Radiotherapy plan dose distribution prediction method based on domain adaptive transfer learning

Abstract

The invention relates to the technical field of transfer learning, and discloses a radiotherapy plan dose distribution prediction method based on domain self-adaptive transfer learning, which comprises the following steps: defining a source domain medical institution and a target domain medical institution, collecting radiotherapy plan data of the source domain medical institution, extracting radiotherapy plan characteristics of the radiotherapy plan data, and predicting the radiotherapy plan dose distribution of the target domain medical institution; training an initial dose distribution analysis model of the source domain medical institution; constructing a federated learning network of the source domain medical institution and the target domain medical institution, and outputting a dose distribution analysis map of the initial dose distribution analysis model; identifying a domain difference between the source domain medical institution and the target domain medical institution; determining an optimization target of the initial dose distribution analysis model, and performing domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; a target dose distribution for the patient in the target domain medical institution is generated. According to the invention, the efficiency and accuracy of radiotherapy plan dose distribution prediction can be improved.

Description

Technical Field

[0001] The present invention relates to a method for predicting radiation therapy plan dose distribution based on domain adaptive transfer learning, and belongs to the technical field of transfer learning. Background Art

[0002] The radiation dose distribution in a radiotherapy plan is a detailed three-dimensional representation of the radiation dose distribution within the patient's body (specifically the tumor area and surrounding tissues) calculated and generated based on the tumor's location, size, and shape, as well as the sensitivity of surrounding normal organs, during the development of a radiotherapy plan. This prediction allows for a quick and accurate estimation of the ultimate radiation dose distribution within the patient before the treatment plan is officially implemented, enabling more precise, safer, and personalized tumor radiotherapy, improving patient prognosis and quality of life.

[0003] Traditional radiotherapy plan dose distribution prediction is mainly based on physics-based calculations, which repeatedly optimize beam parameters (angle, energy, intensity, etc.) to achieve the ideal dose distribution. Since this method requires repeated optimization and calculation, the prediction process is complex and time-consuming, which limits the efficiency of quickly trying different plans and optimizing them. Summary of the Invention

[0004] The present invention provides a method for predicting the dose distribution of radiotherapy plans based on domain adaptive transfer learning, the main purpose of which is to improve the efficiency and accuracy of radiotherapy plan dose distribution prediction.

[0005] To achieve the above objectives, the present invention provides a method for predicting radiation therapy dose distribution based on domain adaptive transfer learning, comprising: Identify a source domain medical institution and a target domain medical institution, collect radiotherapy plan data of the source domain medical institution, and extract radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location, and dose distribution, so as to train an initial dose distribution analysis model for the source domain medical institution; Constructing a federated learning network between the source domain medical institution and the target domain medical institution, distributing the initial dose distribution analysis model to the target domain medical institution based on the federated learning network, and outputting a dose distribution analysis map of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution; identifying domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data; Based on the domain differences, clarifying optimization objectives of the initial dose distribution analysis model, wherein the optimization objectives include: dose analysis loss, domain adversarial loss, and feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; A target dose distribution for patients in the target domain medical institution is generated according to the local dose distribution analysis model.

[0006] Optionally, extracting radiotherapy plan features from the radiotherapy plan data includes: Unifying the format of the radiotherapy plan data to obtain standardized radiotherapy plan data; De-noising the standardized radiotherapy plan data to obtain de-noised radiotherapy plan data; performing coordinate system alignment on the denoised radiotherapy plan data to obtain aligned radiotherapy plan data; identifying key structures of the aligned radiotherapy planning data, and calculating geometric features and anatomical relationship features of the key structures; determining the anatomical structure features and tumor position of the aligned radiotherapy planning data according to the geometric features and the anatomical relationship features; calculating target dose characteristics and organ-at-risk dose characteristics at the tumor location based on the aligned radiotherapy plan data; determining a dose distribution at the tumor location based on the target volume dose characteristics and the organ-at-risk dose characteristics; The radiotherapy planning features of the radiotherapy planning data are determined according to the dose distribution, the anatomical structure features and the tumor position.

[0007] Optionally, the training of the initial dose distribution analysis model of the source domain medical institution includes: Dividing the radiotherapy plan features corresponding to the source domain medical institution into a training set and a validation set; Generate a model framework of the source domain medical institution; Based on the training set, the model framework is trained to obtain a training analysis model; Calculating the mean square error of the training analysis model based on the validation set; When the mean square error is greater than a preset error threshold, after adjusting the model learning parameters of the training analysis model, returning to the step of training the model framework based on the training set to obtain a training analysis model; When the mean square error is less than a preset error threshold, the training analysis model is used as the initial dose distribution analysis model of the source domain medical institution.

[0008] Optionally, the constructing of a federated learning network of the source domain medical institution and the target domain medical institution includes: Clarifying the source domain client node of the source domain medical institution and the target domain client node of the target domain medical institution; Establishing a communication protocol and a dynamic encryption protocol between the source domain client node and the target domain client node; Building a central server for the source domain client node and the target domain client node based on the communication protocol and the dynamic encryption protocol; uniformly configuring the client environments of the source domain client node and the target domain client node; Defining a model training loop rule for the source domain client node, the target domain client node, and the central server according to the client environment; According to the model training cycle rules, a federated learning network of the source domain medical institution and the target domain medical institution is constructed.

[0009] Optionally, outputting the dose distribution analysis atlas of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution includes: Collecting local radiotherapy plan data of the target domain medical institution; extracting key information of the local radiotherapy planning data; Converting the key information into input format information; Inputting the input format information into the initial dose distribution analysis model to obtain dose distribution data; A dose distribution analysis map of the patient corresponding to the target domain medical institution is generated based on the dose distribution data.

[0010] Optionally, the identifying domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data includes: extracting the model analysis dose of the dose distribution analysis atlas and the local plan dose of the local radiotherapy plan data; uniformly identifying the anatomical structure of the dose distribution analysis atlas and the local radiotherapy planning data; The domain difference between the source domain medical institution and the target domain medical institution is calculated according to the anatomical structure, the model analysis dose, and the local planned dose.

[0011] Optionally, clarifying an optimization target of the initial dose distribution analysis model based on the domain difference includes: analyzing difference features of the domain differences, wherein the difference features include dose volume histogram differences and dose distribution differences; determining an anatomical structure weight, a dose level weight, and a voxel importance weight of the initial dose distribution analysis model according to the difference characteristics; determining a dose analysis loss of the initial dose distribution analysis model based on the anatomical structure weight, the dose level weight, and the voxel importance weight; constructing a domain classifier of the initial dose distribution analysis model; extracting input features of the initial dose distribution analysis model, inputting the input features into the domain classifier, and obtaining a classifier analysis value; Calculating a domain adversarial loss of the initial dose distribution analysis model according to the classifier analysis value; extracting source domain data features and target domain data features of the initial dose distribution analysis model respectively; respectively calculating a source domain kernel matrix between the source domain data features and a target domain kernel matrix between the target domain data features; Calculating a kernel matrix between the source domain data features and the target domain data features; determining a feature alignment loss of the initial dose distribution analysis model according to the source domain kernel matrix, the target domain kernel matrix, and the kernel matrix; An optimization target of the initial dose distribution analysis model is determined according to the feature alignment loss, the dose analysis loss, and the domain confrontation loss.

[0012] Optionally, performing domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model includes: According to the optimization target of the initial dose distribution analysis model, the initial dose distribution analysis model is trained using preset training data to obtain a training dose distribution analysis model; Calculating the main task output and domain classification output of the trained dose distribution analysis model; Calculating a model main task loss and a model domain adversarial loss of the training dose distribution analysis model according to the main task output and the domain classification output; When the main task loss of the model is greater than a preset main task loss threshold, after adjusting the model parameters of the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a training dose distribution analysis model; When the model domain adversarial loss is less than a preset domain adversarial loss threshold, after adjusting the classifier parameters of the domain classifier corresponding to the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a trained dose distribution analysis model; When the model main task loss is less than a preset main task loss threshold and the model domain adversarial loss is greater than a preset domain adversarial loss threshold, the training dose distribution analysis model is used as the training dose distribution analysis model.

[0013] Optionally, generating a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model includes: Inputting the patient data corresponding to the patient into the local dose distribution analysis model, and processing the patient data through the convolution layer of the local dose distribution analysis model to obtain high-level features and low-level detail features; The high-level features and the low-level detail features are subjected to feature fusion through a feature fusion layer of the local dose distribution analysis model to obtain fused features; Calculating a dose distribution analysis map of the patient through a dose calculation layer of the local dose distribution analysis model according to the fusion feature; According to the dose distribution analysis graph, the target dose distribution of the patient is output through the output layer of the local dose distribution analysis model.

[0014] In order to solve the above problems, the present invention also provides a radiotherapy plan dose distribution prediction system based on domain adaptive transfer learning, the system comprising: A source domain model training module is used to identify a source domain medical institution and a target domain medical institution, collect radiotherapy plan data of the source domain medical institution, and extract radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location, and dose distribution, so as to train an initial dose distribution analysis model for the source domain medical institution; A local data prediction module is used to construct a federated learning network between the source domain medical institution and the target domain medical institution, distribute the initial dose distribution analysis model to the target domain medical institution based on the federated learning network, and output a dose distribution analysis map of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution; a domain difference analysis module, configured to identify domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data; a local model training module, configured to clarify an optimization target of the initial dose distribution analysis model based on the domain difference, wherein the optimization target includes: dose analysis loss, domain adversarial loss, and feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; The target dose distribution module is used to generate a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model.

[0015] Compared with the problems described in the background technology, the embodiment of the present invention can directly generate a higher quality initial plan by training the initial dose distribution analysis model of the source domain medical institution, reduce the number and magnitude of subsequent optimization adjustments, and shorten the overall planning time; optionally, the embodiment of the present invention can achieve cross-institutional data collaboration under the premise of strictly protecting data privacy by constructing a federated learning network of the source domain medical institution and the target domain medical institution, thereby improving the generalization ability, applicability and overall performance of the initial dose distribution analysis model; the embodiment of the present invention outputs the dose distribution analysis map of the initial dose distribution analysis model through the local radiotherapy plan data of the target domain medical institution, which can preliminarily judge the applicability and accuracy of the initial model on the target domain data, and provide a basis for whether fine-tuning is needed and how to fine-tune it later; the embodiment of the present invention combines the dose distribution analysis model with the target domain medical institution to obtain the dose distribution analysis map of the initial dose distribution analysis model. By analyzing the atlas and the local radiotherapy plan data, identifying the domain differences between the source domain medical institution and the target domain medical institution, it is possible to intuitively and quantitatively identify the systematic differences in dose distribution between the two institutions in similar cases; the embodiment of the present invention performs domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model, which can enable the local dose distribution analysis to adapt to a specific local environment, thereby overcoming the challenges brought by inconsistent data distribution while maintaining or even improving performance, and achieving more accurate, more efficient and more practical localized applications; finally, the embodiment of the present invention generates the target dose distribution of patients in the target domain medical institution according to the local dose distribution analysis model, which can enhance the adaptability and robustness of the model in different medical institution environments, thereby generating dose distributions for target domain patients more reliably and accurately. Therefore, the radiotherapy plan dose distribution prediction method based on domain adaptive transfer learning provided by the embodiment of the present invention can improve the efficiency and accuracy of radiotherapy plan dose distribution prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic flow chart of a method for predicting dose distribution in radiotherapy plans based on domain adaptive transfer learning according to an embodiment of the present invention; Figure 2A schematic diagram of a module for implementing the method for predicting dose distribution of radiotherapy plans based on domain adaptive transfer learning provided in one embodiment of the present invention.

[0017] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] The embodiment of the present application provides a method for predicting the dose distribution of a radiotherapy plan based on domain adaptive transfer learning. The execution subject of the method for predicting the dose distribution of a radiotherapy plan based on domain adaptive transfer learning includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for predicting the dose distribution of a radiotherapy plan based on domain adaptive transfer learning can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0020] Example 1: Reference Figure 1 FIG2 is a flow chart of a method for predicting dose distribution of radiotherapy plans based on domain adaptive transfer learning according to an embodiment of the present invention. In this embodiment, the method for predicting dose distribution of radiotherapy plans based on domain adaptive transfer learning includes: S1. Identify the source domain medical institution and the target domain medical institution, collect the radiotherapy plan data of the source domain medical institution, and extract the radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location and dose distribution, so as to train the initial dose distribution analysis model of the source domain medical institution.

[0021] By specifying source and target domain medical institutions, this embodiment of the present invention allows them to leverage the vast amount of data in the source domain to rapidly construct an initial model, significantly reducing the time and data required to train the target domain from scratch. Source domain medical institutions are those that provide the raw data used to train the initial model. Target domain medical institutions are those that deploy trained dose prediction models (typically trained based on source domain data and potentially adjusted) to improve the efficiency and quality of their own radiotherapy plans.

[0022] By collecting radiotherapy planning data from the source medical institution, the present invention can provide a large amount of training data for subsequent model training. The radiotherapy planning data refers to a collection of various information directly related to the treatment plan generated during the development and implementation of radiotherapy for patients in a medical institution.

[0023] Optionally, the radiotherapy planning data of the source medical institution can be obtained by exporting it from a radiotherapy planning system.

[0024] The embodiments of the present invention can achieve structuring and quantification of the plan data by extracting the radiotherapy plan features from the radiotherapy plan data, thereby supporting the training and application of machine learning models. The radiotherapy plan features refer to information extracted from the original radiotherapy plan data that can quantitatively describe the key attributes or characteristics of the plan. The anatomical structure features refer to the morphology, location, and range of various organs, tissues, bones, tumors, etc. inside the patient's body as presented by medical imaging (the most common is CT scan, and sometimes also includes MRI, PET, etc.). The tumor location refers to the specific spatial coordinates and anatomical region of the tumor inside the patient's body. The dose distribution refers to the spatial dose intensity pattern formed by the energy deposited in the patient's body (or on the surface) by the rays generated by the radiotherapy equipment (such as a linear accelerator) when formulating and evaluating the radiotherapy plan.

[0025] As an embodiment of the present invention, extracting the radiotherapy plan features of the radiotherapy plan data includes: Unifying the format of the radiotherapy plan data to obtain standardized radiotherapy plan data; De-noising the standardized radiotherapy plan data to obtain de-noised radiotherapy plan data; performing coordinate system alignment on the denoised radiotherapy plan data to obtain aligned radiotherapy plan data; identifying key structures of the aligned radiotherapy planning data, and calculating geometric features and anatomical relationship features of the key structures; determining the anatomical structure features and tumor position of the aligned radiotherapy planning data according to the geometric features and the anatomical relationship features; calculating target dose characteristics and organ-at-risk dose characteristics at the tumor location based on the aligned radiotherapy plan data; determining a dose distribution at the tumor location based on the target volume dose characteristics and the organ-at-risk dose characteristics; The radiotherapy planning features of the radiotherapy planning data are determined according to the dose distribution, the anatomical structure features and the tumor position.

[0026] Standardized radiotherapy plan data refers to radiotherapy plan data that, after undergoing standardization steps, conforms to specific standards, has a unified format, and is of controlled quality. Denoised radiotherapy plan data refers to data after noise and interference have been removed or reduced using filtering techniques. Aligned radiotherapy plan data refers to radiotherapy plan data that has undergone coordinate alignment or registration. Key structures refer to specific anatomical regions of clinical and anatomical significance during radiotherapy, such as tumor targets and vital organs. Geometric features refer to quantitative descriptions of the shape and spatial position of anatomical structures as delineated by medical imaging (primarily CT scans). Anatomical relationship features refer to quantitative descriptions of the relative position, distance, and contact between different anatomical structures in three-dimensional space within a radiotherapy plan. Target dose features refer to a series of quantitative indicators used to measure and describe the radiation dose received by the tumor target in a radiotherapy plan. Organ-at-risk dose features refer to a series of quantitative indicators used to measure and describe the radiation dose received by normal tissues (organs-at-risk) in a radiotherapy plan.

[0027] Optionally, the denoised radiotherapy plan data may be obtained through a filter, such as a mean filter, a median filter, etc.

[0028] Optionally, the geometric features of the key structure may be calculated by a boundary detection algorithm, such as Roberts Cross operator, Sobel operator, Prewitt operator, etc.

[0029] By training the initial dose distribution analysis model for the source medical institution, embodiments of the present invention can directly generate a higher-quality initial plan, reducing the number and magnitude of subsequent optimization adjustments and shortening the overall planning time. The initial dose distribution analysis model refers to a machine learning-based method for generating, evaluating, or optimizing a model for the initial dose distribution corresponding to a radiotherapy plan.

[0030] As an embodiment of the present invention, the training of the initial dose distribution analysis model of the source domain medical institution includes: Dividing the radiotherapy plan features corresponding to the source domain medical institution into a training set and a validation set; Generate a model framework of the source domain medical institution; Based on the training set, the model framework is trained to obtain a training analysis model; Calculating the mean square error of the training analysis model based on the validation set; When the mean square error is greater than a preset error threshold, after adjusting the model learning parameters of the training analysis model, returning to the step of training the model framework based on the training set to obtain a training analysis model; When the mean square error is less than a preset error threshold, the training analysis model is used as the initial dose distribution analysis model of the source domain medical institution.

[0031] Among them, the training set refers to a part of data divided from the radiotherapy plan characteristic data of the source domain medical institution for training the initial dose distribution analysis model. The validation set refers to a part of data divided from the radiotherapy plan characteristic data of the source domain medical institution for evaluating and verifying the performance of the model during the training process. The model framework refers to the computational structure used to construct the initial dose distribution analysis model. The training analysis model refers to a model used to evaluate and optimize the performance of the initial dose distribution analysis model during the model training process. The mean square error refers to an indicator used to measure the degree of difference between the model output value and the true value of the verification set. The preset error threshold refers to a limit value set in advance to determine whether the error generated by the model is within an acceptable range. The model learning parameters refer to the values ​​that the model automatically adjusts through learning data during the training process, such as learning rate, weight, bias, etc.

[0032] Optionally, the model framework of the source domain medical institution can be generated by linear regression, such as ridge regression, Lasso, neural network, etc.

[0033] As another embodiment, the mean square error of the training analysis model can be calculated by the following formula:

[0034] in, represents the mean square error, Indicates the number of data sets for the validation set, Indicates the validation set The output value of the group data in the training analysis model, Indicates the validation set The true value of the group data.

[0035] S2. Construct a federated learning network between the source domain medical institution and the target domain medical institution. Based on the federated learning network, distribute the initial dose distribution analysis model to the target domain medical institution. Output the dose distribution analysis map of the initial dose distribution analysis model through the local radiotherapy plan data of the target domain medical institution.

[0036] By constructing a federated learning network between the source and target medical institutions, this embodiment of the present invention enables cross-institutional data collaboration while strictly protecting data privacy, thereby improving the generalization, applicability, and overall performance of the initial dose distribution analysis model. The federated learning network refers to a machine learning architecture and computing framework that allows multiple participants (in this scenario, source and target medical institutions) to jointly train a machine learning model without sharing their raw data.

[0037] As an embodiment of the present invention, the step of constructing a federated learning network of the source domain medical institution and the target domain medical institution includes: Clarifying the source domain client node of the source domain medical institution and the target domain client node of the target domain medical institution; Establishing a communication protocol and a dynamic encryption protocol between the source domain client node and the target domain client node; Building a central server for the source domain client node and the target domain client node based on the communication protocol and the dynamic encryption protocol; uniformly configuring the client environments of the source domain client node and the target domain client node; Defining a model training loop rule for the source domain client node, the target domain client node, and the central server according to the client environment; According to the model training cycle rules, a federated learning network of the source domain medical institution and the target domain medical institution is constructed.

[0038] The source domain client node refers to a medical institution client that participates in federated learning, provides an initial model or baseline model, and may contribute some training data. The target domain client node refers to a medical institution client that participates in federated learning and uses federated learning to improve the performance of the target domain model. The communication protocol refers to a set of predefined rules, standards, and conventions that regulate how information is exchanged and communicated between nodes in the federated learning network (e.g., source domain client nodes, target domain client nodes, and central server). The dynamic encryption protocol refers to an encryption protocol that adjusts or changes in real time based on communication conditions, time, data characteristics, or communication context. The central server refers to a node responsible for coordinating and managing the entire distributed learning process. The client environment refers to all the software and hardware conditions and configurations required for each client node (e.g., source domain client node, target domain client node) participating in the federated learning network to run federated learning tasks on its local device. The model training loop rules refer to a set of steps that regulate how the model is trained over multiple rounds of iterations throughout the federated learning process.

[0039] Optionally, the dynamic encryption protocol between the source domain client node and the target domain client node can be constructed by a dynamic differential privacy algorithm, such as determining the communication content, the number of participating nodes and potential attackers of the data transmitted between the source domain client node and the target domain client node through the dynamic differential privacy algorithm, analyzing the leakage risk coefficient of the transmitted data based on the communication content, the number of participating nodes and the potential attackers, and constructing the dynamic encryption protocol between the source domain client node and the target domain client node based on the leakage risk coefficient.

[0040] Optionally, the model training loop rules of the source domain client node, the target domain client node, and the central server may be defined by an iterative optimization algorithm, such as a gradient descent algorithm, a federated averaging algorithm, and the like.

[0041] The embodiment of the present invention distributes the initial dose distribution analysis model to the target domain medical institutions based on the federated learning network, which can accelerate deployment, improve initial performance, support local personalized optimization, realize knowledge sharing, and lower the usage threshold, laying a good foundation for subsequent more accurate and more local demand-oriented dose distribution analysis in the target domain.

[0042] This embodiment of the present invention uses the local radiotherapy plan data of the target domain medical institution to output a dose distribution analysis map of the initial dose distribution analysis model, which can preliminarily determine the applicability and accuracy of the initial model for the target domain data, providing a basis for determining whether and how to perform subsequent fine-tuning. The dose distribution analysis map is a graphical tool used in radiotherapy planning to intuitively display the distribution of radiation dose in the target area.

[0043] As an embodiment of the present invention, the step of outputting the dose distribution analysis graph of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution includes: Collecting local radiotherapy plan data of the target domain medical institution; extracting key information of the local radiotherapy planning data; Converting the key information into input format information; Inputting the input format information into the initial dose distribution analysis model to obtain dose distribution data; A dose distribution analysis map of the patient corresponding to the target domain medical institution is generated based on the dose distribution data.

[0044] The local radiotherapy planning data refers to all data related to the patient's radiotherapy plan generated by the target medical institution (such as a hospital or clinic) during the radiotherapy process. Key information refers to the core information necessary for the model to perform dose distribution analysis, such as target volume and organ contour information, imaging data, and radiotherapy planning data. Input format information refers to the extracted key information converted and formatted according to the input format required by the initial dose distribution analysis model. Dose distribution data refers to numerical data describing the deposition or distribution of energy (typically radiation energy) within the patient's body.

[0045] Optionally, the key information of the local radiotherapy plan data can be extracted through DICOM RT standard parsing, such as pydicom and dcm2jpg in Python, DcmObject in Java, DCMTK in C++, etc.

[0046] Optionally, the dose distribution analysis map of the patient corresponding to the target domain medical institution can be generated by image superposition and fusion technology, such as superimposing the dose distribution volume data and the anatomical structure image corresponding to the patient to obtain a dose distribution superposition map, and slicing and rendering the dose distribution superposition map according to the dose value corresponding to the dose distribution volume data to obtain a dose distribution analysis map.

[0047] S3. Combining the dose distribution analysis atlas and the local radiotherapy plan data, identifying domain differences between the source domain medical institution and the target domain medical institution.

[0048] By combining the dose distribution analysis atlas with the local radiotherapy planning data, embodiments of the present invention can identify domain differences between the source and target medical institutions, thereby intuitively and quantitatively identifying systematic differences in dose distributions between the two institutions for similar cases. Domain differences refer to systematic differences in the distribution, characteristics, statistical properties, or objectives of data or models from different sources.

[0049] As an embodiment of the present invention, the identifying domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data includes: extracting the model analysis dose of the dose distribution analysis atlas and the local plan dose of the local radiotherapy plan data; uniformly identifying the anatomical structure of the dose distribution analysis atlas and the local radiotherapy planning data; Calculate the domain difference between the source domain medical institution and the target domain medical institution according to the anatomical structure, the model analysis dose, and the local planned dose: The model analysis dose refers to the dose value extracted from the dose distribution analysis atlas for analysis. The local plan dose refers to the actual clinical radiotherapy dose distribution data developed by the target domain medical institution for its patients. The anatomical structure refers to a region or organ with specific physiological or anatomical significance defined in the radiotherapy plan. The voxel refers to the point in three-dimensional space that represents the smallest volume unit of an anatomical structure.

[0050] As another implementation, based on the above technical solution, the domain difference can be calculated using the following formula:

[0051] in, Indicates domain differences, represents the total number of anatomical structures, Indicates the local radiotherapy plan data The volume of the anatomical structure (unit: ), Indicates the The total number of voxels of the anatomical structure, Indicates the local radiotherapy plan data The anatomical structure corresponds to The local planned dose per voxel, Indicates the dose distribution analysis map The anatomical structure corresponds to Model analysis dose per voxel, Indicates the local radiotherapy plan data The local planned dose to each anatomical structure, Indicates the dose distribution analysis map Dose analysis using a model of an anatomical structure.

[0052] For example, there are two anatomical structures A and B, and the volume of A is 100 , the ratio of the intersection and union of the local planned dose and the model analysis dose in the anatomical structure A is 0.7 (i.e. =0.7), A contains 3 voxels, whose local planning doses are 2, 2.2 and 1.9 respectively, and the model analysis doses are 1.8, 2 and 2.1 respectively. The volume of B is 150 , the ratio of the intersection and union of the local planned dose and the model analysis dose in anatomical structure B is 0.6 (i.e. =0.6), B contains 3 voxels, whose local planned doses are 1.5, 1.8 and 1.4 respectively, and the model analysis doses are 1.6, 1.7 and 1.5 respectively. When put into the formula, the domain difference between the source domain medical institution and the target domain medical institution is 3.7.

[0053] Optionally, the model analysis dose of the dose distribution analysis map can be extracted by dose-volume histogram analysis.

[0054] S4. Based on the domain differences, clarify the optimization target of the initial dose distribution analysis model, wherein the optimization target includes: dose analysis loss, domain adversarial loss, and feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model.

[0055] The embodiment of the present invention can make the model better adapt to the subtle differences in the target domain data distribution by clarifying the optimization target of the initial dose distribution analysis model based on the domain differences, thereby improving the portability and practicality of the model between different medical institutions. The optimization target refers to a series of loss functions or a combination of objective functions used to guide the initial dose distribution analysis model to adjust its own parameters to adapt to the target domain medical institution data during the domain adaptive transfer learning process. The dose analysis loss refers to a loss function used to measure the difference between the model analysis dose predicted by the model and the local planned dose of the local actual plan of the target domain medical institution. The domain adversarial loss refers to a loss function that reduces the difference in feature distribution between the source domain and the target domain through adversarial training. The feature alignment loss refers to a loss function used to measure the distribution difference between the source domain and the target domain data in the feature space extracted by the model.

[0056] As an embodiment of the present invention, clarifying the optimization target of the initial dose distribution analysis model based on the domain difference includes: analyzing difference features of the domain differences, wherein the difference features include dose volume histogram differences and dose distribution differences; determining an anatomical structure weight, a dose level weight, and a voxel importance weight of the initial dose distribution analysis model according to the difference characteristics; determining a dose analysis loss of the initial dose distribution analysis model based on the anatomical structure weight, the dose level weight, and the voxel importance weight; constructing a domain classifier of the initial dose distribution analysis model; extracting input features of the initial dose distribution analysis model, inputting the input features into the domain classifier, and obtaining a classifier analysis value; Calculating a domain adversarial loss of the initial dose distribution analysis model according to the classifier analysis value; extracting source domain data features and target domain data features of the initial dose distribution analysis model respectively; respectively calculating a source domain kernel matrix between the source domain data features and a target domain kernel matrix between the target domain data features; Calculating a kernel matrix between the source domain data features and the target domain data features; determining a feature alignment loss of the initial dose distribution analysis model according to the source domain kernel matrix, the target domain kernel matrix, and the kernel matrix; An optimization target of the initial dose distribution analysis model is determined according to the feature alignment loss, the dose analysis loss, and the domain confrontation loss.

[0057] The difference features refer to specific, quantifiable differences in dose distribution and allocation between the source domain data and the local medical institution data. The dose-volume histogram difference refers to the systematic deviation in the dose-volume histograms between the source and target domains for the same type of anatomical structure. The dose distribution difference refers to the inconsistency in the distribution of dose values ​​in three-dimensional space between the source and target domain data. The anatomical structure weight refers to the weight coefficient associated with a specific anatomical structure assigned to the difference between the model analysis result and the true dose distribution during dose analysis loss calculation. The dose level weight refers to the weight coefficient associated with a specific dose value assigned to the difference between the model prediction result and the true dose distribution during dose analysis loss calculation. The voxel importance weight refers to the weight coefficient associated with a single voxel in the image assigned to the difference between the model prediction result and the true dose distribution during dose analysis loss calculation. The domain classifier refers to a neural network module that facilitates the fusion of source and target domain data features. The classifier analysis value refers to the analysis probability output by the domain classifier after performing domain classification on the input features. The source domain data features refer to the intrinsic feature representations extracted from the source domain data that can represent the data, such as DICOM images of the source domain organization, patient anatomical structures, and treatment plan parameters. The target domain data features refer to the intrinsic feature representations extracted from the target domain data that can represent the data, such as DICOM images of the target domain organization, patient anatomical structures, and treatment plan parameters. The source domain kernel matrix refers to the similarity between each feature in the source domain dataset. The target domain kernel matrix refers to the correlation measure between each feature in the target domain dataset. The kernel matrix refers to the similarity between each feature in the target domain dataset and the source domain dataset.

[0058] Optionally, the anatomical structure weight, dose level weight, and voxel importance weight of the initial dose distribution analysis model may be determined by a statistical difference analysis method, such as using mean square error, structural similarity index SSIM, and the like.

[0059] Optionally, the domain classifier of the initial dose distribution analysis model can be constructed by a machine learning classification algorithm, such as support vector machine, logistic regression, random forest, K-nearest neighbor, etc.

[0060] Optionally, the source domain kernel matrix between the source domain data features may be calculated using a kernel function, such as a linear kernel function, a polynomial kernel function, a Gaussian kernel function, and the like.

[0061] The embodiments of the present invention perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model, which can adapt the local dose distribution analysis to the specific local environment. This overcomes the challenges brought by inconsistent data distribution while maintaining or even improving performance, and achieves more accurate, efficient, and practical localized applications. The local dose distribution analysis model refers to a model obtained through domain adaptive transfer learning based on the initial dose distribution analysis model.

[0062] As an embodiment of the present invention, performing domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model includes: According to the optimization target of the initial dose distribution analysis model, the initial dose distribution analysis model is trained using preset training data to obtain a training dose distribution analysis model; Calculating the main task output and domain classification output of the trained dose distribution analysis model; Calculating a model main task loss and a model domain adversarial loss of the training dose distribution analysis model according to the main task output and the domain classification output; When the main task loss of the model is greater than a preset main task loss threshold, after adjusting the model parameters of the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a training dose distribution analysis model; When the model domain adversarial loss is less than a preset domain adversarial loss threshold, after adjusting the classifier parameters of the domain classifier corresponding to the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a trained dose distribution analysis model; When the model main task loss is less than a preset main task loss threshold and the model domain adversarial loss is greater than a preset domain adversarial loss threshold, the training dose distribution analysis model is used as the training dose distribution analysis model.

[0063] The preset training data refers to a dataset prepared before training the initial dose distribution analysis model for model learning and optimization. The trained dose distribution analysis model refers to a dose distribution analysis model that has been optimized to a certain extent, obtained after a specific training step. The primary task output refers to the dose distribution map analyzed by the trained dose distribution analysis model based on the input information. The domain classification output refers to the result generated by the classifier used to determine the source domain of the input data. The model primary task loss refers to a metric used to measure the degree of difference between the analysis results and the true results when the model performs its core analysis task. The model domain adversarial loss refers to a metric that measures the accuracy of the domain classifier in the model in determining the source domain of the data. The preset primary task loss threshold refers to a pre-set value used to determine whether the model's performance on the primary task (i.e., dose distribution analysis) has reached the target level. Model parameters refer to numerical values ​​that can be learned and adjusted within the model, such as model weights and model biases. The domain adversarial loss threshold refers to a pre-set numerical limit used to assess whether the model's domain generalization capability meets requirements. Classifier parameters refer to numerical values ​​that can be learned and adjusted within the domain classifier, such as classifier weights and classifier biases.

[0064] Optionally, the model main task loss of the training dose distribution analysis model can be obtained by calculating the structural similarity index, such as calculating the output local mean, output variance and output covariance of the main task output, and analyzing the true local mean, true variance and true covariance of the training data corresponding to the training dose distribution analysis model, and calculating the structural similarity index of the main task output and the training data based on the output local mean, the output variance, the output covariance, the true local mean, the true variance and the true covariance to determine the model main task loss of the training dose distribution analysis model.

[0065] Optionally, the model parameters of the training dose distribution analysis model can be adjusted by an optimization algorithm, such as a stochastic gradient descent algorithm, a cosine annealing algorithm, etc.

[0066] S5. Generate a target dose distribution for patients in the target domain medical institution based on the local dose distribution analysis model.

[0067] By generating a target dose distribution for patients in the target medical institution based on the local dose distribution analysis model, embodiments of the present invention can enhance the adaptability and robustness of the model across diverse medical institution environments, thereby more reliably and accurately generating a dose distribution for the target patients. The target dose distribution refers to the distribution of radiation dose received by each point in three-dimensional space within the patient's body, as calculated by the treatment planning system.

[0068] As an embodiment of the present invention, generating a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model includes: Inputting the patient data corresponding to the patient into the local dose distribution analysis model, and processing the patient data through the convolution layer of the local dose distribution analysis model to obtain high-level features and low-level detail features; The high-level features and the low-level detail features are subjected to feature fusion through a feature fusion layer of the local dose distribution analysis model to obtain fused features; Calculating a dose distribution analysis map of the patient through a dose calculation layer of the local dose distribution analysis model according to the fusion feature; According to the dose distribution analysis graph, the target dose distribution of the patient is output through the output layer of the local dose distribution analysis model.

[0069] The convolutional layer is a specific layer that extracts features from the input patient data. High-level features are highly abstract, information-rich representations of complex anatomical structures, spatial relationships, and patterns extracted from the patient data (CT images and contours) by the convolutional layer. These representations include the overall shape and contours of organs, the relative position of tumor regions and surrounding organs at risk, and the spatial relationships between organs. The skip connection layer extracts basic visual elements from the patient data, such as bone edges and tissue texture. The feature fusion layer is a module specifically responsible for combining and integrating feature representations from different sources and with different characteristics. The fused feature is a new, more comprehensive feature representation obtained by processing and combining high-level and low-level detail features through the feature fusion layer. The dose calculation layer utilizes the information extracted and integrated by all previous layers to convert it into a dose distribution map within the patient's body. The dose distribution analysis map is the preliminary, analyzed dose distribution result generated by the model after the dose calculation layer completes its work.

[0070] Optionally, the high-level features and the low-level detail features can be obtained through a high-level semantic path and a low-level detail path, such as determining the dilation rate and the convolution kernel weight of the convolution layer, and constructing the high-level semantic path and the low-level detail path of the convolution layer according to the dilation rate and the convolution kernel weight to extract the high-level features and the low-level detail features of the patient data.

[0071] Compared with the problems described in the background technology, the embodiment of the present invention can directly generate a higher quality initial plan by training the initial dose distribution analysis model of the source domain medical institution, reduce the number and magnitude of subsequent optimization adjustments, and shorten the overall planning time; optionally, the embodiment of the present invention can achieve cross-institutional data collaboration under the premise of strictly protecting data privacy by constructing a federated learning network of the source domain medical institution and the target domain medical institution, thereby improving the generalization ability, applicability and overall performance of the initial dose distribution analysis model; the embodiment of the present invention outputs the dose distribution analysis map of the initial dose distribution analysis model through the local radiotherapy plan data of the target domain medical institution, which can preliminarily judge the applicability and accuracy of the initial model on the target domain data, and provide a basis for whether fine-tuning is needed and how to fine-tune it later; the embodiment of the present invention combines the dose distribution analysis model with the target domain medical institution to obtain the dose distribution analysis map of the initial dose distribution analysis model. By analyzing the atlas and the local radiotherapy plan data, identifying the domain differences between the source domain medical institution and the target domain medical institution, it is possible to intuitively and quantitatively identify the systematic differences in dose distribution between the two institutions in similar cases; the embodiment of the present invention performs domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model, which can enable the local dose distribution analysis to adapt to a specific local environment, thereby overcoming the challenges brought by inconsistent data distribution while maintaining or even improving performance, and achieving more accurate, more efficient and more practical localized applications; finally, the embodiment of the present invention generates the target dose distribution of patients in the target domain medical institution according to the local dose distribution analysis model, which can enhance the adaptability and robustness of the model in different medical institution environments, thereby generating dose distributions for target domain patients more reliably and accurately. Therefore, the radiotherapy plan dose distribution prediction method based on domain adaptive transfer learning provided by the embodiment of the present invention can improve the efficiency and accuracy of radiotherapy plan dose distribution prediction.

[0072] Example 2: like Figure 2 As shown in the figure, it is a functional module diagram of a radiotherapy plan dose distribution prediction system based on domain adaptive transfer learning of the present invention.

[0073] The radiation therapy plan dose distribution prediction system 200 based on domain adaptive transfer learning described in the present invention can be installed in an electronic device. According to the functions implemented, the radiation therapy plan dose distribution prediction system based on domain adaptive transfer learning can include a source domain model training module 201, a local data prediction module 202, a domain difference analysis module 203, a local model training module 204, and a target dose distribution module 205. The module described in the present invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can perform fixed functions, which are stored in the memory of the electronic device.

[0074] In the embodiment of the present invention, the functions of each module / unit are as follows: The source domain model training module 201 is used to identify a source domain medical institution and a target domain medical institution, collect radiotherapy plan data of the source domain medical institution, and extract radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location, and dose distribution, so as to train an initial dose distribution analysis model of the source domain medical institution; The local data prediction module 202 is configured to construct a federated learning network between the source domain medical institution and the target domain medical institution, distribute the initial dose distribution analysis model to the target domain medical institution based on the federated learning network, and output a dose distribution analysis map of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution; The domain difference analysis module 203 is configured to identify domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data; The local model training module 204 is configured to determine an optimization target of the initial dose distribution analysis model based on the domain difference, wherein the optimization target includes a dose analysis loss, a domain adversarial loss, and a feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; The target dose distribution module 205 is configured to generate a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model.

[0075] In detail, the modules in the radiotherapy plan dose distribution prediction system 200 based on domain adaptive transfer learning in the embodiment of the present invention are used in the same manner as above. Figure 1 The same technical means are used as the method for predicting the dose distribution of radiotherapy plans based on domain adaptive transfer learning described in , and can produce the same technical effects, so they will not be repeated here.

[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0077] Finally, it should be noted that among the multiple embodiments described above, each embodiment can be combined with each other or be independent, and deleting any one of them will not affect the technical implementation of other embodiments. The above embodiments are only used to illustrate the technical solution of the present invention and are not limiting. Although the present invention is described in detail with reference to the preferred embodiments, ordinary technicians in this field should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting radiotherapy plan dose distribution based on domain adaptive transfer learning, characterized in that: The method comprises: Identify a source domain medical institution and a target domain medical institution, collect radiotherapy plan data of the source domain medical institution, and extract radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location, and dose distribution, so as to train an initial dose distribution analysis model for the source domain medical institution; Constructing a federated learning network between the source domain medical institution and the target domain medical institution, distributing the initial dose distribution analysis model to the target domain medical institution based on the federated learning network, and outputting a dose distribution analysis map of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution; identifying domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data; Based on the domain differences, clarifying optimization objectives of the initial dose distribution analysis model, wherein the optimization objectives include: dose analysis loss, domain adversarial loss, and feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; A target dose distribution for patients in the target domain medical institution is generated according to the local dose distribution analysis model.

2. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The extracting the radiotherapy plan features of the radiotherapy plan data includes: Unifying the format of the radiotherapy plan data to obtain standardized radiotherapy plan data; De-noising the standardized radiotherapy plan data to obtain de-noised radiotherapy plan data; performing coordinate system alignment on the denoised radiotherapy plan data to obtain aligned radiotherapy plan data; identifying key structures of the aligned radiotherapy planning data, and calculating geometric features and anatomical relationship features of the key structures; determining the anatomical structure features and tumor position of the aligned radiotherapy planning data according to the geometric features and the anatomical relationship features; calculating target dose characteristics and organ-at-risk dose characteristics at the tumor location based on the aligned radiotherapy plan data; determining a dose distribution at the tumor location based on the target volume dose characteristics and the organ-at-risk dose characteristics; The radiotherapy planning features of the radiotherapy planning data are determined according to the dose distribution, the anatomical structure features and the tumor position.

3. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The training of the initial dose distribution analysis model of the source domain medical institution includes: Dividing the radiotherapy plan features corresponding to the source domain medical institution into a training set and a validation set; generating a model framework of the source domain medical institution; Based on the training set, the model framework is trained to obtain a training analysis model; Calculating the mean square error of the training analysis model based on the validation set; When the mean square error is greater than a preset error threshold, after adjusting the model learning parameters of the training analysis model, returning to the step of training the model framework based on the training set to obtain a training analysis model; When the mean square error is less than a preset error threshold, the training analysis model is used as the initial dose distribution analysis model of the source domain medical institution.

4. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The constructing of the federated learning network of the source domain medical institution and the target domain medical institution includes: Clarifying the source domain client node of the source domain medical institution and the target domain client node of the target domain medical institution; Establishing a communication protocol and a dynamic encryption protocol between the source domain client node and the target domain client node; Building a central server for the source domain client node and the target domain client node based on the communication protocol and the dynamic encryption protocol; uniformly configuring the client environments of the source domain client node and the target domain client node; Defining a model training loop rule for the source domain client node, the target domain client node, and the central server according to the client environment; According to the model training cycle rules, a federated learning network of the source domain medical institution and the target domain medical institution is constructed.

5. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: Outputting the dose distribution analysis graph of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution includes: Collecting local radiotherapy plan data of the target domain medical institution; extracting key information of the local radiotherapy planning data; Converting the key information into input format information; Inputting the input format information into the initial dose distribution analysis model to obtain dose distribution data; A dose distribution analysis map of the patient corresponding to the target domain medical institution is generated based on the dose distribution data.

6. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The identifying the domain difference between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data includes: extracting the model analysis dose of the dose distribution analysis atlas and the local plan dose of the local radiotherapy plan data; uniformly identifying the anatomical structure of the dose distribution analysis atlas and the local radiotherapy planning data; The domain difference between the source domain medical institution and the target domain medical institution is calculated according to the anatomical structure, the model analysis dose, and the local planned dose.

7. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The step of clarifying the optimization goal of the initial dose distribution analysis model based on the domain differences includes: analyzing difference features of the domain differences, wherein the difference features include dose volume histogram differences and dose distribution differences; determining an anatomical structure weight, a dose level weight, and a voxel importance weight of the initial dose distribution analysis model according to the difference characteristics; determining a dose analysis loss of the initial dose distribution analysis model based on the anatomical structure weight, the dose level weight, and the voxel importance weight; constructing a domain classifier of the initial dose distribution analysis model; extracting input features of the initial dose distribution analysis model, inputting the input features into the domain classifier, and obtaining a classifier analysis value; Calculating a domain adversarial loss of the initial dose distribution analysis model according to the classifier analysis value; extracting source domain data features and target domain data features of the initial dose distribution analysis model respectively; respectively calculating a source domain kernel matrix between the source domain data features and a target domain kernel matrix between the target domain data features; Calculating a kernel matrix between the source domain data features and the target domain data features; determining a feature alignment loss of the initial dose distribution analysis model according to the source domain kernel matrix, the target domain kernel matrix, and the kernel matrix; An optimization target of the initial dose distribution analysis model is determined according to the feature alignment loss, the dose analysis loss, and the domain confrontation loss.

8. The method for predicting radiation therapy dose distribution based on domain adaptive transfer learning according to claim 1, wherein: The performing domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model includes: According to the optimization target of the initial dose distribution analysis model, the initial dose distribution analysis model is trained using preset training data to obtain a training dose distribution analysis model; Calculating the main task output and domain classification output of the trained dose distribution analysis model; Calculating a model main task loss and a model domain adversarial loss of the training dose distribution analysis model according to the main task output and the domain classification output; When the main task loss of the model is greater than a preset main task loss threshold, after adjusting the model parameters of the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a training dose distribution analysis model; When the model domain adversarial loss is less than a preset domain adversarial loss threshold, after adjusting the classifier parameters of the domain classifier corresponding to the training dose distribution analysis model, returning to the optimization target of the initial dose distribution analysis model, and training the initial dose distribution analysis model using preset training data to obtain a trained dose distribution analysis model; When the model main task loss is less than a preset main task loss threshold and the model domain adversarial loss is greater than a preset domain adversarial loss threshold, the training dose distribution analysis model is used as the training dose distribution analysis model.

9. The method for predicting radiation therapy plan dose distribution based on domain adaptive transfer learning according to claim 8, wherein: Generating a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model includes: Inputting the patient data corresponding to the patient into the local dose distribution analysis model, and processing the patient data through the convolution layer of the local dose distribution analysis model to obtain high-level features and low-level detail features; The high-level features and the low-level detail features are subjected to feature fusion through a feature fusion layer of the local dose distribution analysis model to obtain fused features; Calculating a dose distribution analysis map of the patient through a dose calculation layer of the local dose distribution analysis model according to the fusion feature; According to the dose distribution analysis graph, the target dose distribution of the patient is output through the output layer of the local dose distribution analysis model.

10. A radiotherapy plan dose distribution prediction system based on domain adaptive transfer learning, the system implementing the method according to claim 1, characterized in that: The system comprises: A source domain model training module is used to identify a source domain medical institution and a target domain medical institution, collect radiotherapy plan data of the source domain medical institution, and extract radiotherapy plan features of the radiotherapy plan data, wherein the radiotherapy plan features include: anatomical structure features, tumor location, and dose distribution, so as to train an initial dose distribution analysis model for the source domain medical institution; A local data prediction module is used to construct a federated learning network between the source domain medical institution and the target domain medical institution, distribute the initial dose distribution analysis model to the target domain medical institution based on the federated learning network, and output a dose distribution analysis map of the initial dose distribution analysis model using the local radiotherapy plan data of the target domain medical institution; a domain difference analysis module, configured to identify domain differences between the source domain medical institution and the target domain medical institution by combining the dose distribution analysis atlas and the local radiotherapy plan data; a local model training module, configured to clarify an optimization target of the initial dose distribution analysis model based on the domain difference, wherein the optimization target includes: dose analysis loss, domain adversarial loss, and feature alignment loss, so as to perform domain adaptive transfer learning on the initial dose distribution analysis model to obtain a local dose distribution analysis model; The target dose distribution module is used to generate a target dose distribution for patients in the target domain medical institution according to the local dose distribution analysis model.

Images