An intelligent plan optimization method and system

Abstract

The present application relates to the technical field of radiotherapy, in particular to an intelligent plan optimization method and system, which initially uses the combination features of distance and delineation, the principal component analysis method for dimension reduction of weighted delineation, and the model training prediction system of classification first and then regression. The acquisition of the distance image and the combination of the distance image and the delineation image are respectively subjected to feature extraction operation and feature conversion operation, which can improve the prediction accuracy; the data is subjected to dimension reduction processing, which can reduce the training difficulty. The present application can complete automatic plan optimization based on dose prediction, and improve the practical value.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and in particular to an intelligent planning optimization method and system. Background Technology

[0002] In intensity-modulated radiotherapy (IMRT), physicists must repeatedly adjust parameters for the target area and organs at risk to improve and optimize results to meet clinical standards. In existing technologies, to address the problem of predicting the three-dimensional dose distribution, a deep learning-based method for predicting IMRT dose (CN114155934A) uses distance images along with CT images and delineated images as features to predict the three-dimensional dose distribution. However, this method does not consider the combined relationship between distance and delineation, resulting in a problem and drawback of excessively large input feature dimensions. Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose an intelligent planning optimization method, including a method for predicting dose based on distance features and an automatic planning optimization method based on predicted dose, in order to solve the problems of the prior art.

[0004] To achieve the above objectives, the present invention provides a method for predicting dose based on distance features, comprising the following steps:

[0005] Acquire CT images, delineated images, and dose data of the case;

[0006] Resample CT images, delineated images, and dose data at the same resolution (e.g., 2.5mm × 2.5mm × 2.5mm);

[0007] Obtain the coordinates of the target region's outline, calculate the shortest distance from the coordinates of each voxel of each endangered organ to the coordinates of the target region's outline, and construct a distance image at the same resolution.

[0008] The distance image of the organ at risk and the corresponding delineation image of the organ at risk are combined to form a distance-weighted delineation image.

[0009] Dimensionality reduction is performed on all weighted delineation images. The three-dimensional delineation image of each organ at risk is expanded into a one-dimensional vector. Then, principal component analysis is used to reduce the dimension. This one-dimensional vector is converted into principal components through orthogonal transformation. Based on the variance, the largest number of principal components are selected so that the cumulative contribution rate of the variance reaches a set ratio. The three-dimensional image is then restored to obtain the dimensionality-reduced weighted delineation image.

[0010] The CT images, target delineation images, dimensionality-reduced weighted delineation images, and corresponding dose data were used to perform regression training on the 3D-Unet model to obtain the trained regression model.

[0011] The CT image to be predicted, the target area delineation image, and the obtained dimensionality-reduced weighted delineation image are input into the trained regression model to obtain the dose prediction results.

[0012] Preferably, before using the 3D-Unet model for regression training, this method further includes:

[0013] The dose data is converted from numerical description to categorical description. A binary classification model is trained to determine the classification result of each voxel. Data from regions with dose are selected as input data for regression training.

[0014] Preferably, the formula for calculating the shortest distance from the coordinates of each voxel of each endangered organ to the contour coordinates of the target area is:

[0015]

[0016] Among them, V OARS and Π PTV Let represent the three-dimensional coordinate set of the organ at risk and the three-dimensional coordinate set on the target contour, respectively. t represents the coordinates of the voxel at (i,j,k) in the organ at risk, s represents the coordinates of a point on the target contour, d represents the distance function, and l, h, and w represent the number of slices, length, and width of the distance image, respectively.

[0017] Preferably, combining the distance image of the organ at risk and the delineated image to form a distance-weighted delineated image includes:

[0018] The reciprocal of the distance image is taken, and then multiplied element-wise with the corresponding delineation image to form a distance-weighted delineation image. The rules for taking the reciprocal are as follows: if the element is non-zero (i.e., the coordinates are within the delineation area, other locations outside the delineation / target area, or within the target area), the element is directly taken as the reciprocal; if the element is zero (i.e., the coordinates are on the target area outline), the reciprocal is taken as a larger value N. The formula is as follows:

[0019] C′=(c i，j，k *d′ i，j，k ) l×h×w ,

[0020]

[0021]

[0022] Where C' is the weighted image of the organs at risk, c i,j,k This represents the mask value for outlining, which can be either 0 or 1, R. OAR d' represents the area to be outlined with a mask value of 1, and d' represents the reciprocal of the distance.

[0023] Preferably, regression training using the 3D-Unet model includes:

[0024] The 3D-Unet model used has a "U"-shaped structure, including 4 downsampling operations and 4 upsampling operations;

[0025] During training, each downsampling step includes convolution, BN, and ReLU operations, and each upsampling step includes concatenation. The Adam optimizer is used for training, and an early stopping training method is adopted. When the loss on the revalidation set no longer decreases within a certain number of steps, training is stopped, and the model parameters with the minimum loss on the validation set are saved, thus obtaining the trained regression model.

[0026] This invention also provides an automated treatment planning optimization method based on predicted dose. Based on the aforementioned predicted dose method, after obtaining the dose prediction result, the objective functions for the target area and organs at risk are calculated according to the dose prediction result to complete the automated treatment planning optimization. The objective function formula is as follows:

[0027]

[0028] Where d j The calculated dose for the j-th point is obtained from the field intensity, p j For the predicted dose at point j, α j It is a constraint and punishment factor; for organs at risk, d j >p j At that time, the point that violates the constraint, α j It is a non-zero factor, d j <p j At that time, the point that satisfies the constraint, α j It is 0;

[0029] For the target area, d j <p j At that time, the point that violates the constraint, α j It is a non-zero factor, d j >p j At that time, the point that satisfies the constraint, α j It is 0.

[0030] The present invention also provides an intelligent treatment planning optimization system, including a system for predicting dose based on distance features and an automatic treatment planning optimization system based on predicted dose, wherein the system for predicting dose based on distance features includes the following steps:

[0031] The input module is used to acquire CT images, outline images, and dose data of the case.

[0032] A feature processing module, configured to obtain the coordinates of the contour of the target area after resampling, calculate the shortest distance from the voxel coordinates of each organ at risk to the contour coordinates of the target area, form a distance image at the same resolution, and then combine the distance image of the organ at risk and the corresponding delineation image of the organ at risk to form a weighted delineation image based on distance weighting. Then, dimensionality reduction is performed on all the weighted delineation images. The three-dimensional weighted delineation image of each organ at risk is unfolded into a one-dimensional vector, and then the principal component analysis method is used for dimensionality reduction. The principal components are selected according to the cumulative contribution rate of variance and restored to three-dimensional images to obtain the weighted delineation image after dimensionality reduction;

[0033] A training module, configured to perform regression training on the CT image, the target area delineation image, the weighted delineation image after dimensionality reduction, and the corresponding dose data using a 3D-Unet model to obtain a trained regression model;

[0034] A prediction module, configured to input the CT image to be predicted, the target area delineation image, and the weighted delineation image after dimensionality reduction into the trained regression model to obtain a dose prediction result.

[0035] The present invention further provides an automatic plan optimization system based on predicted dose. On the basis of including the above system for predicting dose, it further includes an optimization module. The optimization module is configured to, after obtaining the dose prediction result, calculate the objective function of the target area and the organ at risk to complete automatic plan optimization. The objective function formula is

[0036]

[0037] where d j is the calculated dose at the jth point, p j is the predicted dose at the jth point, and α j is a constraint penalty factor. For an organ at risk, when dj>pj, for the points violating the constraint, α j is a non-zero factor. When dj<pj, for the points satisfying the constraint, αj is 0;

[0038] For the PTV, when d j <p j , for the points violating the constraint, α j is a non-zero factor. When d j >p j , for the points satisfying the constraint, α j is 0;

[0039] Calculate the minimum value of the objective function F obj to obtain the calculated dose d j at each point. Through d jThe optimized plan is obtained. The beneficial effects of this invention are as follows: This invention proposes a feature processing system and a model training and prediction system for dose prediction. It innovatively uses a combination of distance and delineation features, principal component analysis to reduce the dimensionality of the weighted delineation, and a model training and prediction system that first classifies and then regresses. Obtaining the distance image involves feature extraction, and combining it with the delineation image is equivalent to feature transformation of the delineation, which improves prediction accuracy. In regression training, the dimensionality reduction of the data reduces training difficulty. This invention can also automatically optimize the plan based on dose prediction, improving its practical value. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the intelligent planning optimization method according to an embodiment of the present invention;

[0042] Figure 2 This is a block diagram of the intelligent planning optimization system according to an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0044] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0045] like Figure 1 As shown in the embodiments of this specification, an intelligent plan optimization method is provided, including the following steps:

[0046] Acquire case data, including CT images, delineated images, and dosage data;

[0047] The acquired data undergoes feature processing, and the processing steps are as follows:

[0048] CT images, delineated images, and dose data are resampled at the same resolution (e.g., 2.5mm).

[0049] (×2.5mm×2.5mm). First, the CT scan is resampled; then, the dose data is interpolated to the z-axis position of the CT slice. Next, the dose data is resampled, and the resampled dose data is filled to the CT size with 0; then, the outline is resampled, and the unoutlined positions are filled with zero slices; finally, all 3D data are cropped to a fixed size (e.g., 128×128×128).

[0050] Obtain the coordinates of the target region's contour, calculate the shortest distance from the coordinates of each voxel of each endangered organ to the coordinates of the target region's contour, and construct a distance image at the same resolution. The formula for calculating the shortest distance is:

[0051]

[0052] Among them, V OARS and Π PTV Let represent the three-dimensional coordinate set of the organ at risk and the three-dimensional coordinate set on the target contour, respectively; t represents the coordinate of the voxel at (i,j,k) in the organ at risk; s represents the coordinate of a point on the target contour; d represents the distance function; and l, h, and w represent the number of slices, length, and width of the distance image, respectively.

[0053] The distance images of organs at risk are combined with the corresponding delineated images of those organs to form a distance-weighted delineated image. The specific combination method is as follows: First, the reciprocal of each element of the distance image is taken, and then element-wise multiplied with the delineated image. The rules for taking the reciprocal are as follows: If the element is non-zero (i.e., the coordinates are within the delineated area, other locations outside the delineated / non-target area, or within the target area), the element is directly taken as its reciprocal; if the element is zero (i.e., the coordinates are on the target area outline), the reciprocal is taken as a larger value N. When the reciprocals of elements located outside the delineated area are multiplied with the delineated image, they all become zero. Therefore, the weighting is ultimately applied only to the data within the delineated area, as shown in the following formula:

[0054] C′=(c i，j，k *d′ i，j，k ) l×h×w ,

[0055]

[0056]

[0057] Where C' is the weighted image of the organs at risk, c i,j,k This represents the mask value for outlining, which can be either 0 or 1, R. OAR This indicates the area to be outlined with a mask value of 1. d' represents the reciprocal of the distance.

[0058] Dimensionality reduction is performed on all weighted delineation images. Each weighted 3D delineation image of an organ at risk is unfolded into a 1D vector. Then, principal component analysis is used for dimensionality reduction. Through orthogonal transformation, this set of vectors that may be correlated is transformed into a set of linearly uncorrelated vectors. This set of vectors is called principal components. The principal components with the largest variances are selected in descending order of variance, and the cumulative variance contribution rate reaches a certain percentage (e.g., more than 95%). The 3D image is then restored to obtain the dimensionality-reduced weighted delineation image.

[0059] The model is then trained and used for prediction on the feature-processed data. The steps include:

[0060] The case data was divided into training, validation, and test sets. The CT images, target delineation images, dimensionality-reduced weighted delineation images, and corresponding dose data in the training set were used for regression training using the 3D-Unet model. The 3D-Unet model has a "U"-shaped structure, including four downsampling operations and four upsampling operations. Specifically, the 3D-Unet model was constructed using PyTorch: 1. The input image size was 6×128×128×128, where 6 represents the number of channels; 2. It then passed through two convolutional layers with a kernel size of 3, padding of 1, and a default stride of 1. Each convolutional layer was followed by BN and ReLU operations, resulting in an output image size of 16×128×128×128; 3. A pooling layer was then used for downsampling, followed by two more identical convolutional operations, resulting in an output image size of 32×64×64×64; 4. Step 3 was repeated three more times, resulting in an output image size of 256×8×8×64×64. 8; 5. Next, perform upsampling using transposed convolution with a kernel of 2 and a stride of 2. This is concatenated with the input before the last downsampling, resulting in an image size of 256×16×16×16. Then, perform two more identical convolution operations, outputting an image size of 128×16×16×16. 6. Repeat step 5 three times, outputting an image size of 16×128×128×128. Finally, pass the image through a convolutional layer with a kernel of 1 and 1 output channel, resulting in an image size of 1×128×128×128, corresponding to dose prediction. Using the Adam optimizer with a learning rate of 3e-4 and weight decay of 1e-4, an early stopping training method is employed. Training stops when the loss on the revalidation set no longer decreases within a certain number of steps, and the model parameters with the minimum loss on the validation set are saved, resulting in the trained regression model. Because the data dimensionality is reduced, the training difficulty is lowered.

[0061] The CT images, target delineation images, and the resulting dimensionality-reduced weighted delineation images from the test set are input into the trained regression model to obtain the dose prediction results. Obtaining the distance image involves feature extraction, and combining it with the delineation image is equivalent to performing feature transformation on the delineation. This kind of manually designed feature extraction is a common operation in machine learning and can also improve prediction accuracy in deep learning.

[0062] After obtaining the dose prediction results, this method also includes automatic plan optimization, including calculating the objective function for the target area and organs at risk. The formula for the objective function F is as follows:

[0063]

[0064] Where, d j The calculated dose for the j-th point is obtained from the field intensity, p j For the predicted dose at point j, α j It is a constraint and punishment factor; for organs at risk, d j >p j At that time, the point that violates the constraint, α j It is a non-zero factor, d j <p j At that time, the point that satisfies the constraint, α j It is 0; for the target area PTV, d j <p j At that time, the point that violates the constraint, α j It is a non-zero factor, d j >p j At that time, the point that satisfies the constraint, α j It is 0; calculate the objective function F. obj The minimum value is used to obtain the calculated dose d at each point. j , through d j The optimized plan was obtained.

[0065] As one implementation method, considering the large number of dose-free regions in dose resampling, the dose prediction method also includes the following before using the 3D-Unet model for regression training:

[0066] The dose data is converted from numerical descriptions to categorical descriptions, with no-dose regions represented by 0 and dose-containing regions by 1. A binary classification model is trained to determine the classification result for each voxel, and regions with a prediction result of 1 (i.e., dose-containing regions) are selected for regression training. Specifically, the mask data indicating whether a dose is present obtained from the classification is multiplied element-wise with the dose data output from the regression model, and finally, a loss is calculated with the actual dose data. The input data for both the classification and regression models is the same. This further reduces the amount of training required and improves training efficiency.

[0067] like Figure 2 As shown in the embodiments of this specification, an intelligent planning optimization system is also provided, including a system for predicting dose based on distance features and an automatic planning optimization system based on predicted dose, wherein the system for predicting dose based on distance features includes:

[0068] The input module is used to acquire CT images, outline images, and dose data of the case.

[0069] The feature processing module is used to obtain the coordinates of the target region contour after resampling, calculate the shortest distance from the coordinates of each voxel of each organ at risk to the coordinates of the target region contour, and construct a distance image at the same resolution. Then, the distance image of the organ at risk and the corresponding delineation image of the organ at risk are combined to form a distance-weighted delineation image. Then, all weighted delineation images are dimensionality reduced. The weighted three-dimensional delineation image of each organ at risk is unfolded into a one-dimensional vector. Then, principal component analysis is used to reduce the dimensionality. The principal components are selected according to the cumulative contribution rate of variance and restored to a three-dimensional image to obtain the dimensionality-reduced weighted delineation image.

[0070] The training module is used to perform regression training on the 3D-Unet model using CT images, target delineation images, dimensionality-reduced weighted delineation images and corresponding dose data to obtain a trained regression model.

[0071] The prediction module is used to input the CT image to be predicted, the target area delineation image, and the obtained dimensionality-reduced weighted delineation image into the trained regression model to obtain the dose prediction result.

[0072] This specification also provides an automatic medical plan optimization system based on predicted dose, including the aforementioned predicted dose system, and further including an optimization module. The optimization module, after obtaining the dose prediction result, calculates the objective function for the target area and organs at risk based on the dose prediction result, thereby completing the automatic medical plan optimization. The objective function formula is:

[0073]

[0074] Where d j For the calculated dose at point j, p j For the predicted dose at point j, α j It is a constraint and punishment factor; for organs at risk, d j >p j At that time, the point that violates the constraint, α j It is a non-zero factor, d j <p j At that time, the point that satisfies the constraint, α j It is 0; for PTV, d j <p j At that time, the point that violates the constraint, αj It is a non-zero factor, d j >p j At that time, the point that satisfies the constraint, α j It is 0. Calculate the objective function F. obj The minimum value is used to obtain the calculated dose d at each point. j , through d j The optimized plan was obtained.

[0075] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0076] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for intelligent planning optimization, characterized in that... A method for predicting dose based on distance features and an automatic planning optimization method based on predicted dose; wherein, the method for predicting dose based on distance features includes the following steps: Acquire CT images, delineated images, and dose data of the case, including target area delineated images and organs at risk delineated images; CT images, delineated images, and dose data are resampled at the same resolution; Obtain the coordinates of the target region's outline, calculate the shortest distance from the coordinates of each voxel of each organ at risk to the coordinates of the target region's outline, and construct a distance image of the organs at risk. The formula for calculating the shortest distance from the coordinates of each voxel of each organ at risk to the target contour coordinates is as follows: ， Among them, V OARS and Π PTV Let t represent the set of three-dimensional coordinates of the organ at risk and the set of three-dimensional coordinates on the target area contour, respectively. i，j，k The coordinates of the voxel at (i,j,k) in the organ at risk are represented, s represents the coordinates of a point on the target region contour, d represents the distance function, and l, h, and w represent the number of slices, length, and width of the distance image, respectively. The distance image of the organ at risk and the corresponding delineation image of the organ at risk are combined to form a distance-weighted delineation image. The process of combining the distance image and the delineated image of the organ at risk to form a distance-weighted delineated image includes: Take the reciprocal of the distance image and then perform element-wise multiplication with the corresponding delineation image to construct a distance-weighted delineation image. The formula for the weighted delineation image is: ， ， ， Where C' is the weighted image of the organs at risk, c i,j,k This represents the mask value for outlining, which can be either 0 or 1, R. OAR d' represents the outlined area with a mask value of 1. i,j,k This represents the reciprocal of the distance at (i,j,k) in the organ at risk, where N is a constant; Dimensionality reduction is performed on all weighted delineation images. The three-dimensional delineation image of each organ at risk is expanded into a one-dimensional vector. Then, principal component analysis is used to reduce the dimension. This one-dimensional vector is converted into principal components through orthogonal transformation. Based on the variance, the largest number of principal components are selected so that the cumulative contribution rate of the variance reaches a set ratio. The three-dimensional image is then restored to obtain the dimensionality-reduced weighted delineation image. The CT images, target delineation images, dimensionality-reduced weighted delineation images, and corresponding dose data were used to perform regression training on the 3D-Unet model to obtain the trained regression model. The CT image to be predicted, the target area delineation image, and the obtained dimensionality-reduced weighted delineation image are input into the trained regression model to obtain the dose prediction results.

2. The method for predicting dose based on distance features according to claim 1, characterized in that, The CT images, delineated images, and dose data are resampled at the same resolution. First, the CT images are resampled. Then, the dose data is interpolated to the z-axis position of the CT slice. Next, the dose data is resampled, and the resampled dose data is filled with 0 to the size of the CT image. Then, the delineated images are resampled, and zero slices are filled in the un-delineated positions. Finally, all 3D data are cropped to a fixed size.

3. The method for predicting dose based on distance features according to claim 1, characterized in that, Before using the 3D-Unet model for regression training, the method further includes: The dose data is converted from numerical description to categorical description. A binary classification model is trained to determine the classification result of each voxel. Data from regions with dose are selected as input data for regression training.

4. The method for predicting dose based on distance features according to claim 1, characterized in that, The regression training using the 3D-Unet model includes: the 3D-Unet model used has a "U" shaped structure, including 4 downsampling operations and 4 upsampling operations; During training, each downsampling step includes convolution, BN, and ReLU operations, and each upsampling step includes concatenation. The Adam optimizer is used for training, and an early stopping training method is adopted. When the loss on the revalidation set no longer decreases within a certain number of steps, training is stopped, and the model parameters with the minimum loss on the validation set are saved, thus obtaining the trained regression model.

5. The automatic treatment planning optimization method based on predicted dose according to claim 1, characterized in that, Based on the dose prediction method as described in claims 1-4, after obtaining the dose prediction results, the objective functions for the target area and organs at risk are calculated according to the dose prediction results to complete automatic planning optimization. The objective function formula is as follows: ， Where d j For the calculated dose at point j, p j For the predicted dose at point j, α j It is a constraint penalty factor; calculate the objective function F. obj The minimum value is used to obtain the calculated dose d at each point. j , through d j The optimized plan was obtained.

6. An intelligent planning optimization system, characterized in that... A system for predicting dose based on distance characteristics and an automated planning optimization system based on predicted dose, wherein the system for predicting dose based on distance characteristics includes the following steps: The input module is used to acquire CT images, outline images, and dose data of the case. The feature processing module is used to obtain the coordinates of the target region's contour after resampling, calculate the shortest distance from the coordinates of each voxel of each organ at risk to the coordinates of the target region's contour, and construct a distance image at the same resolution. The formula for calculating the shortest distance from the coordinates of each voxel of each organ at risk to the target contour coordinates is as follows: ， Among them, V OARS and Π PTV Let t represent the set of three-dimensional coordinates of the organ at risk and the set of three-dimensional coordinates on the target area contour, respectively. i，j，k Let represent the coordinates of the voxel at (i,j,k) in the organ at risk, s represent the coordinates of a point on the target region contour, d represent the distance function, and l, h, and w represent the number of slices, length, and width of the distance image, respectively. The distance image of the organ at risk and the corresponding delineated image of the organ at risk are combined to form a distance-weighted delineated image. The process of combining the distance image of the organ at risk and the delineated image to form a distance-weighted delineated image includes: Take the reciprocal of the distance image and then perform element-wise multiplication with the corresponding delineation image to construct a distance-weighted delineation image. The formula for the weighted delineation image is: ， ， ， Where C' is the weighted image of the organs at risk, c i,j,k This represents the mask value for outlining, which can be either 0 or 1, R. OAR d' represents the outlined area with a mask value of 1. i,j,k This represents the reciprocal of the distance at (i,j,k) in the organ at risk, where N is a constant; Dimensionality reduction is performed on all weighted delineation images. The three-dimensional delineation image of each organ at risk is expanded into a one-dimensional vector. Then, principal component analysis is used to reduce the dimension. This one-dimensional vector is converted into principal components through orthogonal transformation. Based on the variance, the largest number of principal components are selected so that the cumulative contribution rate of the variance reaches a set ratio. The three-dimensional image is then restored to obtain the dimensionality-reduced weighted delineation image. The training module is used to perform regression training on the 3D-Unet model using CT images, target delineation images, dimensionality-reduced weighted delineation images and corresponding dose data to obtain a trained regression model. The prediction module is used to input the CT image to be predicted, the target area delineation image, and the obtained dimensionality-reduced weighted delineation image into the trained regression model to obtain the dose prediction result.

7. An automated treatment planning optimization system based on predicted dose, characterized in that, The system for predicting dose as described in claim 6 further includes an optimization module, which, after obtaining the dose prediction results, calculates objective functions for the target area and organs at risk based on the dose prediction results to complete automatic plan optimization. The objective function formula is: ， Where d j For the calculated dose at point j, p j For the predicted dose at point j, α j It is a constraint penalty factor, used to calculate the objective function F. obj The minimum value is used to obtain the calculated dose d at each point. j , through d j The optimized plan was obtained.

Images