Dosage prediction methods, devices, equipment and media

Abstract

This application discloses a dose prediction method, apparatus, device, and medium, relating to the field of intelligent medical technology. The method includes: acquiring a scan training image and a radiation dose distribution image; training a first parameter in an initial network model based on the scan training image and the radiation dose distribution image to obtain at least two teacher network models; transmitting information from the at least two teacher network models to a student network model through adversarial learning, adjusting the second parameter of the student network model to obtain a dose prediction model; and predicting the radiation dose distribution corresponding to a target scan image using the dose prediction model to obtain a radiation dose prediction image corresponding to the target scan image. By transmitting information from multiple teacher network models to the student network model through adversarial learning, a radiation dose prediction model for predicting the dose distribution of scan images is obtained, improving the reliability of dose distribution prediction.

Description

Technical Field

[0001] This application relates to the field of intelligent medical technology, and in particular to a dose prediction method, device, equipment and medium. Background Technology

[0002] Radiotherapy has become an important means of treating and controlling tumors. Before radiotherapy, a radiotherapy plan needs to be designed, and the quality of the radiotherapy plan design affects the effect of the radiotherapy process. When designing a radiotherapy plan, it is necessary to ensure the radiation dose to the target area while minimizing the radiation dose to normal tissues and organs.

[0003] To achieve the ideal dose distribution, radiotherapy plans are usually designed by professional planners. This means that planners need to carefully adjust the treatment plan parameters and weights through multiple repeated experiments and iterations to control the trade-offs between clinical goals.

[0004] However, the above process is very time-consuming and labor-intensive, and it can also result in significant differences due to the different experience of different planners. This means that the efficiency of radiotherapy plan design is relatively low, and the reliability of the design is affected by the difference in human experience. Summary of the Invention

[0005] This application provides a dose prediction method, apparatus, device, and medium, which can improve the reliability of dose distribution prediction for scanned images. The technical solution is as follows:

[0006] On the one hand, a dose prediction method is provided, the method comprising:

[0007] Acquire scan training images and radiation dose distribution images, wherein the radiation dose distribution images are used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the scan training images;

[0008] The first parameter in the initial network model is trained based on the scan training image and the radiation dose distribution image to obtain at least two teacher network models;

[0009] Information from the at least two teacher network models is transmitted to the student network model through adversarial learning, and the second parameter of the student network model is adjusted to obtain the dose prediction model.

[0010] The radiation dose distribution corresponding to the target scan image is predicted by the dose prediction model to obtain the radiation dose prediction image corresponding to the target scan image.

[0011] On the other hand, a dose prediction device is provided, the device comprising:

[0012] An acquisition module is used to acquire scan training images and radiation dose distribution images, wherein the radiation dose distribution images are used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the scan training images;

[0013] The training module is used to train the first parameters in the initial network model based on the scanned training image and the radiation dose distribution image to obtain at least two teacher network models;

[0014] The training module is also used to transfer information from the at least two teacher network models to the student network model through adversarial learning, adjust the second parameter of the student network model, and obtain a dose prediction model.

[0015] The prediction module is used to predict the radiation dose distribution corresponding to the target scan image through the dose prediction model, so as to obtain the radiation dose prediction image corresponding to the target scan image.

[0016] On the other hand, a computer device is provided, the terminal including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement any of the dose prediction methods described in the embodiments of this application.

[0017] On the other hand, a computer-readable storage medium is provided, wherein at least one piece of program code is stored in the computer-readable storage medium, the program code being loaded and executed by a processor to implement any of the dose prediction methods described in the embodiments of this application.

[0018] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the dose prediction methods described in the above embodiments.

[0019] The technical solution provided in this application includes at least the following beneficial effects:

[0020] The first parameters of the initial network model are trained based on the scanned training images and radiation dose distribution images to obtain at least two teacher network models. Then, the student network model is trained based on the at least two teacher network models. Based on adversarial learning, the information from the at least two teacher network models is transferred to the student network model through knowledge distillation to adjust the second parameters of the student network model, resulting in a dose prediction model. This dose prediction model has higher performance and accuracy. It combines the simple structure of the student network model with the prediction accuracy of the teacher network model, and obtains more accurate prediction results when predicting the dose distribution of the target scanned image, thus improving the reliability of dose distribution prediction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of this application;

[0023] Figure 2 This is a flowchart of a dose prediction method provided in an exemplary embodiment of this application;

[0024] Figure 3 This is a schematic diagram of the U-NAS framework provided in an exemplary embodiment of this application;

[0025] Figure 4 This is a schematic diagram of the network topology of a downsampling unit provided in an exemplary embodiment of this application;

[0026] Figure 5 This is a schematic diagram of the network topology of an upsampling unit provided in an exemplary embodiment of this application;

[0027] Figure 6 This is a flowchart of a dose prediction method provided in another exemplary embodiment of this application;

[0028] Figure 7 This is a schematic diagram of an HM topology provided in an exemplary embodiment of this application;

[0029] Figure 8 This is a flowchart of a dose prediction method provided in another exemplary embodiment of this application;

[0030] Figure 9 This is a diagram of the LE-NAS framework provided in an exemplary embodiment of this application;

[0031] Figure 10 This is a schematic diagram of the structure of a dose prediction device provided in an exemplary embodiment of this application;

[0032] Figure 11 This is a schematic diagram of the structure of a dose prediction device provided in another exemplary embodiment of this application;

[0033] Figure 12 This is a schematic diagram of the structure of a server provided in an exemplary embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0035] First, a brief introduction to the terms used in the embodiments of this application:

[0036] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0037] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0038] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instruction-based learning.

[0039] Computer Vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in tasks such as target recognition, tracking, and measurement, and further performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, OCR (Optical Character Recognition), video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), autonomous driving, intelligent transportation, and other technologies, as well as common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0040] In this embodiment of the application, computer vision technology is used to identify and predict the scanned image to obtain a radiation dose distribution image corresponding to the scanned image, thereby realizing the design of a radiotherapy plan. In the process of realizing the intelligent design of radiotherapy plans, the efficiency and reliability of radiotherapy plan design are improved.

[0041] Based on the above definitions, the implementation environment of the embodiments in this application will be described. Please refer to... Figure 1 This diagram illustrates a structural block diagram of a computer system provided in an exemplary embodiment of this application. The computer system 100 includes: a terminal 110, a server 120, and a communication network 130.

[0042] Terminal 110 is used to acquire scanned images and upload them to server 120. Server 120 predicts the radiation dose from the scanned images and returns the corresponding radiation dose distribution image to terminal 110, thus providing the radiation dose distribution image corresponding to the scanned image to the doctor to assist in designing a radiotherapy plan. In some embodiments, terminal 110 includes a scanning device 111, a display device 112, and a storage device 113. The scanning device 111 is used to scan a patient or a sample of tissue or organ to be tested to obtain scanned images. The display device 112 is used to display the scanned images and the radiation dose distribution image. The storage device 113 is used to store and process the information data collected during the scan. In this embodiment, taking a CT device as an example, the scanning device 111 can be a device composed of an X-ray tube, a detector, and a scanning gantry. Illustratively, the scanning device 111, display device 112, and storage device 113 can be implemented as the same physical device or as multiple physical devices; no limitation is made here. The aforementioned terminal 110 can be a mobile phone, tablet computer, desktop computer, portable laptop computer, or other terminal devices, and this application embodiment does not limit it to any particular type.

[0043] Server 120 is used to train dose prediction model 121 using the dose prediction method provided in this application embodiment, and to perform dose distribution prediction on the scanned image uploaded by terminal 110 using the trained dose prediction model 121, thereby obtaining a dose distribution prediction image corresponding to the scanned image, and feeding back the dose distribution prediction image to terminal 110.

[0044] It is worth noting that the aforementioned server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0045] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. Based on the cloud computing business model, cloud technology encompasses network technology, information technology, integration technology, management platform technology, and application technology. It can form resource pools, providing flexible and convenient on-demand access. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data at different levels will be processed separately, and various industry data will require robust system support, which can only be achieved through cloud computing.

[0046] In some embodiments, the server 120 can also be implemented as a node in a blockchain system. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and cryptographic algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0047] To illustrate, terminal 110 and server 120 are connected via communication network 130.

[0048] Based on the above introduction to terminology and implementation environment, the application scenarios involved in the embodiments of this application will be described.

[0049] Radiation therapy planning is a crucial aspect of cancer treatment, requiring the allocation of radiation dose to the target area while minimizing radiation dose to normal tissues and organs. To achieve an ideal dose distribution, planners must meticulously adjust treatment plan parameters and weights through numerous iterative experiments to control the trade-offs between clinical goals. This process is extremely time-consuming and labor-intensive, and can vary significantly due to differences in the experience of different planners. In recent years, with the development of intelligent medical technologies, deep learning has been increasingly applied to radiation oncology. Data-driven methods can directly predict dose distribution; for example, the U-Net network can be used to predict the dose distribution during radiation therapy for prostate cancer. However, existing dose distribution prediction networks cannot guarantee applicability to various physicians, diseases, and clinical environments.

[0050] This application proposes a learnable model ensemble method, LE-NAS (Learning-based Ensemble Approach With Neural Architecture Search), to predict radiation dose corresponding to scanned images. Employing a teacher-student network paradigm, it integrates the different outputs of multiple automatically searching neural networks to guide student network training. This includes training the teacher network using a U-shaped differentiable neural network search framework (U-NAS) to ensure flexibility and diversity, and transferring the integrated information from the teacher network to the student network using a knowledge distillation network with adversarial learning (KDA-Net) framework. The dose prediction method provided in this application improves the reliability of radiation dose prediction for scanned images. Furthermore, this dose prediction model has greater applicability to different disease types and clinical environments.

[0051] Please refer to Figure 2 This illustrates a dose prediction method according to one embodiment of the present application, in which the method is applied to, for example, Figure 1 In the server shown, the method includes:

[0052] Step 201: Obtain the scan training image and radiation dose distribution image.

[0053] There is a correspondence between the training scan image and the radiation dose distribution image. The radiation dose distribution image is used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the training scan image.

[0054] Training images are medical images obtained through scanning equipment, such as computed tomography (CT) images, X-ray images, and ultrasound images.

[0055] In this embodiment, the scan training images are CT images of past patients, and the radiation dose distribution images are clinical dose distribution images corresponding to the CT images. These scan training images and radiation dose distribution images serve as the dataset for model training.

[0056] In one example, the dataset above, after preprocessing, can be divided into a training set, a validation set, and a test set. The image data in the training set, validation set, and test set can be either overlapping or non-overlapping. The training set is used for the model training process, the validation set is used for the model validation process, and the test set is used for the model testing process.

[0057] Indicatively, scanned training images correspond to Planning Target Volume (PTV) data. The planning target volume is a geometric concept linking the patient and machine coordinate systems, specifically used for treatment planning and execution. The planning target volume includes the Clinical Target Volume (CTV) itself, patient organ movement during irradiation (represented by the Internal Target Volume (ITV)), and the extent of tissue irradiation caused by factors such as daily positioning, changes in target position and volume during treatment, ensuring that the CTV receives the prescribed therapeutic dose.

[0058] To illustrate, the scan training images also correspond to organ-at-risk (OAR) data, which is the risk dose to normal organs surrounding the radiation-treated area.

[0059] As an illustration, during training, PTV data and OAR data are fed into the model in parallel with scanned training images.

[0060] Step 202: Train the first parameters in the initial network model based on the scanned training images and radiation dose distribution images to obtain at least two teacher network models.

[0061] For illustrative purposes, the initial network model can be a U-Net network structure, a U-NAS network structure, or other network structures, without any limitation here.

[0062] In this embodiment, the training of the teacher network model using the U-NAS framework is used as an example for illustration. The U-NAS framework proposed in this application follows the U-Net structure and has four downsampling units (Down Cell, DC) and four upsampling units (Up Cell, UC). The number of downsampling and upsampling units is illustrated only with four as an example; however, it can also consist of other numbers of sampling units.

[0063] Please refer to Figure 3The diagram shows a schematic of the U-NAS framework. The U-NAS framework 300 includes two convolutional layers 310, located at the beginning and end of the structure, respectively. It also includes four downsampling units 320 and four upsampling units 330. The downsampling units 320 are connected in series, and the four downsampling units 320 and the four upsampling units 330 are connected in series. The downsampling units 320 and the upsampling units 330 are connected in parallel through CAT units 340.

[0064] Schematic illustration: Each downsampling unit consists of at least one operation. The operations corresponding to each downsampling unit can be the same or different. Similarly, the operations constituting each upsampling unit can be the same or different, and no limitation is imposed here. Each downsampling unit consists of a certain number of operations, including but not limited to at least one of the following: convolution, squeeze-and-excitation convolution, dilated convolution, depthwise-separable convolution, max pooling, average pooling, trilinear interpolation, and identity operations.

[0065] Optionally, different initial network models can be obtained through different combinations of operations. For example, the downsampling unit in initial network model A includes operation 1, and the upsampling unit includes operation 2; the downsampling unit in initial network model B includes operation 3, and the downsampling unit includes operation 4. By training the parameters corresponding to each operation in the initial network, teacher network model A corresponding to initial network model A and teacher network model B corresponding to initial network model B are obtained.

[0066] Optionally, different initial network models may have the same combination of operations, but the weights of each operation in the combination may be different.

[0067] Indicative, such as Figure 4 As shown, it illustrates an exemplary network topology 400 for a downsampling unit, where pre401 is a preprocessing operation, N402 is a normal operation, D403 is a downsampling operation, and C404 is a connection operation. Figure 5 As shown, it illustrates an exemplary network topology 500 for an upsampling unit, where pre501 is a preprocessing operation, N502 is a regular operation, U503 is an upsampling operation, C504 is a connection operation, and post505 is an output operation.

[0068] Schematic illustration: The downsampling and upsampling units include basic computational units called Hybrid Modules (HMs). Each HM is a sum of different operations. There are four types of HMs: upsampling, downsampling, and upsampling. Figure 4 and Figure 5 The search space includes regular operations (N), down operations (D), up operations (U), and connection operations (C), which correspond to different combinations of operations in the search space. Illustratively, the operation combinations corresponding to each HM are shown in Table 1. All convolutions in Table 1 use 3x3x3 convolution kernels and employ the Convolution-Instance Normalization-ReLU (Rectified Linear Units) activation function. Furthermore, each HM uses 3x3x3 convolutions and 1x1x1 convolutions at both ends to control the number of channels.

[0069] Table 1

[0070] Standard operation Downward operation Up operation Connection operation direct connection Average pooling convolution Residual connection convolution Max pooling Squeeze-excited convolution No connection Squeeze-excited convolution convolution dilated convolution dilated convolution Squeeze-excited convolution Depthwise separable convolution Depthwise separable convolution dilated convolution Trilinear interpolation Depthwise separable convolution

[0071] To illustrate, the operation combinations are the same across the initial network models, but the weights corresponding to the operations in each HM are different within each initial network model.

[0072] In this embodiment, the first parameter of n initial network models is adjusted based on the scanned training image and the radiation dose distribution image, and n network structures are converged. Based on the n network structures, the target operation combination corresponding to the first parameter is determined from the operation combination. Based on the target operation combination, n teacher network models are obtained, where n≥2 and n is an integer.

[0073] Step 203: Information from at least two teacher network models is transferred to the student network model through adversarial learning, and the second parameter of the student network model is adjusted to obtain the dose prediction model.

[0074] In this embodiment, at least two teacher network models learned through U-NAS are used to perform knowledge distillation on the at least two teacher network models to assist in the training of the student network model. Knowledge distillation refers to using the supervised information output by the teacher network models as knowledge, and then using the supervised information learned by the student network model from the teacher network models as the distillation process. Supervised training using the teacher network models achieves the distillation purpose, ultimately resulting in a higher-performance and more accurate student model.

[0075] For illustrative purposes, the student network model is a single network model, and its network complexity is lower than that of the teacher network model. The student network model can be a U-Net network or other networks; no specific limitation is made here.

[0076] In this embodiment, the student network model is a U-Net network, which includes four downsampling units and four upsampling units. The downsampling units perform max pooling, and the upsampling units perform transposed convolution. Illustratively, the second parameter in the student network model is randomly initialized. Based on adversarial learning, knowledge distillation is performed on at least two teacher network models, transferring information from the teacher network models to the student network model to adjust the randomly initialized parameters. The trained student network model then generates a dose prediction model.

[0077] The second parameter of the student network model is trained using a second loss function. This second loss function is used to determine the error between the dose prediction result output by the student network model and the radiation dose distribution image during the training process. That is, the second loss function is obtained, and the second parameter of the student network model is adjusted based on the second loss function to obtain the dose prediction model.

[0078] In this embodiment, the training of the student network model includes a knowledge distillation part and an adversarial learning part. Therefore, the second loss function is jointly determined by the target loss function corresponding to the knowledge distillation part and the adversarial loss function corresponding to the adversarial learning part. The target loss function is used to constrain the outputs of the teacher network model and the student network model. Illustratively, the target loss function and the adversarial loss function are obtained, and then added together with the target weight configuration to obtain the first loss function.

[0079] Indicatively, the mean of predictions obtained from at least two teacher network models is used as auxiliary information for training the student network model. That is, the second parameter of the student network model is adjusted based on the second loss function and the mean of predictions to obtain the dose prediction model.

[0080] Step 204: The radiation dose distribution corresponding to the target scan image is predicted by the dose prediction model to obtain the radiation dose prediction image corresponding to the target scan image.

[0081] After training the student network model, a dose prediction model is obtained. This model is then used to predict the radiation dose distribution of the input target scan image, resulting in a corresponding radiation dose prediction image. The target scan image can be a medical image from a real-world application or a scan image from a test set.

[0082] In a schematic manner, target scan images and radiation effect data are acquired, including the planned target area corresponding to the tissue and organ regions in the target scan image; high-dimensional feature extraction is performed on the target scan image using a dose prediction model to obtain the tissue and organ features in the target scan image; feature restoration of the tissue and organ features is performed using the dose prediction model to determine the radiation dose corresponding to the tissue and organ in the target scan image; and a radiation dose prediction image is generated based on the radiation dose.

[0083] In summary, the dose prediction method provided in this application trains the first parameters of an initial network model based on scanned training images and radiation dose distribution images to obtain at least two teacher network models. Then, based on these at least two teacher network models, a student network model is trained. Adversarial learning is used to transfer information from the at least two teacher network models to the student network model through knowledge distillation, adjusting the second parameters of the student network model to obtain a dose prediction model. This dose prediction model has higher performance and accuracy, combining the simple structure of the student network model with the prediction accuracy of the teacher network model. When predicting the dose distribution of a target scanned image, it obtains more accurate prediction results, improving the reliability of dose distribution prediction.

[0084] Please refer to Figure 6 This document illustrates a dose prediction method according to an embodiment of this application. The embodiment illustrately describes the training process of a teacher network model, and the method includes:

[0085] Step 601: Obtain the scan training image and radiation dose distribution image.

[0086] There is a correspondence between the training scan image and the radiation dose distribution image. The radiation dose distribution image is used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the training scan image.

[0087] In this embodiment, the scan training images are CT images of past patients, and the radiation dose distribution images are clinical dose distribution images corresponding to the CT images. These scan training images and radiation dose distribution images serve as the dataset for model training.

[0088] Indicatively, scanned training images correspond to Planning Target Volume (PTV) data. The planning target volume is a geometric concept linking the patient and machine coordinate systems, specifically used for treatment planning and execution. The planning target volume includes the Clinical Target Volume (CTV) itself, patient organ movement during irradiation (represented by the Internal Target Volume (ITV)), and the extent of tissue irradiation caused by factors such as daily positioning, changes in target position and volume during treatment, ensuring that the CTV receives the prescribed therapeutic dose.

[0089] To illustrate, the scan training images also correspond to organ-at-risk (OAR) data, which is the risk dose to normal organs surrounding the radiation-treated area.

[0090] As an illustration, during training, PTV data and OAR data are fed into the model in parallel with scanned training images.

[0091] Step 602: Obtain the search space corresponding to the initial network model.

[0092] The search space includes combinations of operations used to compose the initial network model. In this embodiment, the search space is shown in Table 1.

[0093] Schematic, the initial network model is a U-NAS network with four downsampling units, four upsampling units, and convolutional layers at the beginning and end. The downsampling units are connected in series, as are the four downsampling units and four upsampling units. The downsampling and upsampling units are connected in parallel via CAT units. The network topology of the downsampling units has been described... Figure 4 As shown, the network drag structure of the upsampling unit has been... Figure 5 As shown, it will not be elaborated here. The HM that makes up the downsampling unit or upsampling unit is composed of multiple operations. Among them, the above-mentioned mixed operations are the combination of operations in the search space. That is, the regular operation (N) consists of direct connection, convolution, squeeze-activated convolution, dilation convolution, and depthwise separable convolution; the down operation (D) consists of average pooling, max pooling, convolution, squeeze-activated convolution, dilation convolution, and depthwise separable convolution; the up operation (U) consists of convolution, squeeze-activated convolution, dilation convolution, depthwise separable convolution, and trilinear interpolation; and the connection operation (C) consists of residual connection and no connection.

[0094] The training process for the initial network model includes a search process and a retraining process, wherein the model structure of the initial network model is determined based on the search space.

[0095] Step 603: Determine the n weight relationships corresponding to the operation combinations in the search space. The weight relationships correspond to the first parameter.

[0096] Where n≥2, and n is an integer.

[0097] Indicative, such as Figure 7 As shown, it illustrates the topology of the HM700, where O0, O1, ..., O N 710 refers to the various operations indicated in the search space. This structure also includes the first parameter α0, α1, ..., α... N 720. Among them, the operation O in each HM i The weights are determined by the first parameter α. i ∈α is determined, 0≤i≤N, where N and i are integers. The logistic regression function (softmax) corresponding to HM is shown in Equation 1, where, Mark O i Contributions to HM.

[0098] Formula 1:

[0099] In a schematic way, the parameters α and weights w are learned alternately to obtain n weight relationships corresponding to the operation combination.

[0100] Step 604: Based on the model structure indicated by the n weight relationships corresponding to the operation combination, generate n initial network models.

[0101] Based on the HMs indicated in the search space and the operations O in the HMs i The corresponding weights can generate n initial network models, where the operation structures of the n initial network models are the same, but the operation weights are different.

[0102] Step 605: Adjust the first parameters of the n initial network models based on the scanned training images and radiation dose distribution images, and converge to obtain n network structures.

[0103] The scanned training images are input into various initial network models. The search process is repeated using different initial network models to converge to different local optima and obtain different network structures.

[0104] Step 606: Based on the n network structures, determine the target operation combination corresponding to the first parameter from the operation combinations.

[0105] In each search process, the i-th initial network model can determine the operation with the highest probability corresponding to the first parameter α in each HM, obtain the target operation combination, and obtain n teacher network models based on the target operation combination.

[0106] Step 607: Determine n teacher network models by combining the target operations.

[0107] The i-th network structure is determined based on the operation with the highest probability, that is, the downsampling unit and the upsampling unit are replaced by the operation with the highest probability.

[0108] For example, in the network structure corresponding to the first parameter A, the operations corresponding to the four downsampling units are "convolution → max pooling → max pooling → average pooling", and the operations corresponding to the four upsampling units are "dilated convolution → trilinear interpolation → convolution → convolution". In the network structure corresponding to the first parameter B, the operations corresponding to the four downsampling units are "max pooling → squeeze-excited convolution → convolution → average pooling", and the operations corresponding to the four upsampling units are "dilated convolution → trilinear interpolation → squeeze-excited convolution → convolution".

[0109] Step 608: Input the scanned training images and radiation dose distribution images into n teacher network models, and iteratively train the n teacher network models.

[0110] After the search process is complete, each HM retains only the most probable operation based on the first parameter α, and then replaces DC and UC with the optimal search structure. Next, the network needs to be retrained on the training dataset, that is, the scanned training images and radiation dose distribution images are input again into the n teacher network models, and the teacher network models are trained to convergence using the first loss function to adjust the model parameters of the teacher network models.

[0111] Indicatively, a first loss function is obtained, which is used to determine the error between the dose prediction result output by the teacher network model and the radiation dose distribution image during the training process; based on the scanned training image, a radiation dose prediction image corresponding to the i-th teacher network model is generated, where 0 < i ≤ n; the radiation dose prediction image and the radiation dose distribution image are input into the first loss function to determine the first loss value; in response to the first loss value successfully matching the first numerical requirement, the parameter corresponding to the first loss value is determined as the model parameter of the i-th teacher network model; or, in response to the first loss value failing to match the first numerical requirement, the model parameter of the i-th teacher network model is adjusted for iterative training.

[0112] In one example, the first loss function L dose As shown in Formula 2, where, The dose prediction result is derived from the teacher network model, and y is the target dose indicated in the radiation dose distribution image.

[0113] Formula 2:

[0114] In summary, the dose prediction method provided in this application determines the structure of the initial network model by combining operations within the search space, and trains the initial network model according to the different weight relationships corresponding to the operation combinations, converging to obtain multiple locally optimal network structures. From these network structures, multiple corresponding teacher network models are determined, and iterative training is performed using training data to obtain a teacher network model that can provide sufficient information for the student network model. This results in a dose prediction model trained using both the teacher and student network models, exhibiting higher performance and accuracy. It combines the simple structure of the student network model with the prediction accuracy of the teacher network model, achieving more accurate prediction results when predicting the dose distribution of a target scan image, thus improving the reliability of dose distribution prediction.

[0115] Please refer to Figure 8 This document illustrates a dose prediction method according to an embodiment of this application. In this embodiment, the training process of a student network model is illustratively described. The training process includes a knowledge distillation part and an adversarial learning part. The knowledge distillation part includes steps 801-806, and the adversarial learning part includes steps 807-810. The training processes of the knowledge distillation part and the adversarial learning part can be performed simultaneously. The method includes:

[0116] Step 801: Obtain the mean of predictions for the scanned training images from at least two teacher network models.

[0117] The scanned training images are input into at least two teacher network models. Each teacher network model outputs radiation dose prediction data for the scanned training images, which can be illustrated using a dose distribution image. The prediction data output by all teacher network models are averaged to obtain the prediction mean, which serves as auxiliary information for training the student network model.

[0118] Step 802: Obtain the first intermediate output of the k-th teacher unit in the i-th teacher network model.

[0119] The teacher network model consists of m teacher units, and the student network model consists of m student units, where m is a positive integer and 0 < k ≤ m.

[0120] In this embodiment, the teacher network model includes four downsampled teacher units and four upsampled teacher units, and the student network model includes four downsampled student units and four upsampled student units. The downsampled student units in the student network model perform max pooling, and the upsampled student units perform transposed convolution.

[0121] Step 803: Obtain the second intermediate output of the k-th student unit in the student network model.

[0122] Step 804: Determine the target loss value between the first intermediate output and the second intermediate output based on the target loss function.

[0123] Schematic, in this embodiment of the application, a target loss function is used to constrain the outputs of the teacher network model and the student network model. In one example, the target loss function... As shown in Formula 3, where, This represents the first intermediate output of the k-th teacher unit. represents the second intermediate output of the k-th student unit, and M represents the total number of teacher network models.

[0124] Formula 3:

[0125] Step 805: In response to the successful match between the target loss value and the second numerical requirement, the parameter corresponding to the target loss value is determined as the candidate unit parameter of the kth student unit in the student network model.

[0126] Indicatively, this second numerical requirement can be that the target loss value is less than a preset value. That is, if the student units in the student network model are trained to convergence during the training process, then the parameters corresponding to the target loss value are determined as candidate unit parameters. Indicatively, gradient descent can be used to adjust the unit parameters.

[0127] Step 806: In response to the failure of the target loss value to match the second numerical requirement, adjust the unit parameters of the k-th student unit in the student network model and perform iterative training.

[0128] When the target loss value does not meet the second numerical requirement, the unit parameters in the student network model are adjusted and iterative training is performed until the second numerical requirement is met.

[0129] Step 807: Obtain the discriminator D corresponding to the k-th student unit. k .

[0130] Furthermore, adversarial learning needs to be incorporated into the knowledge distillation process to force the student and teacher network models to produce more similar features. Among these, the discriminator D... k The output is used to distinguish between the teacher network model and the student network model. Specifically, for the k-th student unit, a discriminator D is designed. k The student network model in the discriminator D k In adversarial learning, it is necessary to generate outputs that are more similar to the teacher's network model, so that the discriminator D... kIt is impossible to distinguish its output from the output of the teacher network model; the two are mutually antagonistic and mutually reinforcing.

[0131] Step 808: Determine the discriminator D based on the adversarial loss function. k The adversarial loss value between the kth student unit and the kth student unit.

[0132] In one example, adversarial loss function As shown in Formula 4, where I k ~P T I represents the first intermediate output of the k-th teacher unit in the teacher network model. k ~P S This represents the second intermediate output of the k-th student unit in the student network model. This represents the expected value of the distribution function.

[0133] Formula 4:

[0134] Step 809: In response to the successful match between the adversarial loss value and the third numerical requirement, the parameter corresponding to the adversarial loss value is determined as the candidate unit parameter of the k-th student unit in the student network model.

[0135] Indicatively, this third numerical requirement can be that the adversarial loss value is less than a preset value. That is, if the student units in the student network model are trained to convergence during the training process, then the parameters corresponding to the adversarial loss value are determined as candidate unit parameters. Indicatively, gradient descent can be used to adjust the unit parameters.

[0136] Step 810: In response to the failure of the adversarial loss value to match the third numerical requirement, adjust the unit parameters of the k-th student unit in the student network model and perform iterative training.

[0137] Step 811: Determine the unit parameters of the m student units of the student network model based on the candidate unit parameters to obtain the dose prediction model.

[0138] In some embodiments, the knowledge distillation process and the adversarial learning process during the training of the student network model are performed simultaneously, i.e., integrated into the same training process. Schematic, the target loss function and the adversarial loss function can be integrated into a second loss function, in which the target loss function and the adversarial loss function have different weights.

[0139] In some embodiments, the training processes of the teacher network model and the student network model are executed concurrently. That is, after training the first parameters of the teacher network model, the second parameters of the student network model are trained. By integrating Equations 2, 3, and 4, a comprehensive loss function is obtained, which is used to train the entire network architecture. (Illustrative example: comprehensive loss function) As shown in Formula 5, where λ1 represents the first loss function. The weighting factor, λ2 represents the target loss function. The weighting factor, λ3 represents the adversarial loss function. The weighting factors, in one example, are λ1 = 1, λ2 = 0.05, and λ3 = 0.01.

[0140] Formula 5:

[0141] In one example, such as Figure 9 The diagram illustrates the LE-NAS framework 900 provided in this application, which includes a U-NAS automatic search model 910, a teacher network model module 920, and a student network model module 930. CT data 901 is input as training data into the model for training, and the model can output radiation dose prediction images 902. Schematic, the CT data 901 includes CT images, PTV, and OAR.

[0142] In summary, the dose prediction method provided in this application trains a student network model based on at least two teacher network models. Using adversarial learning, information from the at least two teacher network models is transferred to the student network model through knowledge distillation to adjust the second parameter of the student network model, resulting in a dose prediction model. This model has higher performance and accuracy, combining the simple structure of the student network model with the prediction accuracy of the teacher network model. When predicting the dose distribution of a target scan image, it yields more accurate prediction results and improves the reliability of dose distribution prediction.

[0143] Schematic illustration: The dose prediction method provided in this application is tested on the publicly available dataset OpenKBP, which comes from the 2020 AAPM (American Academy of Pain Medicine) international competition. This dataset contains 340 CT images of head and neck cancer patients undergoing radiotherapy. The competition organizers divided the data into a training set of 200 images, a validation set of 40 images, and a test set of 100 images. Based on the AAPM competition, the model performance is evaluated using dose error and Dose Volume Histogram (DVH) error, where dose error represents the error between the model's predicted radiation dose and the true value, and DVH error is a performance evaluation method for radiation dose prediction in medicine.

[0144] Please refer to Table 2, which shows the experimental results of the dose prediction model provided in this application on the above dataset and the comparison data with other models. The errors include Mean-Square Error (MSE) and Mean Absolute Error (MAE). Rows 1-5 of Table 2 show the results of the top five methods in the AAPM competition, rows 6-9 show the most mainstream existing models, and rows 10-11 show the optimal model automatically searched by the U-NAS proposed in this application and the results of LE-NAS. It can be seen that in the single-model results, the LE-NAS proposed in this application achieved the best results. Furthermore, in the comparison of cascaded models, the two-stage cascaded model of the U-NAS proposed in this application also outperforms existing methods. Finally, in the five model ensemble experiments, the model ensemble results obtained by the automatic search of the U-NAS network are significantly better than existing methods and also surpass the results of the first-place winner in AAPM. Therefore, the dose prediction model proposed in this application obtains more accurate prediction results when predicting the dose distribution of target scan images, improving the reliability of dose distribution prediction.

[0145] Table 2

[0146]

[0147]

[0148] Please refer to Figure 10 The present application provides a schematic diagram of the structure of a dose prediction device according to an exemplary embodiment, the device comprising:

[0149] The acquisition module 1010 is used to acquire scan training images and radiation dose distribution images, wherein the radiation dose distribution images are used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the scan training images;

[0150] Training module 1020 is used to train the first parameters in the initial network model based on the scanned training image and the radiation dose distribution image to obtain at least two teacher network models;

[0151] The training module 1020 is also used to transfer information from the at least two teacher network models to the student network model through adversarial learning, adjust the second parameter of the student network model, and obtain a dose prediction model.

[0152] The prediction module 1030 is used to predict the radiation dose distribution corresponding to the target scan image through the dose prediction model, so as to obtain the radiation dose prediction image corresponding to the target scan image.

[0153] In an optional embodiment, such as Figure 11 As shown, the training module 1020 further includes:

[0154] The acquisition unit 1021 is used to acquire the search space corresponding to the initial network model, wherein the search space includes the combination of operations used to compose the initial network model;

[0155] The determining unit 1022 is used to determine the model structure of the initial network model based on the search space.

[0156] In an optional embodiment, the determining unit 1022 is further configured to determine n weight relationships corresponding to the operation combination in the search space, where n≥2 and n is an integer, and the weight relationships correspond to the first parameter;

[0157] The training module 1020 further includes:

[0158] The first generation unit 1023 is used to generate n initial network models based on the model structure indicated by the n weight relationships corresponding to the operation combination.

[0159] In an optional embodiment, the first generation unit 1023 is further configured to adjust the first parameters in the n initial network models based on the scan training image and the radiation dose distribution image, and converge to obtain n network structures;

[0160] The determining unit 1022 is further configured to determine, based on the n network structures, a target operation combination corresponding to the first parameter from the operation combinations;

[0161] The determining unit 1022 is also used to obtain n teacher network models based on the target operation combination.

[0162] In an optional embodiment, the determining unit 1022 is further configured to determine the n teacher network models by the target operation combination;

[0163] The training module 1020 further includes:

[0164] The training unit 1024 is used to input the scan training image and the radiation dose distribution image into the n teacher network models and perform iterative training on the n teacher network models.

[0165] In an optional embodiment, the acquisition unit 1021 is further configured to acquire a first loss function, which is used to determine the error between the dose prediction result output by the teacher network model during the training process and the radiation dose distribution image.

[0166] The first generation unit 1023 is further configured to generate a radiation dose prediction image corresponding to the i-th teacher network model based on the scanned training image, where 0 < i ≤ n;

[0167] The determining unit 1022 is further configured to input the radiation dose prediction image and the radiation dose distribution image into the first loss function to determine the first loss value;

[0168] The determining unit 1022 is further configured to, in response to the first loss value successfully matching the first numerical requirement, determine the parameter corresponding to the first loss value as the model parameter of the i-th teacher network model; or, in response to the first loss value failing to match the first numerical requirement, adjust the model parameter of the i-th teacher network model for iterative training.

[0169] In an optional embodiment, the acquisition unit 1021 is further configured to acquire a second loss function, which is used to determine the error between the dose prediction result output by the student network model during training and the radiation dose distribution image.

[0170] The determining unit 1022 is further configured to adjust the second parameter of the student network model based on the second loss function to obtain the dose prediction model.

[0171] In an optional embodiment, the acquisition unit 1021 is further configured to acquire a target loss function, which is used to constrain the outputs of the teacher network model and the student network model;

[0172] The acquisition unit 1021 is further configured to acquire an adversarial loss function, which is used to control the output of the student network model to approach the output of the teacher network model.

[0173] The determining unit 1022 is further configured to add the target loss function and the adversarial loss function with target weight configuration to obtain the first loss function.

[0174] In an optional embodiment, the acquisition unit 1021 is further configured to acquire the mean of the predictions made by the at least two teacher network models for the scanned training image;

[0175] The determining unit 1022 is further configured to adjust the second parameter of the student network model based on the second loss function and the predicted mean, so as to obtain the dose prediction model.

[0176] In an optional embodiment, the teacher network model includes m teacher units, and the student network model includes m student units, where m is a positive integer;

[0177] The acquisition unit 1021 is also used to acquire the first intermediate output of the kth teacher unit in the i-th teacher network model, where 0 < k ≤ m;

[0178] The acquisition unit 1021 is further configured to acquire the second intermediate output of the kth student unit in the student network model;

[0179] The determining unit 1022 is further configured to determine a target loss value between the first intermediate output and the second intermediate output based on the target loss function;

[0180] The determining unit 1022 is further configured to, in response to the successful matching of the target loss value with the second numerical requirement, determine the parameter corresponding to the target loss value as the candidate unit parameter of the kth student unit in the student network model; or, in response to the failure to match the target loss value with the second numerical requirement, adjust the unit parameter of the kth student unit in the student network model for iterative training.

[0181] The determining unit 1022 is further configured to determine the unit parameters of the m student units of the student network model based on the candidate unit parameters, thereby obtaining the dose prediction model.

[0182] In an optional embodiment, the acquisition unit 1021 is further configured to acquire the discriminator D corresponding to the k-th student unit. k The discriminator D k The output used to distinguish between the teacher network model and the student network model;

[0183] The determining unit 1022 is further configured to determine the discriminator D based on the adversarial loss function. kThe adversarial loss value between the kth student unit and the student unit;

[0184] The determining unit 1022 is further configured to, in response to the adversarial loss value successfully matching the third numerical requirement, determine the parameter corresponding to the adversarial loss value as the candidate unit parameter of the kth student unit in the student network model; or, in response to the adversarial loss value failing to match the third numerical requirement, adjust the unit parameter of the kth student unit in the student network model for iterative training.

[0185] In an optional embodiment, the acquisition module 1010 is further configured to acquire the target scan image and radiation effect data, wherein the radiation effect data includes the planned target area corresponding to the tissue and organ region in the target scan image;

[0186] The prediction module 1030 further includes:

[0187] Feature extraction unit 1031 is used to perform high-dimensional feature extraction on the target scan image through the dose prediction model to obtain tissue and organ features in the target scan image;

[0188] The feature restoration unit 1032 is used to restore the features of the tissues and organs through the dose prediction model, and determine the radiation dose corresponding to the tissues and organs in the target scan image.

[0189] The second generation unit 1033 is used to generate the radiation dose prediction image based on the radiation dose.

[0190] In summary, the dose prediction device provided in this application trains the first parameters of the initial network model based on the scanned training image and the radiation dose distribution image to obtain at least two teacher network models. Then, based on these at least two teacher network models, a student network model is trained. Using adversarial learning, information from the at least two teacher network models is transferred to the student network model through knowledge distillation to adjust the second parameters of the student network model, resulting in a dose prediction model. This dose prediction model has higher performance and accuracy, combining the simple structure of the student network model with the prediction accuracy of the teacher network model. When predicting the dose distribution of the target scanned image, it obtains more accurate prediction results, improving the reliability of dose distribution prediction.

[0191] It should be noted that the dose prediction device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the dose prediction device and the dose prediction method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0192] Figure 12 A schematic diagram of the structure of a server provided in an exemplary embodiment of this application is shown.

[0193] Specifically, server 1200 includes a central processing unit (CPU) 1201, a system memory 1204 including random access memory (RAM) 1202 and read-only memory (ROM) 1203, and a system bus 1205 connecting the system memory 1204 and the CPU 1201. Server 1200 also includes a mass storage device 1206 for storing the operating system 1213, application programs 1214, and other program modules 1215.

[0194] Mass storage device 1206 is connected to central processing unit 1201 via a mass storage controller (not shown) connected to system bus 1205. Mass storage device 1206 and its associated computer-readable media provide non-volatile storage for server 1200. That is, mass storage device 1206 may include computer-readable media (not shown) such as hard disk or compact disc read-only memory (CD-ROM) drives.

[0195] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 1204 and mass storage device 1206 described above can be collectively referred to as memory.

[0196] According to various embodiments of this application, server 1200 can also be connected to a remote computer on a network, such as the Internet. That is, server 1200 can be connected to network 1212 via network interface unit 1211 connected to system bus 1205, or network interface unit 1211 can be used to connect to other types of networks or remote computer systems (not shown).

[0197] The aforementioned memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.

[0198] Embodiments of this application also provide a computer device, which includes a processor and a memory. The memory stores at least one instruction, at least one program, code set, or instruction set. The at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the dose prediction method provided in the above-described method embodiments.

[0199] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the dose prediction method provided in the above-described method embodiments.

[0200] Embodiments of this application also provide a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the dose prediction methods described in the above embodiments.

[0201] Optionally, the computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM). The sequence numbers of the embodiments in this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0202] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0203] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A dose prediction method, characterized in that, The method includes: Acquire scan training images and radiation dose distribution images, wherein the radiation dose distribution images are used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the scan training images; At least two initial network models are obtained based on the weights corresponding to each operation in at least two operation combinations, and the at least two initial network models correspond one-to-one with the at least two operation combinations; wherein, different initial network models correspond to different operation combinations, and / or, different initial network models correspond to the same operation combinations, but the weight combinations between operations in the same operation combinations are different; Based on the scanned training image and the radiation dose distribution image, the first parameter in the at least two initial network models is adjusted to converge at least two network structures. Based on the at least two network structures, the target operation combination corresponding to the first parameter is determined from the operation combination. Based on the aforementioned combination of target operations, at least two teacher network models are obtained; Information from the at least two teacher network models is transmitted to the student network model through adversarial learning, and the second parameter of the student network model is adjusted to obtain the dose prediction model. The radiation dose distribution corresponding to the target scan image is predicted by the dose prediction model to obtain the radiation dose prediction image corresponding to the target scan image.

2. The method according to claim 1, characterized in that, Before training the first parameters of the initial network model based on the scanned training image and the radiation dose distribution image to obtain at least two teacher network models, the process further includes: Obtain the search space corresponding to the initial network model, wherein the search space includes the combination of operations used to compose the initial network model; Based on the search space, the model structure of the initial network model is determined.

3. The method according to claim 2, characterized in that, The step of determining the model structure of the initial network model based on the search space includes: Determine at least two weight relationships corresponding to the operation combination in the search space, wherein the weight relationships correspond to the first parameter; Based on the model structure indicated by at least two weight relationships corresponding to the operation combination, at least two initial network models are generated.

4. The method according to claim 3, characterized in that, Based on the combination of the target operations, at least two teacher network models are obtained, including: The at least two teacher network models are determined by the combination of the target operations; The scan training image and the radiation dose distribution image are input into the at least two teacher network models, and the at least two teacher network models are iteratively trained.

5. The method according to claim 4, characterized in that, The iterative training of the at least two teacher network models includes: Obtain a first loss function, which is used to determine the error between the dose prediction result output by the teacher network model during the training process and the radiation dose distribution image; Based on the scanned training images, a radiation dose prediction image corresponding to the i-th teacher network model is generated, where 0 < i ≤ n, and n is the total number of teacher network models. The radiation dose prediction image and the radiation dose distribution image are input into the first loss function to determine the first loss value; In response to a successful match between the first loss value and the first numerical requirement, the parameter corresponding to the first loss value is determined as the model parameter of the i-th teacher network model; or, in response to a failure to match between the first loss value and the first numerical requirement, the model parameter of the i-th teacher network model is adjusted for iterative training.

6. The method according to any one of claims 1 to 5, characterized in that, The step of transmitting information from at least two teacher network models to a student network model through adversarial learning, and adjusting the second parameter of the student network model to obtain a dose prediction model, includes: Obtain a second loss function, which is used to determine the error between the dose prediction result output by the student network model during training and the radiation dose distribution image; Based on the second loss function, the second parameter of the student network model is adjusted to obtain the dose prediction model.

7. The method according to claim 6, characterized in that, The process of obtaining the second loss function includes: Obtain a target loss function, which is used to constrain the outputs of the teacher network model and the student network model; Obtain an adversarial loss function, which is used to control the output of the student network model to approach the output of the teacher network model; The target loss function and the adversarial loss function are added together with the target weight configuration to obtain the second loss function.

8. The method according to claim 7, characterized in that, The step of adjusting the second parameters of the student network model based on the second loss function to obtain the dose prediction model includes: Obtain the mean of the predictions made by the at least two teacher network models for the scanned training images; Based on the second loss function and the predicted mean, the second parameter of the student network model is adjusted to obtain the dose prediction model.

9. The method according to claim 8, characterized in that, The teacher network model includes m teacher units, and the student network model includes m student units, where m is a positive integer; The step of adjusting the second parameters of the student network model based on the second loss function and the predicted mean to obtain the dose prediction model includes: Obtain the first intermediate output of the k-th teacher unit in the i-th teacher network model, where 0 < k ≤ m; Obtain the second intermediate output of the k-th student unit in the student network model; The target loss value between the first intermediate output and the second intermediate output is determined based on the target loss function; In response to a successful match between the target loss value and the second numerical requirement, the parameter corresponding to the target loss value is determined as the candidate unit parameter of the kth student unit in the student network model; or, in response to a failure to match between the target loss value and the second numerical requirement, the unit parameter of the kth student unit in the student network model is adjusted for iterative training. Based on the candidate unit parameters, the unit parameters of the m student units in the student network model are determined to obtain the dose prediction model.

10. The method according to claim 9, characterized in that, Before determining the unit parameters of the m student units of the student network model based on the candidate unit parameters to obtain the dose prediction model, the method further includes: Obtain the discriminator D corresponding to the k-th student unit. k The discriminator D k The output used to distinguish between the teacher network model and the student network model; The discriminator D is determined based on the adversarial loss function. k The adversarial loss value between the kth student unit and the student unit; In response to a successful match between the adversarial loss value and the third numerical requirement, the parameter corresponding to the adversarial loss value is determined as the candidate unit parameter of the kth student unit in the student network model; or, in response to a failure to match between the adversarial loss value and the third numerical requirement, the unit parameter of the kth student unit in the student network model is adjusted for iterative training.

11. The method according to any one of claims 1 to 5, characterized in that, The step of predicting the radiation dose distribution corresponding to the target scan image using the dose prediction model to obtain the radiation dose prediction image corresponding to the target scan image includes: Acquire the target scan image and radiation effect data, wherein the radiation effect data includes the planned target area corresponding to the tissue and organ region in the target scan image; The dose prediction model is used to extract high-dimensional features from the target scan image to obtain the tissue and organ features in the target scan image; The radiation dose corresponding to the tissue or organ in the target scan image is determined by using the dose prediction model to restore the features of the tissue or organ. The radiation dose prediction image is generated based on the radiation dose.

12. A dose prediction device, characterized in that, The device includes: An acquisition module is used to acquire scan training images and radiation dose distribution images, wherein the radiation dose distribution images are used to indicate the radiation dose distribution when radiotherapy is performed on the tissues and organs in the scan training images; A training module is used to obtain at least two initial network models based on the weights corresponding to each operation in at least two operation combinations, wherein the at least two initial network models correspond one-to-one with the at least two operation combinations; wherein different initial network models correspond to different operation combinations, and / or different initial network models correspond to the same operation combinations, but the weight combinations between operations in the same operation combinations are different; the first parameter in the at least two initial network models is adjusted based on the scanned training image and the radiation dose distribution image, and converges to obtain at least two network structures; a target operation combination corresponding to the first parameter is determined from the operation combinations according to the at least two network structures; and at least two teacher network models are obtained based on the target operation combination. The training module is also used to transfer information from the at least two teacher network models to the student network model through adversarial learning, adjust the second parameter of the student network model, and obtain a dose prediction model. The prediction module is used to predict the radiation dose distribution corresponding to the target scan image through the dose prediction model, so as to obtain the radiation dose prediction image corresponding to the target scan image.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processor to implement the dose prediction method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or instruction set is loaded and executed by a processor to implement the dose prediction method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium, which a processor reads from and executes to implement the dose prediction method as described in any one of claims 1 to 11.

Images