Radiotherapy automatic sketching method and sketching agent

Abstract

The invention discloses a radiotherapy automatic sketching method and a sketching intelligent agent, and the method comprises the steps: obtaining image data sets from different parts, and dividing the image data sets into a source domain data set and a target domain data set; constructing a multi-part image sketching training model containing a shared encoder and a decoder, and performing supervised training by using the source domain data set to obtain an initial automatic sketching model; inputting sample features of the source domain data set and the target domain data set into a dynamic weighted full-connection network, calculating contribution weights of source domain data set samples to target domain data set samples, and generating weighted sample feature representations of the source domain data set and the target domain data set based on the contribution weights, performing domain confrontation training on the sample features of the weighted source domain data set and the target domain data set, and optimizing the initial sketching model to obtain an optimized automatic sketching model; and performing automatic sketching on the target domain image by using the optimized sketching model.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, specifically to an automatic radiotherapy delineation method and a delineation intelligent agent. Background Technology

[0002] Radiotherapy is a crucial treatment for tumors, and its success hinges on the precise delineation of the target area and surrounding organs at risk. In recent years, deep learning-based automated delineation technology has been widely applied in clinical practice to improve delineation efficiency and consistency. However, existing automated delineation methods still have significant limitations when dealing with images of multiple sites, cross-site locations, or irregular areas.

[0003] In existing technologies, researchers have proposed several solutions to address the challenges of multi-site delineation. For example, Chinese invention patent CN117151162A (publication number) proposes a "self-supervised and expert-gated method for incremental organ segmentation across anatomical regions." This method introduces multiple decoders and utilizes an expert-gated network to automatically select the correct decoder to perform segmentation of specific sites during the testing phase. This approach addresses the problem of multi-site segmentation to some extent.

[0004] However, such existing technical solutions have obvious drawbacks and shortcomings. First, due to the decoder selection strategy employed, when processing a target image, the model essentially only activates one or a few decoders that are most relevant to it, failing to fully and synergistically utilize the images and annotation information of all learned regions (such as the head, chest, and abdomen). This limits the model's knowledge transfer ability, making it unable to effectively fuse common features learned from multiple standard regions, thus affecting the accuracy and robustness of the delineation. Second, this selection mechanism has poor generalization ability for cross-regional and irregular region images that are not fully covered in the training data, because it relies on predefined decoders and is difficult to flexibly adapt to new scenes with unknown or complex anatomical structures.

[0005] Therefore, there is an urgent need in this field for a new automatic delineation technology that can effectively integrate knowledge from multiple source domains without relying on target domain annotation data, and has a strong ability to generalize across regions and irregular regions, in order to overcome the aforementioned problems and shortcomings of existing technologies.

[0006] In view of this, this invention patent is hereby proposed. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides an automatic delineation method and delineation intelligent agent for radiotherapy, specifically employing the following technical solution: An automatic radiotherapy delineation method, comprising: The data processing and partitioning steps involve acquiring image datasets from different parts of the body and partitioning them into source domain datasets and target domain datasets. The multi-part image delineation training steps involve constructing a multi-part image delineation training model that includes a shared encoder and decoder, and performing supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training steps involve inputting the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculating the contribution weights of the source domain dataset samples to the target domain dataset samples, generating weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performing domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizing the initial delineation model, and obtaining the optimized automatic delineation model. The target image delineation step uses an optimized delineation model to automatically delineate the target domain image.

[0008] As an optional embodiment of the present invention, an automatic radiotherapy delineation method of the present invention includes a data processing and segmentation step in which image datasets from different parts of the body are acquired and divided into different parts of the body datasets. The standard parts of the different parts of the body datasets are used as the source domain datasets, and the non-standard parts of the different parts of the body datasets are used as the target domain datasets. The standard parts of the body datasets are head data, chest data, or abdominal data, and the non-standard parts of the body datasets are hip or irregular parts of the body.

[0009] As an optional embodiment of the present invention, in an automatic radiotherapy delineation method of the present invention, the data processing and segmentation steps include: The datasets of different parts are processed uniformly. The standard part dataset with labeled data is used as the supervised training dataset Dsi, where s represents the source domain dataset and i indicates that the dataset is the i-th one in the source domain dataset; the target domain dataset is used as the test dataset Dt, where t represents the target domain.

[0010] As an optional embodiment of the present invention, in an automatic radiotherapy delineation method of the present invention, the multi-site image delineation training step includes: Based on the UNet network, a multi-part image delineation training model with a shared encoder and decoder was constructed. The sample xij in the supervised training dataset Dsi is input into the shared encoder, and the delineation training loss Lseg = Dice(yij,gij) between the output result yij of the decoder and the true annotation gij of the sample is calculated, where Dice represents the Dice loss and ij represents the j-th sample from the i-th source domain dataset. The obtained delineation training loss Lseg is backpropagated and the multi-part image delineation training model is updated. This process is repeated until the delineation training loss Lseg stabilizes and no longer decreases, thus obtaining the initial automatic delineation model.

[0011] As an optional embodiment of the present invention, an automatic delineation method for radiotherapy includes a weighted automatic delineation training step in which sample features of a source domain dataset and a target domain dataset are input into a dynamically weighted fully connected network, the contribution weights of the source domain dataset samples to the target domain dataset samples are calculated, and a weighted representation of the sample features of the source domain dataset and the target domain dataset is generated based on the contribution weights, comprising: Calculate the feature difference sij between the j-th sample feature xij of the i-th part in the source domain dataset and the j-th sample feature tj in the target domain dataset, where sij = xij - tj; The feature difference sij is input into the dynamic weighted fully connected network and mapped to generate the corresponding contribution weight w. The contribution weight w is used to weight the sample features of the source domain dataset and the target domain dataset respectively to obtain the weighted source domain feature representation zxij and the target domain feature representation ztj.

[0012] As an optional embodiment of the present invention, an automatic delineation method for radiotherapy includes a weighted automatic delineation training step in which domain adversarial training is performed on the sample features of the weighted source domain dataset and the target domain dataset to optimize the initial delineation model and obtain an optimized automatic delineation model, including: The weighted source domain feature representation zxij and the weighted target domain feature representation ztj are input together into a domain discriminator with a gradient reversal layer in front. Using the source domain feature representation zxij and the target domain feature representation ztj as classification labels, the domain discriminator performs domain classification, minimizing the classification loss Lce = CE(yz,ydz), where CE represents the cross-entropy loss, yz represents the domain label of the input feature, and ydz is the predicted domain label obtained by the classifier. During backpropagation of the domain classification loss Lce, the gradient inversion layer flips the gradients passing through it to achieve the adversarial training objective of maximizing the feature similarity between the source domain and the target domain, thereby optimizing the parameters of the shared encoder.

[0013] As an optional embodiment of the present invention, in an automatic radiotherapy delineation method of the present invention, the target image delineation step includes: Retain and invoke the shared encoder and decoder optimized by the weighted automatic delineation training step; The unlabeled images in the target domain dataset are input into the shared encoder, and the decoder outputs the final automatic delineation result.

[0014] As an optional embodiment of the present invention, in a radiotherapy automatic delineation method of the present invention, an evaluation and optimization step is performed after the target image delineation step, the evaluation and optimization step including: Receive user ratings and text feedback on the automatic drawing results; Key requirements are extracted based on the text evaluation, and a feature matrix is ​​generated based on the key requirements, wherein the weight of the part dataset corresponding to the key requirements is increased. The feature matrix is ​​normalized using the scores to generate guidance signals for model optimization; Based on the guidance signal, the contribution weight calculation in the weighted automatic delineation training step is adjusted, and the weighted automatic delineation training step and the target image delineation step are re-executed to generate an optimized new delineation result.

[0015] As an optional embodiment of the present invention, an automatic delineation method for radiotherapy includes receiving textual evaluations of the automatic delineation results from users in the evaluation and optimization step, and extracting key requirements based on the textual evaluations, including: Receive natural language text evaluations input by the user through the language model interface; The language model is used to perform semantic parsing on the natural language text evaluation, identify the target anatomical parts and / or delineation quality descriptions mentioned therein that need to be optimized, and output the parsing results in a structured manner as the key requirements.

[0016] This invention also provides an automated delineation intelligent agent for radiotherapy, comprising: The data processing and segmentation module acquires image datasets from different parts of the image and divides them into source domain datasets and target domain datasets. The multi-part image delineation training module constructs a multi-part image delineation training model based on the UNet network, and performs supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training module inputs the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculates the contribution weight of the source domain dataset samples to the target domain dataset samples, generates weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performs domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizes the initial delineation model, and obtains the optimized automatic delineation model. The target image delineation module uses an optimized delineation model to automatically delineate the target domain image.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This method achieves dynamic fusion of cross-regional knowledge: by inputting the features of source and target domain samples into a dynamically weighted fully connected network to calculate contribution weights, it can automatically and dynamically evaluate the contribution of data from different source regions to the target image. This overcomes the problem of rigid knowledge utilization in existing fixed models or simple model selection strategies when processing cross-regional images, enabling the model to intelligently fuse prior knowledge from multiple regions based on the specific features of the target image, thus laying a solid foundation for improving delineation accuracy.

[0018] This method enhances the model's generalization ability to the target domain: By employing the technique of "domain adversarial training on weighted sample features," it proactively reduces the distributional differences between labeled source domain data and unlabeled target domain data at the feature level. This mechanism enables the final optimized automatic delineation model to better adapt to the image characteristics of the target domain (such as cross-regional or irregular regions), significantly improving the model's generalization performance and robustness in unseen scenarios with scarce labeled data.

[0019] A complete and efficient automated workflow has been constructed: This method forms a complete solution through a series of closely linked steps: "data processing and partitioning → multi-part training → weighted and domain adversarial training → target delineation". This workflow ensures that the model fully utilizes source domain supervision information while specifically optimizing its performance on the target domain through specific adaptive modules (weighted and domain adversarial training). The final model does not require redesign or retraining for each new part, achieving efficient delineation capabilities of "one-time training, multi-part adaptation".

[0020] In summary, the automatic radiotherapy delineation method provided by this invention achieves flexible integration of knowledge through dynamic weighting and effective transfer of model capabilities through domain adversarial training. Ultimately, these combined efforts solve the core problems of poor adaptability and weak generalization ability in existing technologies when processing cross-site image delineation. Attached image description: Figure 1 A flowchart of an automatic delineation method for radiotherapy according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] In the description of this invention, it should be noted that the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. These terms are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] See Figure 1 As shown, an automatic radiotherapy delineation method according to this embodiment includes: The data processing and partitioning steps involve acquiring image datasets from different parts of the body and partitioning them into source domain datasets and target domain datasets. The multi-part image delineation training steps involve constructing a multi-part image delineation training model that includes a shared encoder and decoder, and performing supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training steps involve inputting the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculating the contribution weights of the source domain dataset samples to the target domain dataset samples, generating weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performing domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizing the initial delineation model, and obtaining the optimized automatic delineation model. The target image delineation step uses an optimized delineation model to automatically delineate the target domain image.

[0027] The automatic delineation method for radiotherapy provided in this embodiment has the following significant beneficial effects: This method achieves dynamic fusion of cross-regional knowledge: by inputting the features of source and target domain samples into a dynamically weighted fully connected network to calculate contribution weights, it can automatically and dynamically evaluate the contribution of data from different source regions to the target image. This overcomes the problem of rigid knowledge utilization in existing fixed models or simple model selection strategies when processing cross-regional images, enabling the model to intelligently fuse prior knowledge from multiple regions based on the specific features of the target image, thus laying a solid foundation for improving delineation accuracy.

[0028] This method enhances the model's generalization ability to the target domain: By employing the technique of "domain adversarial training on weighted sample features," it proactively reduces the distributional differences between labeled source domain data and unlabeled target domain data at the feature level. This mechanism enables the final optimized automatic delineation model to better adapt to the image characteristics of the target domain (such as cross-regional or irregular regions), significantly improving the model's generalization performance and robustness in unseen scenarios with scarce labeled data.

[0029] A complete and efficient automated workflow has been constructed: This method forms a complete solution through a series of closely linked steps: "data processing and partitioning → multi-part training → weighted and domain adversarial training → target delineation". This workflow ensures that the model fully utilizes source domain supervision information while specifically optimizing its performance on the target domain through specific adaptive modules (weighted and domain adversarial training). The final model does not require redesign or retraining for each new part, achieving efficient delineation capabilities of "one-time training, multi-part adaptation".

[0030] This embodiment provides an automatic radiotherapy delineation method that achieves flexible integration of knowledge through dynamic weighting and effective transfer of model capabilities through domain adversarial training. Ultimately, these combined efforts solve the core problems of poor adaptability and weak generalization ability in existing technologies when processing cross-site image delineation.

[0031] In this embodiment, an automatic radiotherapy delineation method is provided. In the data processing and segmentation step, image datasets from different parts of the body are acquired and divided into different parts of the body datasets. The standard parts of the different parts of the body datasets are used as the source domain datasets, and the non-standard parts of the different parts of the body datasets are used as the target domain datasets. The standard parts of the body datasets are head data, chest data, or abdominal data, and the non-standard parts of the body datasets are hip or irregular parts of the body.

[0032] In this embodiment of the automatic radiotherapy delineation method, the data processing and segmentation steps include: The datasets of different parts are processed uniformly. The standard part dataset with labeled data is used as the supervised training dataset Dsi, where s represents the source domain dataset and i indicates that the dataset is the i-th one in the source domain dataset; the target domain dataset is used as the test dataset Dt, where t represents the target domain.

[0033] In the automatic delineation method for radiotherapy in this embodiment, the data processing and segmentation steps are as follows: This approach clarifies the boundaries of knowledge transfer, providing a structural foundation for domain adaptation. By defining standard body part datasets (head, chest, abdomen) as the source domain and non-standard body part datasets (hip, irregular body parts) as the target domain, this method clearly defines the relationship between "known" and "unknown" at the data level. This establishes clear and reasonable application scenarios and objectives for subsequent weighted automatic delineation training and domain adversarial training, enabling the model optimization process to be targeted and specifically focused on solving the core problem of knowledge transfer from structurally clear known domains to structurally ambiguous or complex unknown domains.

[0034] This method leverages readily available labeled data to address the practical challenge of scarce annotations: it strategically utilizes relatively easy-to-obtain, high-quality labeled data from standard clinical sites such as the head, chest, and abdomen as the knowledge source (source domain), aiming to solve the problem of delineating non-standard sites (target domain) that are difficult to label or have no labels at all. This approach cleverly avoids the dilemma of directly collecting large amounts of labeled data for all rare or complex sites, greatly improving the method's practicality and cost-effectiveness, enabling its rapid application to a wider range of scenarios in clinical practice.

[0035] A model training and evaluation framework simulating real-world clinical challenges was constructed: by dividing the model into a "standard source domain" and a "non-standard target domain," this method simulates the generalization challenges the model would face in the real world during the training phase. This data partitioning method forces the model to learn features with cross-site generalization capabilities, rather than simply memorizing the anatomical structures of standard sites in the source domain, thereby directly improving the robustness and practicality of the final automatic delineation model in dealing with the diversity of real-world clinical situations.

[0036] Therefore, in the automatic radiotherapy delineation method of this embodiment, the data processing and partitioning steps start from the source of the problem definition and construct a training and evaluation framework for the entire invention through a strategic data partitioning. This framework can effectively stimulate and verify the cross-site generalization ability of the invention and is an important prerequisite and foundation for all subsequent technological innovations to be realized and effective.

[0037] In this embodiment of the automatic radiotherapy delineation method, the multi-site image delineation training step includes: Based on the UNet network, a multi-part image delineation training model with a shared encoder and decoder was constructed. The sample xij in the supervised training dataset Dsi is input into the shared encoder, and the delineation training loss Lseg = Dice(yij,gij) between the output result yij of the decoder and the true annotation gij of the sample is calculated, where Dice represents the Dice loss and ij represents the j-th sample from the i-th source domain dataset. The obtained delineation training loss Lseg is backpropagated and the multi-part image delineation training model is updated. This process is repeated until the delineation training loss Lseg stabilizes and no longer decreases, thus obtaining the initial automatic delineation model.

[0038] For an image sample in the target domain dataset, the annotation data from different parts of the source domain dataset may potentially contribute to the delineation of that sample, and their contributions are different. To address this, an automatic delineation method for radiotherapy in this embodiment introduces a weighted automatic delineation training step, which adjusts the weights of each image part according to the similarity contribution of the source domain dataset to the target domain dataset, further optimizing the multi-part image delineation training step.

[0039] In an automatic delineation method for radiotherapy according to this embodiment, the weighted automatic delineation training step involves inputting sample features of the source domain dataset and the target domain dataset into a dynamically weighted fully connected network, calculating the contribution weights of the source domain dataset samples to the target domain dataset samples, and generating weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, including: Calculate the feature difference sij between the j-th sample feature xij of the i-th part in the source domain dataset and the j-th sample feature tj in the target domain dataset, where sij = xij - tj; The feature difference sij is input into the dynamic weighted fully connected network and mapped to generate the corresponding contribution weight w. The contribution weight w is used to weight the sample features of the source domain dataset and the target domain dataset respectively to obtain the weighted source domain feature representation zxij and the target domain feature representation ztj.

[0040] This embodiment introduces a dynamically weighted fully connected network, using the similarity between the features of samples xij and tj after passing through the shared encoder as the input of the network, where xij represents the j-th sample feature from the i-th part dataset in the source domain dataset, and tj represents the j-th sample feature in the target dataset.

[0041] In this embodiment, the feature difference sij between the j-th sample feature xij of the i-th part in the source domain dataset and the j-th sample feature tj in the target domain dataset is calculated as the similarity between the two. When a certain dimension of sij is smaller, it indicates that the dimension is the feature most similar to the sample in the source domain and the sample in the target domain, that is, the feature with higher potential contribution of the sample in the source domain.

[0042] In an automatic delineation method for radiotherapy according to this embodiment, the weighted automatic delineation training step involves performing domain adversarial training on the sample features of the weighted source domain dataset and the target domain dataset to optimize the initial delineation model and obtain an optimized automatic delineation model, including: The weighted source domain feature representation zxij and the weighted target domain feature representation ztj are input together into a domain discriminator with a gradient reversal layer in front. Using the source domain feature representation zxij and the target domain feature representation ztj as classification labels, the domain discriminator performs domain classification, minimizing the classification loss Lce = CE(yz,ydz), where CE represents the cross-entropy loss, yz represents the domain label of the input feature, and ydz is the predicted domain label obtained by the classifier. During backpropagation of the domain classification loss Lce, the gradient inversion layer flips the gradients passing through it to achieve the adversarial training objective of maximizing the feature similarity between the source domain and the target domain, thereby optimizing the parameters of the shared encoder.

[0043] The purpose of the weighted automatic delineation training step described in this embodiment is to narrow the distance between the sample features of the source domain dataset and the target domain dataset, so that the shared encoder can confuse the two domains and transfer the knowledge learned from the trained source domain dataset to the target domain to a greater extent, thereby improving the delineation accuracy and generalization of the target domain dataset.

[0044] This embodiment obtains an optimized automatic delineation model through the weighted automatic delineation training step, which can acquire training knowledge from image delineation information from different parts of the image, thereby guiding the automatic delineation of the target domain image.

[0045] In this embodiment of an automatic radiotherapy delineation method, the target image delineation step includes: Retain and invoke the shared encoder and decoder optimized by the weighted automatic delineation training step; The unlabeled images in the target domain dataset are input into the shared encoder, and the decoder outputs the final automatic delineation result.

[0046] The multi-part image delineation training model and the automatic delineation model will be trained together. The overall loss function can be expressed as Ltotal = Lseg + Lce. The first objective of the overall training is to minimize the loss of the source domain dataset, and the second objective is to reduce the loss of a small portion of the labeled data in the target domain simultaneously without participating in backpropagation.

[0047] In the target image delineation stage, only the shared encoder and decoder modules of the model are retained. Images from the target dataset are input into the shared encoder and the results are obtained from the decoder. Ultimately, the trained model can improve its ability to delineate images across multiple locations or irregular locations, building upon its ability to delineate multi-location images. Compared to existing encoder selection strategies, this method does not require labeled data in the target image dataset for model training, exhibits good generalization, and can better integrate knowledge learned from all datasets, thus improving delineation accuracy.

[0048] In an automatic radiotherapy image delineation method according to this embodiment, an evaluation and optimization step is performed after the target image delineation step. The evaluation and optimization step includes: Receive user ratings and text feedback on the automatic drawing results; Key requirements are extracted based on the text evaluation, and a feature matrix is ​​generated based on the key requirements, wherein the weight of the part dataset corresponding to the key requirements is increased. The feature matrix is ​​normalized using the scores to generate guidance signals for model optimization; Based on the guidance signal, the contribution weight calculation in the weighted automatic delineation training step is adjusted, and the weighted automatic delineation training step and the target image delineation step are re-executed to generate an optimized new delineation result.

[0049] In an automatic radiotherapy delineation method according to this embodiment, the evaluation and optimization step receives textual feedback from the user on the automatic delineation result and extracts key requirements based on the textual feedback, including: Receive natural language text evaluations input by the user through the language model interface; The language model is used to perform semantic parsing on the natural language text evaluation, identify the target anatomical parts and / or delineation quality descriptions mentioned therein that need to be optimized, and output the parsing results in a structured manner as the key requirements.

[0050] Specifically, the delineation agent in this method is a combination of a trained language model and an automatic delineation model. The language model serves as the interface for communication between the agent and the user regarding delineation requirements, while the automatic delineation model acts as the agent's delineation optimization module. After training and fine-tuning with user-annotated information, the delineation agent can communicate delineation information with the user. The user can input their delineation requirements for the target image into the delineation agent, which will optimize the model until a delineation result meeting the user's requirements is obtained. During this process, the user does not need to manually adjust the delineation result, improving delineation efficiency. In actual use, the user first sends the target image to be delineated to the model, which outputs the delineation result. After receiving the delineation result, the user can score the delineation and provide specific feedback to the agent. Upon receiving the user's score and evaluation, the agent adjusts the delineation model. The steps are as follows: (1) The user's evaluation of the target image delineation result is: 90 points, the delineation effect of chest organs is not ideal.

[0051] The agent extracts the score and the characteristic of "enhancing the chest feature delineation", generates a score feature, and uses this score feature as the weight coefficient of the weighted automatic delineation module to update the weighted feature, thereby increasing the potential contribution of the "chest" dataset features to the target feature.

[0052] (2) After the model is re-optimized and updated, it provides users with new drawing results. Users can choose to accept the new drawing results or score and iterate the results again.

[0053] (3) The above score feature generation method is to generate a feature matrix for the extracted effects of each part. For example, if you want to "enhance the outline of the chest feature", the value of the chest dataset in the feature matrix will be increased accordingly. Secondly, the feature matrix is ​​normalized using user scores. The lower the score, the higher the variance of each feature value after normalization, which encourages a greater increase in the weight of the required part and the training effect.

[0054] This embodiment also provides an automated delineation intelligent agent for radiotherapy, including: The data processing and segmentation module acquires image datasets from different parts of the image and divides them into source domain datasets and target domain datasets. The multi-part image delineation training module constructs a multi-part image delineation training model based on the UNet network, and performs supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training module inputs the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculates the contribution weight of the source domain dataset samples to the target domain dataset samples, generates weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performs domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizes the initial delineation model, and obtains the optimized automatic delineation model. The target image delineation module uses an optimized delineation model to automatically delineate the target domain image.

[0055] This embodiment provides an automatic delineation intelligent agent for radiotherapy. The model is adjusted and optimized according to the user's needs, so that the model can combine information on different standard parts and their delineation to improve the automatic delineation accuracy of various cross-part or structurally complex images. At the same time, the target image dataset does not require labeled data to participate in training, which effectively improves the generalization of automatic delineation of different images.

[0056] This embodiment also provides an automated delineation intelligent agent for radiotherapy, and further includes an evaluation and optimization module, wherein the evaluation and optimization steps include: Receive user ratings and text feedback on the automatic drawing results; Key requirements are extracted based on the text evaluation, and a feature matrix is ​​generated based on the key requirements, wherein the weight of the part dataset corresponding to the key requirements is increased. The feature matrix is ​​normalized using the scores to generate guidance signals for model optimization; Based on the guidance signal, the contribution weight calculation in the weighted automatic delineation training step is adjusted, and the weighted automatic delineation training step and the target image delineation step are re-executed to generate an optimized new delineation result.

[0057] This embodiment also provides a computer-readable storage medium storing a computer-executable program, which, when executed, implements the automatic delineation method for radiotherapy as described above.

[0058] The computer-readable storage medium described in this embodiment may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0059] This embodiment also provides an electronic device, including a processor and a memory, wherein the memory is used to store a computer-executable program, and when the computer program is executed by the processor, the processor executes the automatic delineation method for radiotherapy.

[0060] The electronic device is manifested in the form of a general-purpose computing device. It may contain one or more processors that work collaboratively. This invention also does not preclude distributed processing, meaning that processors may be distributed across different physical devices. The electronic device of this invention is not limited to a single entity, but may also be the sum of multiple physical devices.

[0061] The memory stores a computer-executable program, typically machine-readable code. The computer-readable program can be executed by the processor to enable the electronic device to perform the method of the present invention, or at least some steps of the method.

[0062] The memory includes volatile memory, such as random access memory (RAM) and / or cache memory, and may also be non-volatile memory, such as read-only memory (ROM).

[0063] It should be understood that the electronic device of the present invention may also include elements or components not shown in the examples above. For example, some electronic devices also include display units such as a display screen, and some electronic devices also include human-computer interaction elements such as buttons and keyboards. Any electronic device capable of executing a computer-readable program in its memory to implement the method of the present invention or at least some steps of the method can be considered as an electronic device covered by the present invention.

[0064] From the above description of the embodiments, those skilled in the art will readily understand that the present invention can be implemented by hardware capable of executing specific computer programs, such as the system of the present invention, and the electronic processing unit, server, client, mobile phone, control unit, processor, etc. included in the system. The present invention can also be implemented by computer software executing the methods of the present invention, for example, by control software executed by a microprocessor, electronic control unit, client, server, etc. However, it should be noted that the computer software executing the methods of the present invention is not limited to execution in one or a specific set of hardware entities; it can also be implemented in a distributed manner by unspecified hardware. For computer software, the software product can be stored in a computer-readable storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) or distributed across a network, as long as it enables electronic devices to execute the methods according to the present invention.

[0065] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. An automatic delineation method for radiotherapy, characterized in that, include: The data processing and partitioning steps involve acquiring image datasets from different parts of the body and partitioning them into source domain datasets and target domain datasets. The multi-part image delineation training steps involve constructing a multi-part image delineation training model that includes a shared encoder and decoder, and performing supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training steps involve inputting the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculating the contribution weights of the source domain dataset samples to the target domain dataset samples, generating weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performing domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizing the initial delineation model, and obtaining the optimized automatic delineation model. The target image delineation step uses an optimized delineation model to automatically delineate the target domain image.

2. The automatic delineation method for radiotherapy according to claim 1, characterized in that, In the data processing and segmentation steps, image datasets from different body parts are acquired and divided into different body part datasets. The standard body part datasets in the different body part datasets are used as the source domain datasets, and the non-standard body part datasets in the different body part datasets are used as the target domain datasets. The standard body part data are head data, chest data, or abdominal data, and the non-standard body part data are hip or irregular body part data.

3. The automatic delineation method for radiotherapy according to claim 2, characterized in that, The data processing and partitioning steps include: The datasets of different parts are processed uniformly. The standard part dataset with labeled data is used as the supervised training dataset Dsi, where s represents the source domain dataset and i indicates that the dataset is the i-th one in the source domain dataset; the target domain dataset is used as the test dataset Dt, where t represents the target domain.

4. The automatic delineation method for radiotherapy according to claim 3, characterized in that, The multi-part image delineation training steps include: Based on the UNet network, a multi-part image delineation training model with a shared encoder and decoder was constructed. The sample xij in the supervised training dataset Dsi is input into the shared encoder, and the delineation training loss Lseg = Dice(yij,gij) between the output result yij of the decoder and the true annotation gij of the sample is calculated, where Dice represents the Dice loss and ij represents the j-th sample from the i-th source domain dataset. The obtained delineation training loss Lseg is backpropagated and the multi-part image delineation training model is updated. This process is repeated until the delineation training loss Lseg stabilizes and no longer decreases, thus obtaining the initial automatic delineation model.

5. The automatic delineation method for radiotherapy according to claim 4, characterized in that, The weighted automatic delineation training step involves inputting the sample features of the source domain dataset and the target domain dataset into a dynamically weighted fully connected network, calculating the contribution weights of the source domain dataset samples to the target domain dataset samples, and generating weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, including: Calculate the feature difference sij between the j-th sample feature xij of the i-th part in the source domain dataset and the j-th sample feature tj in the target domain dataset, where sij = xij - tj; The feature difference sij is input into the dynamic weighted fully connected network and mapped to generate the corresponding contribution weight w. The contribution weight w is used to weight the sample features of the source domain dataset and the target domain dataset respectively to obtain the weighted source domain feature representation zxij and the target domain feature representation ztj.

6. The automatic delineation method for radiotherapy according to claim 5, characterized in that, The weighted automatic delineation training step involves performing domain adversarial training on the sample features of the weighted source domain dataset and the target domain dataset to optimize the initial delineation model and obtain the optimized automatic delineation model, including: The weighted source domain feature representation zxij and the weighted target domain feature representation ztj are input together into a domain discriminator with a gradient reversal layer in front. Using the source domain feature representation zxij and the target domain feature representation ztj as classification labels, the domain discriminator performs domain classification, minimizing the classification loss Lce = CE(yz,ydz), where CE represents the cross-entropy loss, yz represents the domain label of the input feature, and ydz is the predicted domain label obtained by the classifier. During backpropagation of the domain classification loss Lce, the gradient inversion layer flips the gradients passing through it to achieve the adversarial training objective of maximizing the feature similarity between the source domain and the target domain, thereby optimizing the parameters of the shared encoder.

7. The automatic delineation method for radiotherapy according to claim 6, characterized in that, The target image delineation steps include: Retain and invoke the shared encoder and decoder optimized by the weighted automatic delineation training step; The unlabeled images in the target domain dataset are input into the shared encoder, and the decoder outputs the final automatic delineation result.

8. The automatic delineation method for radiotherapy according to claim 1, characterized in that, Following the target image delineation step, an evaluation and optimization step is performed, which includes: Receive user ratings and text feedback on the automatic drawing results; Key requirements are extracted based on the text evaluation, and a feature matrix is ​​generated based on the key requirements, wherein the weight of the part dataset corresponding to the key requirements is increased. The feature matrix is ​​normalized using the scores to generate guidance signals for model optimization; Based on the guidance signal, the contribution weight calculation in the weighted automatic delineation training step is adjusted, and the weighted automatic delineation training step and the target image delineation step are re-executed to generate an optimized new delineation result.

9. The automatic delineation method for radiotherapy according to claim 8, characterized in that, The evaluation and optimization step receives user text feedback on the automatic drawing results and extracts key requirements based on the text feedback, including: Receive natural language text evaluations input by the user through the language model interface; The language model is used to perform semantic parsing on the natural language text evaluation, identify the target anatomical parts and / or delineation quality descriptions mentioned therein that need to be optimized, and output the parsing results in a structured manner as the key requirements.

10. A radiotherapy automatic delineation intelligent agent, characterized in that, include: The data processing and segmentation module acquires image datasets from different parts of the image and divides them into source domain datasets and target domain datasets. The multi-part image delineation training module constructs a multi-part image delineation training model based on the UNet network, and performs supervised training using the source domain dataset to obtain an initial automatic delineation model. The weighted automatic delineation training module inputs the sample features of the source domain dataset and the target domain dataset into a dynamic weighted fully connected network, calculates the contribution weight of the source domain dataset samples to the target domain dataset samples, generates weighted sample feature representations of the source domain dataset and the target domain dataset based on the contribution weights, performs domain adversarial training on the weighted sample features of the source domain dataset and the target domain dataset, optimizes the initial delineation model, and obtains the optimized automatic delineation model. The target image delineation module uses an optimized delineation model to automatically delineate the target domain image.

Images