AI auxiliary radiotherapy planning method based on diffusion model

Through the deep diffusion generation model and 2.5DCT data slicing strategy, the problem of lack of high-frequency details and high computational burden of radiotherapy plans in the existing technology is solved, and efficient and accurate radiotherapy dose distribution prediction and individualized customization are achieved, suitable for clinical applications. .

CN119007934APending Publication Date: 2024-11-22RIZHAO PEOPLES HOSPITAL OF SHANDONG PROVINCE
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411083713.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing deep learning-based radiotherapy planning methods lack high-frequency detailed predictions, the calculation burden is high, and it is difficult to apply to automatic radiotherapy planning for clinical, especially complex cancers, and it is difficult to consider individual differences, which makes it difficult to achieve precise dose distribution.

Method used

Using an AI-assisted radiotherapy planning method based on deep diffusion generation model, a network model including structural encoder, dose encoder and noise predictor is introduced, combined with the 2.5DCT data slicing strategy, a step-by-step diffusion generation mechanism is introduced to optimize the model structure and performance. , include individualized information, reduce the amount of three-dimensional neural network parameters, and improve computing efficiency.

Benefits of technology

It realizes accurate prediction of high-frequency details, improves the authenticity and computing efficiency of dose distribution, is suitable for resource-constrained medical environments, and can individually customize radiotherapy plans to protect normal tissues and improve the accuracy of tumor dose.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007934A_ABST
    Figure CN119007934A_ABST
Patent Text Reader

Abstract

The invention discloses an AI auxiliary radiotherapy planning method based on a diffusion model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: S1, collecting and processing a data set: collecting a positioning CT image, sketching information and clinical information of a cancer patient, and establishing a cancer radiotherapy dose database; s2, a dose prediction model: establishing a deep learning network based on a deep diffusion generation model based on task and data characteristics, and obtaining a model with higher generalization ability by using the characteristics of strong generation ability, high result authenticity and rich details of the deep learning network; and S3, in combination with a 2.5 D input strategy of individualized information, a 2.5 D mode is adopted, an efficient and lightweight high-precision model is realized, and meanwhile, the model established in the mode is more easily applied to clinic. According to the method, the AI technology is introduced, the slicing strategy of 2.5 DCT data is combined, and the structure and performance of the model are optimized by combining the two-dimensional and three-dimensional characteristics.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted radiotherapy planning method based on a diffusion model. Background Art

[0002] Deep learning-based methods also face the following challenges: Due to the averaging effect of the loss function, these methods usually lack high-frequency detail prediction. Because some networks use 3D inputs or large and complex networks, their computational burden is large and difficult to apply in clinical practice. The results of current methods are still difficult to directly apply in clinical practice, especially for automatic radiotherapy planning for complex cancers.

[0003] Due to the averaging effect of the loss function, these methods usually lack high-frequency detail predictions and are difficult to meet the requirements for accurate radiotherapy dose distribution: the dose distribution of radiotherapy plans is usually required to be relatively accurate. For example, for breast cancer plans, the dose distribution of endangered organs such as the heart, nerves, and lungs has a significant impact on the patient's prognosis and quality of life, while key information such as the uniformity of the target area and the maximum amount determine the patient's actual experience during treatment and the toxicity reaction. The high-frequency details of the dose distribution are crucial for the prediction of the maximum dose point. However, the commonly used loss functions and network structures usually ignore high-frequency details.

[0004] Since some networks use 3D input or large and complex networks, their computational burden is large and difficult to apply in clinical practice: existing methods mostly predict the dose of one layer or the entire three-dimensional image. Predicting one layer will lack three-dimensional information, making it difficult to achieve the required accuracy of dose prediction in the axial position. Predicting the entire three-dimensional image will cause too much computational burden and difficult to be directly applied in clinical practice.

[0005] Since the current implementation of automatic planning usually adopts a standardized method, that is, a consistent field angle is given by default, it is difficult to cope with complex and changing clinical situations: the consistent standard method is difficult to take into account the individual differences between different patients, and it is difficult to accurately take these personalized limits into consideration. Summary of the invention

[0006] The purpose of the present invention is to provide an AI-assisted radiotherapy planning method based on a diffusion model to solve the problems raised in the above background technology.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: an AI-assisted radiotherapy planning method based on a diffusion model, the method comprising the following steps:

[0008] S1. Dataset collection and processing: Collect positioning CT images, contouring information and clinical information of cancer patients to establish a cancer radiotherapy dose database;

[0009] S2. Constructing a dose prediction network model: Based on the characteristics of the task and data, a deep learning network based on the deep diffusion generative model is established, and a model with generalization ability is obtained by taking advantage of its strong generative ability, high authenticity of results and rich details;

[0010] S3. 2.5D input strategy combined with individualized information: Combining the slicing strategy and individualized information of 2.5D CT data, it is difficult to accurately take these personalized limits into consideration by combining the characteristics of two-dimensional and three-dimensional, embedding individualized information to select more appropriate optimization parameters and optimize the structure and performance of the model.

[0011] In step S1,

[0012] The collection and processing of the data set specifically includes the following steps:

[0013] S2-1. Collect 100 cancer cases, including patient disease course information, localization CT images, and delineation information of target areas and organs at risk;

[0014] S2-2. Clean the data to remove invalid and erroneous data, including missing values, format errors, and type errors;

[0015] S2-3, preprocess the data, including image format conversion, contrast adjustment, image standardization and data enhancement;

[0016] S2-4. Create a database to store data in paired and unpaired structures, and record the information of each patient.

[0017] The dose prediction network model includes a structure encoder, a dose encoder and a noise predictor; the model introduces a step-by-step diffusion generation mechanism to solve the problem of mode collapse and instability during training.

[0018] The structural encoder consists of five consecutive steps. In the first step, an embedding layer for parsing individual information is added. Except for the last step, each step contains a residual block and a downsampling block. The residual block consists of two convolution units, each of which includes a 3×3 convolution layer Conv, a group normalization layer GN and an activation function to maintain the residual connection and prevent the problem of gradient disappearance during training. The downsampling block consists of a 3×3 convolution layer Conv with a step size of 2.

[0019] The dose encoder consists of five consecutive steps, which are responsible for extracting features from the dose image, merging structural features, and inputting them into the noise predictor; the input content is encoded into a feature map through a convolutional layer, and then the two sets of feature maps are fused by element-by-element addition, allowing the structural information extracted in the structural encoder to be passed to the dose encoder; the next two downsampling steps continue to use addition operations to complete the fusion of information, while the following three steps use a cross-attention mechanism to achieve similarity-based structural guidance at a deeper level.

[0020] The noise predictor restores the feature representation captured by the dose encoder to the final output, where the skip connections are preserved for the reuse and aggregation of multi-level features.

[0021] The diffusion generation mechanism is specifically as follows:

[0022] The diffusion generation process is divided into two steps, namely forward diffusion generation and backward diffusion generation. In the forward diffusion generation process, the model adjusts the size and direction of the added noise according to the difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the target data distribution. In the backward diffusion generation process, the model adjusts the size and direction of the removed information according to the difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the noise sample.

[0023] The specific update rules of the diffusion generation mechanism are as follows:

[0024]

[0025] Among them, y t-1 and t Represents the images in the diffusion generation process, all of which are images at different noise levels, y t-1 Generate y by forward diffusion t ,y t Generate y by back diffusion t-1 ; α t and γ t are two control parameters, α t In this model, γ is used to adjust the noise level. t Used to control the way noise is added; ε t,θ is the noise term at time step t, which is a function of the model output, including parameters θ; z t is a random noise term, z t Obey standard normal distribution;

[0026] α t and γ t The parameters control the intensity and distribution of the noise, helping the model to transition smoothly between different time steps;

[0027] Noise term ε t,θ and z t The randomness of the generation process is guaranteed, allowing the model to generate diverse images.

[0028] The purpose of the dose prediction model is to train the noise predictor and the structure encoder so that the noise predicted in the reverse process is similar to the noise added in the forward process; to minimize the expected loss, the objective function is defined as follows:

[0029]

[0030] Among them, θ represents the parameter to be optimized; E (x,y) Represents the expected value of the data distribution (x, y); E ε,γ represents the expected value of the noise ε and the random variable γ; f(x feature ,y t ,γ t) is a feature x feature , the target variable y t and the random variable γ t Function of ε t is the noise term, ε t It follows a normal distribution.

[0031] In order to reduce the number of parameters of the three-dimensional neural network model and prevent the two-dimensional neural network from losing the correlation information between different layers when processing two-dimensional image slices, this method adopts a slicing strategy for 2.5DCT data; this strategy aims to optimize the structure and performance of the model by combining the characteristics of two-dimensional and three-dimensional; for the dose prediction of each layer, the upper and lower 5 layers are taken as input, the information in each 2.5D layer is connected in the channel dimension, and the obtained multi-channel data is used as model input.

[0032] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0033] 1. The present invention has excellent high-frequency detail performance: the deep diffusion generation model can capture more high-frequency details, making the predicted dose distribution superior to existing methods in terms of details and authenticity;

[0034] 2. The present invention adopts a 2.5D method and a slicing strategy of 2.5D CT data to optimize the structure and performance of the model by combining the characteristics of two-dimensional and three-dimensional. It achieves a compromise between the axis position and the computational burden, which ensures the computational efficiency without affecting the prediction accuracy. The lightweight design enables the system to operate efficiently in a resource-constrained environment and is suitable for more medical institutions.

[0035] 3. The present invention adds field angle information and personal dose limits to the automatic planning system, realizing personalized AI-assisted planning. By considering the field angle and incorporating the limits of important tissues, we can more accurately control the distribution of doses, thereby better protecting normal tissues and increasing the dose of tumors.

[0036] 4. Based on the characteristics of the task and data, a deep learning network based on the deep diffusion generative model is established, and its strong generation ability, high authenticity of results and rich details are used to obtain a model with generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0038] Figure 1 It is a radiotherapy dose prediction network based on the deep diffusion generation model of the present invention;

[0039] Figure 2 It is a schematic diagram of the steps of an AI-assisted radiotherapy planning method based on a diffusion model of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1-Figure 2 The present invention provides a technical solution: an AI-assisted radiotherapy planning method based on a diffusion model, the method comprising the following steps:

[0042] S1. Dataset collection and processing: Collect positioning CT images, contouring information and clinical information of cancer patients to establish a cancer radiotherapy dose database;

[0043] S2. Constructing a dose prediction network model: Based on the characteristics of the task and data, a deep learning network based on the deep diffusion generative model is established, and a model with generalization ability is obtained by taking advantage of its strong generative ability, high authenticity of results and rich details;

[0044] S3. 2.5D input strategy combined with individualized information: Adopting the slicing strategy of 2.5D CT data, the structure and performance of the model are optimized by combining the characteristics of two-dimensional and three-dimensional.

[0045] In step S1,

[0046] The collection and processing of the data set specifically includes the following steps:

[0047] S2-1. Collect 100 cancer cases, including patient disease course information, localization CT images, and delineation information of target areas and organs at risk;

[0048] S2-2. Clean the data to remove invalid and erroneous data, including missing values, format errors, and type errors;

[0049] S2-3, preprocess the data, including image format conversion, contrast adjustment, image standardization and data enhancement;

[0050] S2-4. Create a database to store data in paired and unpaired structures, and record the information of each patient.

[0051] The dose prediction network model includes a structure encoder, a dose encoder and a noise predictor; the model introduces a step-by-step diffusion generation mechanism to solve the problem of mode collapse and instability during training.

[0052] The structural encoder consists of five consecutive steps. In the first step, an embedding layer for parsing individual information is added. Except for the last step, each step contains a residual block and a downsampling block. The residual block consists of two convolution units, each of which includes a 3×3 convolution layer Conv, a group normalization layer GN and an activation function to maintain the residual connection and prevent the problem of gradient disappearance during training. The downsampling block consists of a 3×3 convolution layer Conv with a step size of 2.

[0053] The dose encoder consists of five consecutive steps, which are responsible for extracting features from the dose image, merging structural features, and inputting them into the noise predictor; the input content is encoded into a feature map through a convolutional layer, and then the two sets of feature maps are fused by element-by-element addition, allowing the structural information extracted in the structural encoder to be passed to the dose encoder; the next two downsampling steps continue to use addition operations to complete the fusion of information, while the following three steps use a cross-attention mechanism to achieve similarity-based structural guidance at a deeper level.

[0054] The noise predictor restores the feature representation captured by the dose encoder to the final output, where the skip connections are preserved for the reuse and aggregation of multi-level features.

[0055] The diffusion generation mechanism is specifically as follows:

[0056] The diffusion generation process is divided into two steps, namely forward diffusion generation and backward diffusion generation. In the forward diffusion generation process, the model adjusts the size and direction of the added noise according to the difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the target data distribution. In the backward diffusion generation process, the model adjusts the size and direction of the removed information according to the difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the noise sample.

[0057] The specific update rules of the diffusion generation mechanism are as follows:

[0058]

[0059] Among them, y t-1 and t Represents the images in the diffusion generation process, all of which are images at different noise levels, y t-1 Generate y by forward diffusion t ,y t Generate y by back diffusion t-1 ; α t and γ t are two control parameters, α t In this model, γ is used to adjust the noise level. t Used to control the way noise is added; ε t,θ is the noise term at time step t, which is a function of the model output, including parameters θ; z t is a random noise term, z t Obey standard normal distribution;

[0060] α t and γ t The parameters control the intensity and distribution of the noise, helping the model to transition smoothly between different time steps;

[0061] Noise term ε t,θ and z t The randomness of the generation process is guaranteed, allowing the model to generate diverse images.

[0062] The purpose of the dose prediction model is to train the noise predictor and the structure encoder so that the noise predicted in the reverse process is similar to the noise added in the forward process; to minimize the expected loss, the objective function is defined as follows:

[0063]

[0064] Among them, θ represents the parameter to be optimized; E (x,y) Represents the expected value of the data distribution (x, y); E ε,γ represents the expected value of the noise ε and the random variable γ; f(x feature ,yt ,γ t ) is a feature x feature , the target variable y t and the random variable γ t Function of ε t is the noise term, ε t It follows a normal distribution.

[0065] In order to reduce the number of parameters of the three-dimensional neural network model and prevent the two-dimensional neural network from losing the correlation information between different layers when processing two-dimensional image slices, this method adopts a slicing strategy for 2.5DCT data; this strategy aims to optimize the structure and performance of the model by combining the characteristics of two-dimensional and three-dimensional; for the dose prediction of each layer, the upper and lower 5 layers are taken as input, the information in each 2.5D layer is connected in the channel dimension, and the obtained multi-channel data is used as model input.

[0066] In an embodiment of the present invention,

[0067] Step S1: Dataset collection and processing;

[0068] S1-1: Collect data:

[0069] One hundred cancer cases were collected to obtain information on the course of the patients' disease, localization CT images, and delineation of target volumes and organs at risk.

[0070] S1-2: Data cleaning:

[0071] Remove invalid and erroneous data, such as blurred images, missing data, etc.

[0072] S1-3: Data preprocessing:

[0073] Image format conversion: Convert all CT images to a unified format, such as DICOM to PNG.

[0074] Contrast Adjustment: Normalize the contrast of CT images to improve image quality.

[0075] Image Normalization: Normalizes image pixel values ​​to be within the same range.

[0076] Data enhancement: Increase the amount of data through operations such as rotation, flipping, and scaling.

[0077] S1-4: Create database:

[0078] storing the processed data in a database, including paired and unpaired data structures, and recording the information of each patient;

[0079] Step S2: Constructing a dose prediction network model

[0080] S2-1: Network structure design:

[0081] Structural encoder: It consists of five steps. The first step adds an embedding layer to parse individual information; the remaining steps include residual blocks and downsampling blocks.

[0082] Residual block: Each residual block consists of two convolutional units, each of which includes a 3×3 convolutional layer (Conv), a group normalization layer (GN), and an activation function.

[0083] Downsampling block: consists of a 3×3 convolutional layer (Conv) with a stride of 2.

[0084] Dose encoder: Responsible for extracting features from the dose image and feeding them into the noise predictor to fuse the features via element-wise addition and cross-attention mechanism.

[0085] The input is encoded into a feature map through a convolutional layer.

[0086] Feature information is fused using addition operation and cross attention mechanism.

[0087] Noise Predictor: Restores the feature representation captured by the encoder to the final output, retaining skip connections for reuse and aggregation of multi-level features.

[0088] S2-2: Diffusion generation mechanism:

[0089] Forward diffusion generation: According to the difference between the current sample and the target data distribution, the size and direction of the added noise are adjusted to achieve gradual convergence to the target data distribution.

[0090] Backward diffusion generation: According to the difference between the current sample and the target data distribution, the size and direction of the removed information are adjusted to achieve gradual convergence to the noise sample.

[0091] The specific update rules of the diffusion generation mechanism are as follows:

[0092]

[0093] Among them, y t-1 and t Represents the images in the diffusion generation process, all of which are images at different noise levels, y t-1 Generate y by forward diffusion t ,y t Generate y by back diffusion t-1 ; α t and γ t are two control parameters, α t In this model, γ is used to adjust the noise level. t Used to control the way noise is added; εt,θ is the noise term at time step t, which is a function of the model output, including parameters θ; z t is a random noise term, z t Obey standard normal distribution;

[0094] The objective function is defined as follows:

[0095]

[0096] Among them, θ represents the parameter to be optimized; E (x,y) Represents the expected value of the data distribution (x, y); E ε,γ represents the expected value of the noise ε and the random variable γ; f(x feature ,y t ,γ t ) is a feature x feature , the target variable y t and the random variable γ t Function of ε t is the noise term, ε t It follows a normal distribution.

[0097] Step S3: 2.5D input strategy combined with individualized information:

[0098] 2.5D data input: For dose prediction at each level, CT images of the upper and lower 5 layers are used as input. The information in each 2.5D layer is connected in the channel dimension to form multi-channel data.

[0099] Multi-channel data is used as model input to maintain inter-layer correlation while reducing the number of model parameters.

[0100] By utilizing the powerful generation capability and rich details of the diffusion model and combining it with the 2.5D input strategy, the number of model parameters can be effectively reduced, making the model more accurate and easier to use in clinical practice.

[0101] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0102] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An AI-assisted radiotherapy planning method based on a diffusion model, characterized by: The method comprises the following steps: S1. Dataset collection and processing: Collect positioning CT images, contouring information and clinical information of cancer patients to establish a cancer radiotherapy dose database; S2. Constructing a dose prediction network model: Based on the characteristics of the task and data, a deep learning network based on the deep diffusion generative model is established, and a model with generalization ability is obtained by taking advantage of its strong generative ability, high authenticity of results and rich details; S3. Combine the slicing strategy of 2.5D data with the field angle information and personal dose limit to achieve personalized AI-assisted planning and change the structure and performance of the model.

2. The AI-assisted radiotherapy planning method based on diffusion model according to claim 1, characterized in that: In step S1, The collection and processing of the data set specifically includes the following steps: S2-1. Collect 100 cancer cases, including patient disease course information, localization CT images, and delineation information of target areas and organs at risk; S2-2. Clean the data to remove invalid and erroneous data, including missing values, format errors, and type errors; S2-3, preprocess the data, including image format conversion, contrast adjustment, image standardization and data enhancement; S2-4. Create a database to store data in paired and unpaired structures, and record the information of each patient.

3. The AI-assisted radiotherapy planning method based on diffusion model according to claim 1, characterized in that: The dose prediction network model includes a structure encoder, a dose encoder and a noise predictor; the model introduces a step-by-step diffusion generation mechanism to solve the problem of mode collapse and instability during training.

4. The AI-assisted radiotherapy planning method based on diffusion model according to claim 3, characterized in that: The structural encoder consists of five consecutive steps. In the first step, an embedding layer for parsing individual information is added. Except for the last step, each step contains a residual block and a downsampling block. The residual block consists of two convolution units, each of which includes a 3×3 convolution layer Conv, a group normalization layer GN and an activation function to maintain the residual connection and prevent the problem of gradient disappearance during training. The downsampling block consists of a 3×3 convolution layer Conv with a step size of 2.

5. The AI-assisted radiotherapy planning method based on diffusion model according to claim 3, characterized in that: The dose encoder consists of five consecutive steps responsible for extracting features from the dose image, incorporating structural features, and feeding them into the noise predictor; The input content is encoded into feature maps through convolutional layers, and then the two sets of feature maps are fused by element-wise addition, allowing the structural information extracted in the structure encoder to be passed to the dose encoder; the next two downsampling steps continue to use addition operations to complete the fusion of information, while the subsequent three steps adopt a cross-attention mechanism to achieve similarity-based structural guidance at a deeper level.

6. The AI-assisted radiotherapy planning method based on diffusion model according to claim 3, characterized in that: The noise predictor restores the feature representation captured by the dose encoder to the final output, where the skip connections are preserved for the reuse and aggregation of multi-level features.

7. The AI-assisted radiotherapy planning method based on diffusion model according to claim 3, characterized in that: The diffusion generation mechanism is specifically as follows: The diffusion generation process is divided into two steps, namely forward diffusion generation and backward diffusion generation. In the forward diffusion generation process, the model adjusts the size and direction of the added noise according to the degree of difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the target data distribution. In the process of back-diffusion generation, the model adjusts the size and direction of the removed information according to the difference between the current sample and the target data distribution, so as to achieve the purpose of gradually converging to the noise sample. The specific update rules of the diffusion generation mechanism are as follows: Where t represents the time step; y t-1 and t Represents the images in the diffusion generation process, all of which are images at different noise levels, y t-1 Generate y by forward diffusion t ,y t Generate y by back diffusion t-1 ; α t and γ t are two control parameters, α t In this model, γ is used to adjust the noise level. t Used to control the way noise is added; t,θ is the noise term at time step t, which is a function of the model output, including parameters θ; z t is a random noise term, z t Obey standard normal distribution; α t and γ t The parameters control the intensity and distribution of the noise, helping the model to transition smoothly between different time steps; Noise term ε t,θ and z t The randomness of the generation process is guaranteed, allowing the model to generate diverse images.

8. The AI-assisted radiotherapy planning method based on diffusion model according to claim 3, characterized in that: The purpose of the dose prediction model is to train the noise predictor and the structure encoder so that the noise predicted in the reverse process is similar to the noise added in the forward process; to minimize the expected loss, the objective function is defined as follows: Where t represents the time step; θ represents the parameter to be optimized; E (x,y) Represents the expected value of the data distribution (x, y); E ε,γ represents the expected value of the noise ε and the random variable γ; f(x feature ,y t ,γ t ) is a feature x feature , the target variable y t and the random variable γ t Function of ε t is the noise term, ε t It follows a normal distribution.

9. The AI-assisted radiotherapy planning method based on diffusion model according to claim 1, characterized in that: The 2.5D input strategy combining individualized information is specifically as follows: In order to reduce the number of parameters of the three-dimensional neural network model and prevent the two-dimensional neural network from losing the correlation information between different layers when processing two-dimensional image slices, this method adopts a slicing strategy for 2.5DCT data; this strategy aims to optimize the structure and performance of the model by combining the characteristics of two-dimensional and three-dimensional; for the dose prediction of each layer, the upper and lower 5 layers are taken as input, the information in each 2.5D layer is connected in the channel dimension, and the obtained multi-channel data is used as model input.

Citation Information

Cited By

  • Local advanced rectal cancer neoadjuvant therapy curative effect prediction system based on AI

    CN119811589A