodiment, the processor is used to run program codes or process data stored in the memory, such as running program codes of the multi-disease radiotherapy dose prediction method.

[0051] Embodiment 4 A computer readable storage medium stores a computer program, the computer program is executed by a processor to make the processor execute steps of a multi-disease radiotherapy dose prediction method.

[0052] The computer readable storage medium stores an interface display program, the interface display program is executed by at least one processor to make the at least one processor execute steps of the multi-disease radiotherapy dose prediction method.

[0053] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server or network device, etc.) execute the multi-disease radiotherapy dose prediction method described in the embodiments of the present application.

Claims

1. A method for predicting radiotherapy dose for multiple diseases, characterized in that, Includes the following steps: Step S1: Obtain sample data; Acquire CT sample images and corresponding critical organ sample images, target area sample images, true dose distribution sample images, and disease labels; Step S2: Construct a dose prediction model; The dose prediction model includes a U-Net network, a global average pooling module, a target-guided attention module, a text encoder pre-training module, and a dynamic decoder head; the U-Net network includes an encoder and a decoder, the target-guided attention module includes a target encoder and a cross-attention sub-module, and the text encoder pre-training module includes a text encoder, a linear mapping layer, and a multilayer perceptron. CT images, critical organ images, and target area images are input into the encoder of the U-Net network to obtain image features. Image features Image features are obtained after inputting into the global average pooling module. The target region image is input into the target region encoder to obtain the target region features. Image features Target features Attention-guided features are obtained after inputting the cross-attention submodule. Attention-guided features Image features After fusion, fusion features are obtained. fusion features Image features are obtained after input into the decoder. After the disease label is input into the pre-trained module of the text encoder, prompt words are obtained and then input into the text encoder. The text encoder outputs language embedding features. Language embedding features The task vector is obtained after projection through the linear mapping layer. Task vector With fusion features After concatenation, the data is input into a multilayer perceptron to obtain the corresponding dynamic convolution kernel and bias. Dynamic convolution kernels and biases Image features Input the dynamic decoding head and obtain the dose prediction value; Step S3: Train the dose prediction model; The dose prediction model constructed in step S2 is trained using the sample data obtained in step S1; Step S4, real-time prediction; The system acquires the CT image to be tested, along with the corresponding critical organ image, target area image, and disease label, and inputs them into the trained dose prediction model. The dose prediction model then outputs the dose prediction value.

2. The method for predicting radiotherapy dose for multiple diseases as described in claim 1, characterized in that, In step S2, image features Target features When inputting the cross-attention submodule, target region features Image features are key and value vectors. As the query vector, the cross-attention submodule is used for feature alignment and fusion to obtain the fused features. .

3. The method for predicting radiotherapy dose for multiple diseases as described in claim 1, characterized in that, In step S2, the dynamic convolution kernel and bias... Image features When inputting the dynamic decoding head, various types of dynamic convolution kernels and biases are used. Divided into three parts , , , , , These correspond to the three convolutional layers of the input dynamic decoding head, respectively.

4. The method for predicting radiotherapy dose for multiple diseases as described in claim 3, characterized in that, Image features After inputting the dynamic decoding head, the three convolutional layers sequentially process the image features. Converted to continuous dose value : ； in, Represents a non-linear activation function. This indicates a 3D convolution operation.

5. The method for predicting radiotherapy dose for multiple diseases as described in claim 1, characterized in that, In step S3, when training the dose prediction model, the loss function is: ； in, Indicates the first Predicted dose for each sample Indicates the first The actual dose of each sample This represents the total number of samples.

6. The method for predicting radiotherapy dose for multiple diseases as described in claim 5, characterized in that, For datasets composed of multiple diseases Each subset The training objective is to minimize the average loss across all samples, expressed as: ； in, Indicates the number of diseases. This represents the number of samples included in the i-th disease category. This represents the CT image of the j-th sample under the i-th disease. This represents the structural mask of the j-th sample under the i-th disease. This represents the task prompt vector. The parameter is Dose prediction model, Represents the loss function. This represents the true dose distribution of the j-th sample under the i-th disease.

7. A multi-disease radiotherapy dose prediction system, characterized in that, include: The sample data acquisition module is used to acquire CT sample images and corresponding critical organ sample images, target area sample images, true dose distribution sample images, and disease labels; The dose prediction model building module is used to build a dose prediction model. The dose prediction model includes a U-Net network, a global average pooling module, a target-guided attention module, a text encoder pre-training module, and a dynamic decoder head. The U-Net network includes an encoder and a decoder. The target-guided attention module includes a target encoder and a cross-attention sub-module. The text encoder pre-training module includes a text encoder, a linear mapping layer, and a multilayer perceptron. CT images, critical organ images, and target area images are input into the encoder of the U-Net network to obtain image features. Image features Image features are obtained after inputting into the global average pooling module. The target region image is input into the target region encoder to obtain the target region features. Image features Target features Attention-guided features are obtained after inputting the cross-attention submodule. Attention-guided features Image features After fusion, fusion features are obtained. fusion features Image features are obtained after input into the decoder. After the disease label is input into the pre-trained module of the text encoder, prompt words are obtained and then input into the text encoder. The text encoder outputs language embedding features. Language embedding features The task vector is obtained after projection through the linear mapping layer. Task vector With fusion features After concatenation, the data is input into a multilayer perceptron to obtain the corresponding dynamic convolution kernel and bias. Dynamic convolution kernels and biases Image features Input the dynamic decoding head and obtain the dose prediction value; The dose prediction model training module is used to train the dose prediction model constructed by the dose prediction model construction module using the sample data obtained by the sample data acquisition module. The real-time prediction module is used to acquire the CT image to be tested, as well as the corresponding critical organ image, target area image, and disease label, and input them into the trained dose prediction model. The dose prediction model outputs the dose prediction value.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images