Interactive radiotherapy plan generation method and device based on deep learning, and electronic equipment

Abstract

The application provides a kind of based on deep learning interactive radiotherapy plan generation method, device, electronic equipment and computer readable storage medium, interactive radiotherapy plan generation method includes collecting radiotherapy plan data, using plan optimization system to the radiotherapy plan data collected reoptimization, obtain the data set for training. Using the data set divided to the deep learning model is trained, obtains interactive dose distribution generation tool. The medical image of patient and initial prompt data are input into interactive dose distribution generation tool, and dose distribution data is generated in real time for evaluation, if it does not meet the radiotherapy standard, then with interactive way input adjusted prompt, repeat this process, until the dose distribution data output meets the radiotherapy standard.The application can accept the adjusted prompt inputted in interactive way by user in the process of generating radiotherapy plan, so that interactive dose distribution generation tool can optimize the dose distribution data output, finally ensure that the dose distribution data meeting the standard is obtained.

Description

Technical Field

[0001] This invention relates to the medical field, and more particularly to a method, processing device, and electronic device for generating interactive radiology plans. Background Technology

[0002] Radiation therapy is a key method in cancer treatment, encompassing various techniques such as intensity-modulated radiation therapy (IMRT), volume-modulated arc therapy (VMAT), and stereotactic body radiation therapy (SBRT). These techniques aim to effectively destroy cancer cells by precisely guiding external radiation (such as photons, electrons, and protons) to the tumor site. However, radiation damage to surrounding healthy tissues is unavoidable during treatment. Therefore, during the treatment planning stage, it is crucial to ensure that the tumor receives an adequate dose while minimizing radiation exposure to surrounding healthy organs to guarantee patient safety. In clinical practice, developing a radiotherapy plan is a meticulous and complex process, requiring operators to continuously adjust dose limits and weights for different regions and optimize treatment engine parameters. Through repeated adjustments and optimizations, the goal is to achieve the best performance of the radiotherapy equipment for a specific patient, thus optimizing the radiotherapy plan.

[0003] Current radiation therapy plans typically use deep learning methods for dose distribution prediction, but their output depends on training data, and the generated dose distribution can only conform to the dose distribution preferences of the training input. However, operators in different hospitals have different preferences when formulating treatment plans. For example, different institutions may use different dose limits for normal organs and minimum dose limits for tumors. Dose limits for normal organs are also adjusted according to the specific patient's condition. For example, when normal organs are close to the tumor, the dose limits for these organs may be appropriately relaxed. Furthermore, when there are multiple adjacent normal organs, different operators may have differing opinions on the dose allocation that each organ can tolerate. Even deep learning models trained using multiple data sources struggle to generate a dose distribution that satisfies all operators at once. While treatment planning software can generate hundreds of alternative plans for operators to choose from, these require individual optimization of each plan, consuming significant computational resources.

[0004] In summary, even when clinical treatment requirements are met, treatment plans for the same patient may exhibit different dosage distributions depending on the operator. Therefore, it is necessary to develop a real-time, interactive tool for optimizing treatment plan dosage distribution. Summary of the Invention

[0005] To achieve the above objectives, this invention first provides an interactive radiotherapy plan generation method based on deep learning, the detailed technical solution of which is as follows:

[0006] An interactive radiotherapy plan generation method based on deep learning includes the following steps:

[0007] Completed radiotherapy plan data of patients are collected. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. The collected radiotherapy plan data is re-optimized using a plan optimization system. The dose limits are adjusted at equal intervals from a pre-set dose range to generate a dataset for training. The dataset is then divided into a training set, a validation set, and a test set.

[0008] A deep learning model is trained using a pre-defined dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from the medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data.

[0009] Input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, input the adjusted prompts interactively, repeat this process until the output dose distribution data meets the radiotherapy standards.

[0010] In some embodiments, the medical image data includes medical images, target area structure images, images of organs at risk, and images of a first auxiliary structure, wherein the first auxiliary structure is an auxiliary structure inherent in the medical image; the image encoder adopts a U-Net convolutional neural network structure, and the input channels of the image encoder include input channels for medical images, input channels for segmentation masks of target areas and organs at risk, and input channels for auxiliary structures.

[0011] In some embodiments, the prompts include the shape and dose limits of the target area structure, organ structure, and auxiliary structures; the prompts indicate that the encoder uses a keyword embedding-based encoding method or a numerical vector-based encoding method.

[0012] In some embodiments, the image decoder includes a feature fusion module, an upsampling module, and an output module, wherein: the feature fusion module is used to accept features input from the image encoder and features input from the cue encoder, and fuse the features input from the image encoder and the features input from the cue encoder through a splicing or attention mechanism to obtain a feature image; the upsampling module is used to restore the feature image to the original medical image size through upsampling or deconvolution operations; the output module is used to constrain the output value to be valid dose distribution data through an activation function.

[0013] In some embodiments, the adjusted prompt is entered interactively, including at least one of the following methods: entering the adjusted prompt in text form; or entering a second auxiliary structure by drawing, wherein the second auxiliary structure is a specific organ or a custom structure, and the second auxiliary structure includes at least the target area for dose increase or decrease.

[0014] In some embodiments, the image range of the second auxiliary structure is converted to "1-6", where "1-3" is the intensity of dose reduction in the target area, and the smaller the value, the greater the intensity of dose reduction; "4-6" is the intensity of dose increase in the target area, and the larger the value, the greater the intensity of dose increase.

[0015] In some embodiments, the planned optimization system is any one of the Photon Optimizer optimization engine, DMPO optimization engine, OpenREGGUI optimization engine, and Monte Carlo-based optimization engine.

[0016] A second aspect of the present invention provides an interactive radiotherapy planning device, comprising:

[0017] The dataset construction module is used to collect completed radiotherapy plan data of patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. The collected radiotherapy plan data is re-optimized using a plan optimization system. The dose limit is adjusted at equal intervals from a pre-set dose range to generate a dataset for training. The dataset is divided into a training set, a validation set, and a test set.

[0018] The model training module is used to train a deep learning model using a pre-defined dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from the medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the cue encoder are used as input to the image decoder, which outputs the corresponding dose distribution data.

[0019] Interactive Radiotherapy Plan Generation Module: This module takes the patient's medical images and initial prompts and inputs them into the interactive dose distribution generation tool. It evaluates the output dose distribution data and, if it does not meet the radiotherapy standards, inputs adjusted prompts interactively. This process is repeated until the output dose distribution data meets the radiotherapy standards.

[0020] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the interactive radiotherapy plan generation method of any of the first aspects of the present invention.

[0021] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the interactive radiotherapy plan generation method of any one of the first aspects of the present invention.

[0022] The aforementioned interactive radiotherapy plan generation method, apparatus, electronic device, and computer-readable storage medium, along with the interactive dose distribution generation tool, allow the user to interactively input adjusted prompts during the radiotherapy plan generation process. This enables the interactive dose distribution generation tool to repeatedly optimize the output dose distribution data until standard dose distribution data is obtained, thereby meeting the diverse treatment needs of different patients and improving the effectiveness of radiotherapy. Attached Figure Description

[0023] Figure 1 This is an execution flowchart of an interactive radiotherapy plan generation method according to one embodiment of the present invention;

[0024] Figure 2 This is an execution block diagram of the interactive radiotherapy plan generation method in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the model usage process in the interactive radiotherapy plan generation method of this invention embodiment;

[0026] Figure 4 This is a structural block diagram of the interactive radiotherapy plan generation device in an embodiment of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide indicative explanations of the invention. It should be observed that all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which the invention pertains.

[0028] The system architecture of the present invention and the solutions in the prior art will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It should be noted that the described embodiments are only for explanation and illustration of the present invention, and not all of the contents. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0029] Example 1

[0030] like Figure 1 As shown, the present invention provides an interactive radiotherapy plan generation method based on deep learning, comprising the following steps:

[0031] S1. Collect completed radiotherapy plan data for patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. Use the plan optimization system to re-optimize the collected radiotherapy plan data. By adjusting the dose limits at equal intervals from the pre-set dose range, a dataset for training is generated, and the dataset is divided into a training set, a validation set, and a test set.

[0032] S2. Train the deep learning model using the partitioned dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data.

[0033] S3. Input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, input the adjusted prompts interactively and repeat this process until the output dose distribution data meets the radiotherapy standards.

[0034] The reason for using the planning optimization system to re-optimize the collected radiotherapy planning data in step S1 above, and using the generated optimized data as the training dataset, is as follows:

[0035] Completed radiotherapy planning data may come from different hospitals, equipment, or clinical standards, resulting in inconsistencies in dose distribution and parameter settings. Furthermore, some radiotherapy planning data may suffer from insufficient dose coverage or inadequate protection of organs at risk due to differences in operational experience or technical limitations. Therefore, some radiotherapy planning data may not represent the optimal radiotherapy plan for the corresponding patient. Consequently, directly using collected radiotherapy planning data to train a deep learning model is unlikely to yield a satisfactory interactive dose distribution generation tool and may also lead to poor model generalization ability.

[0036] By re-optimizing the collected radiotherapy planning data using a planning optimization system, gradually increasing the dose limits from low to high, radiotherapy plans under different dose constraints can be generated, thus better simulating the diverse needs that may arise in clinical practice. This diversity helps the model learn the complex mapping relationship between dose distribution and parameters, enhancing its ability to adapt to different patients. Furthermore, re-optimizing the radiotherapy planning data through the planning optimization system can generate dose distributions that conform to current medical standards (such as target volume dose coverage and adequate protection of organs at risk), eliminating biases in the original data.

[0037] The core task of deep learning models is to learn the mapping between parameter inputs (such as dose limits) and dose distribution outputs. By systematically adjusting dose limits from low to high and re-optimizing them, richer and more accurate parameter-dose mapping data can be generated.

[0038] The core task of the interactive dose distribution generation tool obtained in this invention is to respond to user prompts for adjustments to parameters including radiation dose in an interactive manner, such as increasing the target dose or reducing the dose to organs. The planning optimization system simulates the interaction between clinicians and the system by generating dose distributions under different parameters, thereby enabling the model to learn dynamic adjustment capabilities.

[0039] The planning optimization system used in this invention is any one of the following: Photon Optimizer optimization engine, DMPO (Direct Machine Parameter Optimization) optimization engine, OpenREGGUI optimization engine, and Monte Carlo-based optimization engine.

[0040] In step S1, the dose limit is adjusted at equal intervals from a pre-set dose range. For example, the dose limit can be adjusted at equal intervals in a gradient manner from a pre-set dose range.

[0041] In step S3, the operator (doctor or physicist) can assess whether the output dose distribution data meets the standards for radiotherapy.

[0042] Example 2

[0043] like Figure 1 As shown, the present invention provides an interactive radiotherapy plan generation method based on deep learning, comprising the following steps:

[0044] S1. Collect completed radiotherapy plan data for patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. Use the plan optimization system to re-optimize the collected radiotherapy plan data. By adjusting the dose limits at equal intervals from the pre-set dose range, a dataset for training is generated, and the dataset is divided into a training set, a validation set, and a test set.

[0045] S2. Train the deep learning model using the partitioned dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data.

[0046] S3. Input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, input the adjusted prompts interactively and repeat this process until the output dose distribution data meets the radiotherapy standards.

[0047] The difference between this embodiment and Embodiment 1 is that:

[0048] In step S1, the medical image data includes medical images of the patient, target area structure images, images of organs at risk, and images of the first auxiliary structure acquired by the medical imaging device. The first auxiliary structure is an auxiliary structure that is built into the medical image.

[0049] Medical image data can be one or more medical images acquired by various existing medical imaging devices, such as computed tomography (CT) images and magnetic resonance imaging (MRI) images. Before administering radiotherapy, doctors or physicists will map the target area and the outline of organs at risk onto the medical images. In addition, they may also map one or more custom-designed structures to assist radiotherapy, such as dose drop zones near the target area. These can be considered as the first auxiliary structures inherent in the medical images. Images of the first auxiliary structures also constitute a type of medical image data.

[0050] The parameters and data used in radiotherapy are those set by the physician or physicist during the treatment process. These parameters include, for example, the shape and dose limits of the target area structure, organ structure, and primary and secondary structures. They may also include prescribed dose, treatment preferences, beam parameters, and target functions.

[0051] The prompts in steps S1 and S2 may include prompts about the shape and dose limits of the target area structure, organ structure and accessory structures.

[0052] As mentioned earlier, doctors or physicists will draw the outlines of the target area and organs at risk on medical images, and the parameter data also includes the shape and dose limits of the target area structure, organ structure, and primary auxiliary structures. Therefore, in step S2, when training the model, relevant prompts regarding the shape and dose limits of the target area structure, organ structure, and auxiliary structures can be extracted from the parameter data during medical images or treatment. Of course, during training, in order to improve the model's generalization ability, relevant prompts regarding the shape and dose limits of the target area structure, organ structure, and auxiliary structures are also randomly generated.

[0053] In step S2, both the image encoder and the cue encoder are based on a deep learning model architecture.

[0054] The image encoder can be based on the U-Net convolutional neural network architecture. Its input channels include input channels for medical images, input channels for segmentation masks of target structures and organs at risk, and input channels for auxiliary structural images. The pre-trained weights of the image encoder can be obtained from SAM-Med3D. Since the U-Net-based image encoder only accepts images of size 128×128×128, all training images may need to undergo linear interpolation, cropping, or padding to a size of 128×128×128 before being input into the image encoder. The cue encoder can be based on BERT network architecture, GPT network architecture, etc., and can employ keyword embedding-based encoding methods or numerical vector-based encoding methods.

[0055] In step S2, image decoding includes a feature fusion module, an upsampling module, and an output module, wherein:

[0056] The feature fusion module receives features from both the image encoder and the cue encoder, and fuses them using a concatenation or attention mechanism to obtain a feature image. The resulting feature image incorporates multimodal features, ensuring that the dose distribution is influenced by both image anatomy and artificial parameters.

[0057] The upsampling module is used to restore the feature image to the original medical image size through upsampling or deconvolution operations. For example, the upsampling module consists of multiple deconvolution layers or interpolation upsampling + convolution layers, which are used to progressively restore the spatial resolution of the image (from the low-dimensional feature map to the original medical image size).

[0058] The output module is used to constrain the output values ​​to be valid dose distribution data through an activation function. For example, a 1×1 convolutional layer is used to map the number of channels to 1, outputting a single-channel dose distribution map, where each pixel value represents the radiation dose at that location.

[0059] Example 3

[0060] like Figure 1 As shown, the present invention provides an interactive radiotherapy plan generation method based on deep learning, comprising the following steps:

[0061] S1. Collect completed radiotherapy plan data for patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. Use the plan optimization system to re-optimize the collected radiotherapy plan data. By adjusting the dose limits at equal intervals from the pre-set dose range, a dataset for training is generated, and the dataset is divided into a training set, a validation set, and a test set.

[0062] S2. Train the deep learning model using the partitioned dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data.

[0063] S3. Input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, input the adjusted prompts interactively and repeat this process until the output dose distribution data meets the radiotherapy standards.

[0064] The encoder is prompted to use either keyword embedding-based encoding or numerical vector-based encoding.

[0065] In step S3, the adjusted prompt is entered interactively, including at least one of the following methods:

[0066] Enter the adjusted prompt in text format.

[0067] The second auxiliary structure is input by outlining the contour. The second auxiliary structure is a specific organ or a custom structure. The second auxiliary structure includes at least the target area for dose increase or decrease.

[0068] The interactive dose distribution generation tool allows users to input adjustment parameters via text, such as "reduce vascular dose" or "increase spinal cord dose." Alternatively, users can input a second auxiliary structure by drawing a outline. For example, to reduce vascular dose, users can draw the outline of the blood vessels or any outline covering the blood vessels in the image displayed on the tool's interface. This outline is then input as the second auxiliary structure image into the tool, which converts it into a number from 1 to 6. Numbers 1-3 represent the intensity of dose reduction in the target area (smaller numbers indicate greater dose reduction), while numbers 4-6 represent the intensity of dose increase in the target area (larger numbers indicate greater dose increase).

[0069] Example 4

[0070] To enable those skilled in the art to more clearly understand, for example, the execution process of the deep learning-based interactive radiotherapy planning method of the present invention, this embodiment will be combined with... Figure 2 The execution process of the invention's deep learning-based interactive radiotherapy plan generation method is described by way of example.

[0071] like Figure 2 As shown, the interactive radiotherapy plan generation method in this embodiment includes three stages: training data generation, model training, and model application. Wherein:

[0072] During the training data generation phase, collected radiotherapy planning data, including medical image data (e.g., patient medical images, organ-at-risk contours, etc.) and parameter data during radiotherapy (e.g., PTV (target volume unit) dose limits, organ-at-risk dose limits, etc.), are input into the optimization engine. During optimization, optimization parameter data (e.g., adjusted and newly generated dose limit weights, randomly generated surrounding contours, dose optimization weights, etc.) are also provided to the optimization engine. After optimization, a training dataset is obtained. The training dataset is divided into a training set, a validation set, and a test set; optionally, the ratio of the data size of the training set, validation set, and test set is 7:1:2.

[0073] During the model training phase, the predefined training set, validation set, and test set are input into the deep learning model for training. Among them, medical image data (such as medical images of patients, outlines of organs at risk, etc.) are input into the image encoder, and cue data (such as cue for dosage adjustment direction and degree in the form of text, outline, or custom outline) are input into the cue encoder.

[0074] The output features of the image encoder and the output features of the cue encoder serve as inputs to the image decoder. The image decoder outputs dose distribution data, which is the dose distribution data predicted by the model.

[0075] The loss calculation function adjusts the model's weights by calculating the predicted dose distribution data and the dose distribution data generated by the optimization engine. After multiple rounds of training, a trained model is obtained, which serves as an interactive dose distribution generation tool. Optionally, the Adam optimizer is used for model training, with an initial learning rate of 0.0008. The model is trained for 200 cycles, with the learning rate reduced to one-tenth of the initial rate every 50 cycles. The loss function is the mean squared error (MSE).

[0076] During the model usage phase, the obtained interactive dose distribution generation tool is imported into the treatment planning system. Thus, the interactive interface of the treatment planning system serves as the input port for the interactive dose distribution generation tool. The patient's medical image data (e.g., medical images of the patient, outlines of organs at risk, etc.) and initial prompt data (e.g., prompts for dose adjustment direction and degree in text, outlines, or custom outlines) are input into the interactive dose distribution generation tool. The output dose distribution data is evaluated. If it does not meet radiotherapy standards, the prompt data is adjusted interactively. This process is repeated until the output dose distribution data meets radiotherapy standards.

[0077] like Figure 3 As shown, during the model usage phase, the parameter data can be adjusted interactively via text input, such as "reduce vascular dose," contour input, or custom region (contour) input.

[0078] Example 5

[0079] like Figure 4 As shown, an interactive radiotherapy plan generation device includes:

[0080] Dataset construction module 1 is used to collect radiotherapy plan data of completed patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. The collected radiotherapy plan data is re-optimized using a plan optimization system. The dose limit is adjusted at equal intervals from a pre-set dose range to generate a dataset for training. The dataset is then divided into a training set, a validation set, and a test set.

[0081] Model training module 2 is used to train the deep learning model using a pre-defined dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from the medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data.

[0082] The interactive radiotherapy plan generation module 3 is used to input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, then input the adjusted prompts interactively, repeating this process until the output dose distribution data meets the radiotherapy standards.

[0083] Since the processing procedures of each functional module of the interactive radiotherapy plan generation device in this embodiment are consistent with the processing procedures of the interactive radiotherapy plan generation method provided in the aforementioned method embodiment, this embodiment will not repeat the detailed processing procedures of each functional module of the interactive radiotherapy plan generation device. You can directly refer to the relevant content in the aforementioned method embodiment.

[0084] Example 6

[0085] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor and the memory are connected, for example, via a bus. When the processor executes the program, it implements the interactive radiotherapy planning method of any of the above embodiments.

[0086] Finally, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the interactive radiotherapy plan generation method provided in the preceding method embodiments.

[0087] The interactive radiotherapy plan generation method, apparatus, electronic device, and computer-readable storage medium provided by this invention allow the interactive dose distribution generation tool to accept user-inputted adjustments during the radiotherapy plan generation process. This enables the interactive dose distribution generation tool to repeatedly optimize the output dose distribution data until standard dose distribution data is obtained, thereby meeting the diverse treatment needs of different patients and improving the effectiveness of radiotherapy.

[0088] The present invention has been described above in sufficient detail and with certain specificities. Those skilled in the art should understand that the descriptions in the embodiments are merely exemplary, and all changes made without departing from the true spirit and scope of the invention should fall within the protection scope of the invention. The scope of protection claimed by the present invention is defined by the claims, and not by the above descriptions in the embodiments.

Claims

1. A deep learning-based interactive radiotherapy plan generation method, characterized in that, The The interactive radiotherapy plan generation method includes the following steps: Completed radiotherapy plan data of patients are collected. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. The collected radiotherapy plan data is re-optimized using a plan optimization system. The dose limits are adjusted at equal intervals from a pre-set dose range to generate a dataset for training. The dataset is then divided into a training set, a validation set, and a test set. A deep learning model is trained using a pre-defined dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from the medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the output features of the cue encoder are used as the input to the image decoder, and the image decoder outputs the corresponding dose distribution data. Input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, input the adjusted prompts interactively and repeat this process until the output dose distribution data meets the radiotherapy standards. The medical image data includes medical images, target area structure images, images of organs at risk, and images of a first auxiliary structure, wherein the first auxiliary structure is an auxiliary structure inherent in the medical image. The image encoder adopts a U-Net convolutional neural network structure. The input channels of the image encoder include input channels for medical images, input channels for target area structures and segmentation masks of organs at risk, and input channels for auxiliary structures.

2. The interactive radiotherapy plan generation method based on deep learning as described in claim 1, characterized in that, The prompts include the shape and dose limits of the target area structure, organ structure, and accessory structures; The prompt encoder uses either a keyword embedding-based encoding method or a numerical vector-based encoding method.

3. The interactive radiotherapy plan generation method based on deep learning as described in claim 1, characterized in that, The image decoder includes a feature fusion module, an upsampling module, and an output module, wherein: The feature fusion module is used to accept features input from the image encoder and features input from the prompt encoder, and fuse the features input from the image encoder and the prompt encoder through a splicing or attention mechanism to obtain a feature image. The upsampling module is used to restore the feature image to the original medical image size through upsampling or deconvolution operations; The output module is used to constrain the output value to be valid dose distribution data through an activation function.

4. The interactive radiotherapy plan generation method as described in claim 1, characterized in that, The interactive input of the adjusted prompts includes at least one of the following methods: Enter the adjusted prompt in text format; The second auxiliary structure is input by drawing, which is a specific organ or a custom structure, and the second auxiliary structure includes at least the target area for dose increase or decrease.

5. The interactive radiotherapy planning generation method as described in claim 4, characterized in that, The image range of the second auxiliary structure is converted to "1 to 6", where "1 to 3" is the intensity of dose reduction in the target area, and the smaller the value, the greater the intensity of dose reduction; "4 to 6" is the intensity of dose increase in the target area, and the larger the value, the greater the intensity of dose increase.

6. The interactive radiotherapy planning method as described in claim 1, characterized in that, The optimization system is any one of the following: Photon Optimizer optimization engine, DMPO optimization engine, OpenREGGUI optimization engine, and Monte Carlo-based optimization engine.

7. An interactive radiotherapy plan generation device, characterized in that, The interactive radiotherapy planning device includes: The dataset construction module is used to collect completed radiotherapy plan data of patients. The radiotherapy plan data includes at least the patient's medical image data and parameter data during the radiotherapy process. The collected radiotherapy plan data is re-optimized using a plan optimization system. The dose limit is adjusted at equal intervals from a pre-set dose range to generate a dataset for training. The dataset is divided into a training set, a validation set, and a test set. The model training module is used to train a deep learning model using a pre-defined dataset to obtain an interactive dose distribution generation tool. The deep learning model includes an image encoder, a cue encoder, and an image decoder. The input to the image encoder is medical image data, and the input to the cue encoder is cue extracted from the medical image data or parameter data, or a randomly generated cue. The output features of the image encoder and the cue encoder are used as input to the image decoder, which outputs the corresponding dose distribution data. Interactive Radiotherapy Plan Generation Module: This module is used to input the patient's medical images and initial prompts into the interactive dose distribution generation tool, evaluate the output dose distribution data, and if it does not meet the radiotherapy standards, it will input the adjusted prompts interactively and repeat this process until the output dose distribution data meets the radiotherapy standards. The medical image data includes medical images, target area structure images, images of organs at risk, and images of a first auxiliary structure, wherein the first auxiliary structure is an auxiliary structure inherent in the medical image. The image encoder adopts a U-Net convolutional neural network structure. The input channels of the image encoder include input channels for medical images, input channels for target area structures and segmentation masks of organs at risk, and input channels for auxiliary structures.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the interactive radiotherapy plan generation method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the interactive radiotherapy planning method according to any one of claims 1 to 7.

Images