Multi-criterion medical image dose prediction method

Abstract

The invention discloses a multi-criterion medical image dose prediction method. Firstly, medical images are acquired, and image preprocessing such as merging, cutting and normalization is carried out on the medical images; training a deep neural network to obtain a dose prediction basic model; taking the weight of a part of network layers of the basic model as an individual, and performing optimization by using an evolutionary algorithm to obtain a dose prediction generalization model; and finally, predicting the dose by using the generalization model to obtain a multi-criterion dose prediction distribution diagram. According to the method, the evolutionary algorithm is combined with the deep neural network, and the deep neural network model is optimized through the evolutionary algorithm, so that the universality of the medical image dose prediction model can be improved, and the method can better adapt to different radiation systems.

Description

Technical Field

[0001] The invention relates to the field of medical image dose prediction, in particular to a multi-criteria medical image dose prediction method. Background Art

[0002] Traditional radiation systems will inevitably irradiate the normal areas around the target area, causing damage to them. Then, conformal radiation systems were developed, which use collimators to obtain the shape of the radiation field, thereby protecting the normal areas around the target area. However, due to the irregular shape of the target area, such as bulges in some areas, the conformal radiation system cannot ensure uniform dose distribution within the target area. To address this problem, the intensity modulated radiation system was developed.

[0003] The basic steps of radiation planning by intensity modulated radiation system include: obtaining medical images to be processed, outlining the target area PTV and organs at risk OARs, dose prediction, reverse optimization, plan verification and adaptive adjustment. Dose prediction plays an important role in the process of radiation plan formulation. It can be used to provide parameters close to the optimal reverse plan, greatly reducing the number of times the physicist adjusts the parameters and the time of reverse optimization, thereby greatly improving the efficiency of the plan; at the same time, high-quality dose prediction can provide personalized planning guidance, which will greatly reduce the harm of the radiation system to normal areas.

[0004] With the booming development of deep neural networks (DNN), radiation systems have begun to use deep neural networks to automatically predict dose distribution. Using deep neural networks to predict dose distribution can reduce the workload of physicists, greatly shorten the time for formulating radiation plans, and improve radiation quality. Although deep neural networks have been widely used in dose prediction in recent years and have achieved many results, due to differences in data sets, the trained deep neural network models often lack universality. For example, a deep neural network model trained using medical image data from a certain institution is deployed in the radiation system of another institution. Due to differences caused by different brands of scanners, machine parameter settings, image post-processing, and different medical images, the effect may be unsatisfactory or the dose prediction function may fail completely. Therefore, due to the limited generalization ability of deep neural network models, medical image dose prediction still faces great challenges. Summary of the invention

[0005] The present invention aims to solve the deficiencies of the above-mentioned prior art and propose a multi-criteria medical image dose prediction method in order to improve the versatility of medical image dose prediction and thus better adapt to different radiation systems.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The multi-criteria medical image dose prediction method of the present invention is characterized in that it is performed according to the following steps:

[0008] Step 1: Obtain medical images and perform preprocessing;

[0009] Step 1.1, obtaining a CT image of the head and neck, and performing cropping and normalization to obtain a preprocessed CT image;

[0010] Step 1.2, obtaining the PTV images of all the target areas of the head and neck and merging them into one target area PTV image, and recording the binary mask of the target area PTV image as PTVs;

[0011] Step 1.3, obtaining characteristic information sets OARs of several organs at risk of the head and neck, and recording the binary mask of the characteristic information set OARs as OMs;

[0012] Step 1.4: Define the actual dose distribution map of the head and neck ,make A single voxel The dose value in the real dose distribution diagram is ;

[0013] Step 2: Use the preprocessed medical images to train the dose prediction network to obtain a dose prediction basic model;

[0014] Step 2.1: Construct a dose prediction network based on the U-net model, and input the preprocessed CT image, the target binary mask PTVs, and the binary mask OMs of OARs into the dose prediction network to obtain a predicted dose distribution map. , where a single voxel In the predicted dose distribution map The dose value in is recorded as ;

[0015] Step 2.2: Divide the dose distribution map using equations (1) and (2) , and obtain the target dose distribution map , Dose distribution map of organs at risk :

[0016] (1)

[0017] (2)

[0018] In formula (1) and formula (2), Represents a bitwise AND operation;

[0019] Step 2.3: Use formula (3) to construct the predicted dose distribution map Optimal dose map The loss function between , updating the weight parameters of the dose prediction network through a back-propagation algorithm, and stopping the network training when the maximum number of iterations is reached, thereby obtaining a dose prediction basic model;

[0020] (3)

[0021] In formula (3), represents any dose distribution diagram, and , , express The number of voxels in ;

[0022] Step 3: Construct constraints and fitness function;

[0023] Step 3.1: Use equation (4) to construct the minimum dose equation for all voxels :

[0024] (4)

[0025] In formula (4), Represents a single voxel The dosage, is the minimum dose of the set dose constraint;

[0026] Step 3.2: Use equation (5) to construct the maximum dose constraint equation for all voxels :

[0027] (5)

[0028] In formula (5), is the maximum dose of the set dose constraint;

[0029] Step 3.3: Use equation (6) to construct the optimal dose constraint equation for all voxels :

[0030] (6)

[0031] In formula (6), is the optimal dose for the set dose constraint;

[0032] Step 3.4: Use equation (7) to construct the dose-volume constraint equation for all voxels. :

[0033] (7)

[0034] In formula (7), is the maximum dose of the dose-volume constraint, express The maximum voxel dose in the medium dose;

[0035] Step 3.5: Use formula (8) to construct the first objective function :

[0036] (8)

[0037] Step 3.6: Use formula (9) to construct the second objective function :

[0038] (9)

[0039] Step 3.7: Use formula (10) to construct the fitness function :

[0040] (10)

[0041] In formula (10), and is 2 weights;

[0042] Step 4: using an evolutionary algorithm to optimize the dose prediction basic model to obtain a dose prediction generalization model for performing multi-criteria dose prediction on the medical image;

[0043] Step 4.1, initialize the population;

[0044] Step 4.1.1, define the maximum number of iterations as , the current number of iterations is , and initialize ;

[0045] Step 4.1.2: Select the weight vector of some layers in the dose prediction basic model And as a basic individual, after N random perturbations on the basic individual, the G generation population is obtained ,in, represents the i-th individual of the G-th generation population, and each individual contains t decision variables, ,in, express The j-th decision variable in ;

[0046] Step 4.2, calculation Fitness ;

[0047] Step 4.3: According to the fitness of each individual, the G generation population Perform non-dominated sorting and calculate the crowding distance of the non-dominated front after sorting, so as to sort the individuals in ascending order according to the crowding distance, and finally obtain the G-generation population in ascending order;

[0048] Step 4.4: Perform crossover mutation on the ascending G-th generation population to generate the offspring population of the G-th generation population. ,in represents the i-th individual in the offspring population;

[0049] Step 4.5 Calculate according to formula (4)-formula (10) The fitness of each individual;

[0050] Step 4.6: Merge and According to the fitness of each individual, the G-th generation population after the merger is non-dominated and the crowding distance of the non-dominated frontier after sorting is calculated. Then, the individuals in the G-th generation population after the merger are sorted in ascending order according to the crowding distance, and finally the ascending non-dominated frontier is obtained, and the first N individuals are selected as the first Generation population ;

[0051] Step 4.7, assign G+1 to G, and repeat steps 4.3 until G> So far, we get Generation population ;in, Indicates The i-th individual in the generation population;

[0052] Step 4.8: Each individual in is successively reduced to the weight of the corresponding layer in the basic model of dose prediction, thereby obtaining N generalized models of dose prediction, which are used to predict the preprocessed CT images, binary masks of target areas PTVs, and binary masks of OARs OMs, and obtain N predicted dose distribution maps of multiple criteria.

[0053] The multi-criteria cancer image dose prediction method described in the present invention is also characterized in that the U-net model in step 2 is composed of an input convolution layer, an output convolution layer, M downsampling convolution modules and M upsampling modules, wherein each downsampling convolution layer is composed of a downsampling layer, a convolution layer, an instance normalization layer and a rectified linear unit, and each upsampling module is composed of an upsampling layer, a convolution layer, an instance normalization layer and a rectified linear unit;

[0054] The feature map output by the mth downsampling convolution layer is mapped to the input of the corresponding mth upsampling deconvolution layer, and the feature map output by the mth downsampling convolution layer and the feature map output by the m-1th upsampling deconvolution layer are connected through the channel dimension to form the mth skip connection structure.

[0055] Further, the step 4.2 includes:

[0056] Step 4.2.1. Restored to the weights of the corresponding layers in the dose prediction basic model , thus obtaining the Gth generation i-th dose prediction generalization model, and predicting the preprocessed CT images, target binary mask PTVs and OARs binary mask OMs to obtain the Gth generation i-th dose distribution map ;

[0057] Step 4.2.2: Use equations (1) and (2) to obtain the dose distribution map of the i-th target area of ​​the G-th generation: , the i-th organ at risk in the Gth generation , and then according to formula (4)-formula (10) we can get Fitness .

[0058] Further, the step 4.4 includes:

[0059] Step 4.4.1: Based on the non-dominated frontier and its crowding distance, use the binary tournament method to randomly select the i-th individual from the G-th generation population in ascending order. and the oth individual , and after randomly selecting the crossover point k, exchange and The decision variables after the kth position, thus corresponding to the generation of the i-th offspring individual and the oth offspring individual ;

[0060] Step 4.4.2 Calculate using formula (11) and The variation value of each decision variable in , and in Generate random numbers in a range ,like , then use formula (12) to mutate the decision variable at the corresponding position to obtain the mutated decision variable, otherwise, do not mutate, and thus obtain the mutated i-th offspring individual and the oth offspring individual ;in, is the mutation probability;

[0061] (11)

[0062] In formula (11), Represents a custom non-negative real number; Indicated in Random numbers in the interval;

[0063] (12)

[0064] In formula (12), express Middle The decision variables at each location, express Middle The decision variables after mutation at each position.

[0065] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the multi-criteria cancer image dose prediction method, and the processor is configured to execute the program stored in the memory.

[0066] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program executes the steps of the multi-criteria cancer image dose prediction method when the computer program is executed by a processor.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1) Deep neural networks are widely used in medical image dose prediction, but due to the differences in data sets, the versatility of these models is limited. This paper combines evolutionary algorithms with deep neural networks and optimizes the basic dose prediction model through evolutionary algorithms to improve the versatility of medical image dose prediction models.

[0069] 2) Using evolutionary algorithms to optimize the dose prediction basic model is a large-scale problem that requires a lot of computing resources and time. In order to improve the efficiency of the optimization process, the method of the present invention takes the weights of some network layers of U-net as the initial individuals of the evolutionary algorithm, thereby speeding up the optimization process and improving the efficiency of dose prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of the overall method of the present invention;

[0071] Figure 2 A schematic diagram of a deep neural network used in the present invention;

[0072] Figure 3 A schematic diagram of a deep neural network weight vector used in the present invention;

[0073] Figure 4 This is a flow chart of the evolutionary algorithm for optimizing the basic model for dose prediction in the present invention. DETAILED DESCRIPTION

[0074] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0075] In this embodiment, a multi-criteria medical image dose prediction method is provided. Figure 1 As shown, the following steps are followed:

[0076] Step 1: Obtain medical images and perform preprocessing;

[0077] Step 1.1, obtain the CT image of the head and neck, and then crop and normalize it to obtain a preprocessed CT image, the pixel size of the preprocessed CT image is 128×128×128;

[0078] Step 1.2, obtain PTV images of all target areas of the head and neck. In this embodiment, there are three target area PTV images, namely PTV70, PTV63 and PTV56. The target areas PTV70, PTV63 and PTV56 are merged into one target area PTV image, and the binary mask of the target area PTV image is recorded as PTVs. After processing, the pixel size of the PTV image is 128×128×128, and the pixel size of the binary mask PTVs of the PTV image is 128×128×128.

[0079] Step 1.3: Obtain a set of characteristic information of several organs at risk in the head and neck , and the binary mask of the feature information set OARs is recorded as ;

[0080] Step 1.4: Define the real dose distribution map of the head and neck ,make A single voxel The dose value in the real dose distribution diagram is ;

[0081] In this embodiment, the head and neck case data used is from the database provided by the American Association of Physicists in Medicine (AAPM) Open Knowledge Planning Grand Challenge (OpenKBP). The database contains data information of 340 patients with head and neck cancer, of which 200 cases are used as training sets, 40 cases are used as validation sets, and 100 cases are used as test sets; the information of each case data includes CT image tensor, contour structure (describing the area, shape and size of PTV and OARs) and dose distribution (Gy).

[0082] Step 2: Use the preprocessed medical images to train the dose prediction network to obtain a dose prediction basic model;

[0083] Step 2.1: Construct a dose prediction network based on the U-net model, and input the preprocessed CT image, the target area binary mask PTVs, and the OARs binary mask OMs into the dose prediction network. After the downsampling convolution layer extracts the information, the upsampling convolution layer restores the image to obtain the predicted dose distribution map. , where a single voxel In the predicted dose distribution map The dose value in is recorded as ;

[0084] In this embodiment, the U-net model consists of an input convolution layer (convolution kernel size is 3×3×3, step size is 1, and padding is 3), an output convolution layer (convolution kernel size is 1×1×1, step size is 1, and number is 1), M downsampling convolution modules (in this implementation example, M is set to 4, the convolution kernel size is 3×3×3, the step size is 2, the padding is 1, and the number of feature maps is 128, 64, 32, and 16 respectively) and M upsampling modules (in this implementation example, M is set to 4, the convolution kernel size is 3×3×3, the step size is 2, the padding is 1, and the number of feature maps is 16, 32, 64, and 128 respectively), as shown in FIG. Figure 2 As shown in the figure, each downsampling convolution layer consists of a downsampling layer, a convolution layer, an instance normalization layer, and a rectified linear unit, and each upsampling module consists of an upsampling layer, a convolution layer, an instance normalization layer, and a rectified linear unit. The feature map output by the mth downsampling convolution layer is mapped to the input of the corresponding mth upsampling deconvolution layer, and the feature map output by the mth downsampling convolution layer and the feature map output by the m-1th upsampling deconvolution layer are connected through the channel dimension to form the mth skip connection structure.

[0085] Step 2.2: Divide the dose distribution map using equations (1) and (2) , and obtain the target dose distribution map , Dose distribution map of organs at risk :

[0086] (1)

[0087] (2)

[0088] In formula (1) and formula (2), Represents a bitwise AND operation.

[0089] Step 2.3: Use formula (3) to construct the predicted dose distribution map Optimal dose map The loss function between , the weight parameters of the dose prediction network are updated through the back-propagation algorithm, and the dose prediction network is iteratively trained using the Adam optimizer (batchsize is set to 1, weight decay is set to 3e-5, initial learning rate is set to 1e-4 for optimization, and the maximum number of iterations is set to 100), and the network training is stopped when the maximum number of iterations is reached, thereby obtaining a dose prediction basic model;

[0090] (3)

[0091] In formula (3), represents any dose distribution diagram, and , , express The number of voxels in .

[0092] Step 3: Construct constraints and fitness function;

[0093] Step 3.1: Use equation (4) to construct the minimum dose equation for all voxels :

[0094] (4)

[0095] In formula (4), Represents a single voxel The dosage, is the minimum dose of the set dose constraint;

[0096] Step 3.2: Use equation (5) to construct the maximum dose constraint equation for all voxels :

[0097] (5)

[0098] In formula (5), The maximum dose of the set dose constraint.

[0099] Step 3.3: Use equation (6) to construct the optimal dose constraint equation for all voxels :

[0100] (6)

[0101] In formula (6), is the optimal dose for the set dose constraint;

[0102] Step 3.4: Use equation (7) to construct the dose-volume constraint equation for all voxels. :

[0103] (7)

[0104] In formula (7), is the maximum dose of the dose-volume constraint, express The maximum voxel dose in the medium dose;

[0105] Step 3.5: Use formula (8) to construct the first objective function :

[0106] (8)

[0107] Step 3.6: Use formula (9) to construct the second objective function :

[0108] (9)

[0109] Step 3.7: Use formula (10) to construct the fitness function :

[0110] (10)

[0111] In formula (10), and is 2 weights, in this implementation example and Both are set to 0.5.

[0112] Step 4: Use evolutionary algorithms to optimize the basic dose prediction model and obtain a dose prediction generalization model for multi-criteria dose prediction of medical images, such as Figure 4 As shown;

[0113] Step 4.1, initialize the population;

[0114] Step 4.1.1, define the maximum number of iterations as , the current number of iterations is , and initialize ;

[0115] Step 4.1.2: Select the weight vector of some layers in the dose prediction basic model And as a basic individual, in this implementation example, the last convolution layer is selected, and the convolution kernel of the last convolution layer is flattened into a weight vector as a basic individual, such as Figure 3 As shown, the optimization process is accelerated; after N random perturbations of the basic individuals, the G generation population is obtained ,in, represents the i-th individual of the G-th generation population, and each individual contains t decision variables, ,in, express is the jth decision variable in .

[0116] Step 4.2, individual fitness calculation;

[0117] Step 4.2.1. Restored to the weights of the corresponding layers in the dose prediction basic model , thus obtaining the Gth generation i-th dose prediction generalization model, and predicting the preprocessed CT images, target binary mask PTVs and OARs binary mask OMs to obtain the Gth generation i-th dose distribution map ;

[0118] Step 4.2.2: Use equation (1) and equation (2) to obtain the dose distribution map of the i-th target area of ​​the G-th generation: , the i-th organ at risk in the Gth generation , and then according to formula (4)-formula (10) we can get Fitness .

[0119] Step 4.3: According to the fitness of each individual, the G generation population Perform non-dominated sorting and calculate the crowding distance of the non-dominated front after sorting, so as to sort the individuals in ascending order according to the crowding distance, and finally obtain the G-generation population in ascending order;

[0120] Step 4.4: Perform crossover mutation on the ascending G-th generation population to generate the offspring population of the G-th generation population. ,in Represents the i-th individual in the offspring population.

[0121] Step 4.4.1: Based on the non-dominated frontier and its crowding distance, use the binary tournament method to randomly select the i-th individual from the G-th generation population in ascending order. and the oth individual , and after randomly selecting the crossover point k, exchange and The decision variables after the kth position, thus corresponding to the generation of the i-th offspring individual and the oth offspring individual ;

[0122] Step 4.4.2 Calculate using formula (11) and The variation value of each decision variable in , and in Generate random numbers in a range ,like , then use formula (12) to mutate the decision variable at the corresponding position to obtain the mutated decision variable, otherwise, do not mutate, and thus obtain the mutated i-th offspring individual and the oth offspring individual ;in, is the mutation probability;

[0123] (11)

[0124] In formula (11), Represents a user-defined non-negative real number. In this implementation example, Set to 1; Indicated in A random number in the interval.

[0125] (12)

[0126] In formula (12), express Middle The decision variables at each location, express Middle The decision variables after mutation at each position.

[0127] Step 4.5 Calculate according to formula (4)-formula (10) The fitness of each individual;

[0128] Step 4.6: Merge and According to the fitness of each individual, the G-th generation population after the merger is non-dominated and the crowding distance of the non-dominated frontier after sorting is calculated. Then, the individuals in the G-th generation population after the merger are sorted in ascending order according to the crowding distance, and finally the ascending non-dominated frontier is obtained, and the first N individuals are selected as the first Generation population .

[0129] Step 4.7, assign G+1 to G, and repeat steps 4.3 until G> So far, we get Generation population ;in, Indicates The i-th individual in the generation population;

[0130] Step 4.8: Each individual in is successively reduced to the weight of the corresponding layer in the basic model of dose prediction, thereby obtaining N generalized models of dose prediction, which are used to predict the preprocessed CT images, the binary masks of the target areas PTVs, and the binary masks of OARs OMs, and obtain N multi-criteria predicted dose distribution maps, which meet the needs of the intensity modulated radiation system to formulate radiation plans and improve the versatility of the medical image dose prediction model.

[0131] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0132] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0133] The above are specific embodiments of the present invention, but the protection scope of the present invention should not be limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention, so the protection scope of the present invention should be based on the protection scope defined in the claims.

Claims

1. A multi-criteria medical image dose prediction method, characterized in that: The steps are as follows: Step 1: Obtain medical images and perform preprocessing; Step 1.1, obtaining a CT image of the head and neck, and performing cropping and normalization to obtain a preprocessed CT image; Step 1.2, obtaining the PTV images of all the target areas of the head and neck and merging them into one target area PTV image, and recording the binary mask of the target area PTV image as PTVs; Step 1.3, obtaining characteristic information sets OARs of several organs at risk of the head and neck, and recording the binary mask of the characteristic information set OARs as OMs; Step 1.4: Define the actual dose distribution map of the head and neck ,make A single voxel The dose value in the real dose distribution diagram is ; Step 2: Use the preprocessed medical images to train the dose prediction network to obtain a dose prediction basic model; Step 2.1: Construct a dose prediction network based on the U-net model, and input the preprocessed CT image, the target binary mask PTVs, and the binary mask OMs of OARs into the dose prediction network to obtain a predicted dose distribution map. , where a single voxel In the predicted dose distribution map The dose value in is recorded as ; Step 2.2: Divide the dose distribution map using equations (1) and (2) , and obtain the target dose distribution map , Dose distribution map of organs at risk : (1) (2) In formula (1) and formula (2), Represents a bitwise AND operation; Step 2.3: Use formula (3) to construct the predicted dose distribution map Optimal dose map The loss function between , updating the weight parameters of the dose prediction network through a back-propagation algorithm, and stopping the network training when the maximum number of iterations is reached, thereby obtaining a dose prediction basic model; (3) In formula (3), represents any dose distribution diagram, and , , express The number of voxels in ; Step 3: Construct constraints and fitness function; Step 3.1: Use equation (4) to construct the minimum dose equation for all voxels : (4) In formula (4), Represents a single voxel The dosage, is the minimum dose of the set dose constraint; Step 3.2: Use equation (5) to construct the maximum dose constraint equation for all voxels : (5) In formula (5), is the maximum dose of the set dose constraint; Step 3.3: Use equation (6) to construct the optimal dose constraint equation for all voxels : (6) In formula (6), is the optimal dose for the set dose constraint; Step 3.4: Use equation (7) to construct the dose-volume constraint equation for all voxels. : (7) In formula (7), is the maximum dose of the dose-volume constraint, express The maximum voxel dose in the medium dose; Step 3.5: Use formula (8) to construct the first objective function : (8) Step 3.6: Use formula (9) to construct the second objective function : (9) Step 3.7: Use formula (10) to construct the fitness function : (10) In formula (10), and is 2 weights; Step 4: using an evolutionary algorithm to optimize the dose prediction basic model to obtain a dose prediction generalization model for performing multi-criteria dose prediction on the medical image; Step 4.1, initialize the population; Step 4.1.1, define the maximum number of iterations as , the current number of iterations is , and initialize ; Step 4.1.2: Select the weight vector of some layers in the dose prediction basic model And as a basic individual, after N random perturbations on the basic individual, the G generation population is obtained ,in, represents the i-th individual of the G-th generation population, and each individual contains t decision variables, ,in, express The j-th decision variable in ; Step 4.2, calculation Fitness ; Step 4.3: According to the fitness of each individual, the G generation population Perform non-dominated sorting and calculate the crowding distance of the non-dominated front after sorting, so as to sort the individuals in ascending order according to the crowding distance, and finally obtain the G-generation population in ascending order; Step 4.4: Perform crossover mutation on the ascending G-th generation population to generate the offspring population of the G-th generation population. ,in represents the i-th individual in the offspring population; Step 4.5 Calculate according to formula (4)-formula (10) The fitness of each individual; Step 4.6: Merge and According to the fitness of each individual, the G-th generation population after the merger is non-dominated and the crowding distance of the non-dominated frontier after sorting is calculated. Then, the individuals in the G-th generation population after the merger are sorted in ascending order according to the crowding distance, and finally the ascending non-dominated frontier is obtained, and the first N individuals are selected as the first Generation population ; Step 4.7, assign G+1 to G, and repeat steps 4.3 until G> So far, we get Generation population ;in, Indicates The i-th individual in the generation population; Step 4.8: Each individual in is successively reduced to the weight of the corresponding layer in the basic model of dose prediction, thereby obtaining N generalized models of dose prediction, which are used to predict the preprocessed CT images, binary masks of target areas PTVs, and binary masks of OARs OMs, and obtain N predicted dose distribution maps of multiple criteria.

2. A multi-criteria cancer image dose prediction method according to claim 1, characterized in that: The U-net model in step 2 is composed of an input convolution layer, an output convolution layer, M downsampling convolution modules and M upsampling modules, wherein each downsampling convolution layer is composed of a downsampling layer, a convolution layer, an instance normalization layer and a rectified linear unit, and each upsampling module is composed of an upsampling layer, a convolution layer, an instance normalization layer and a rectified linear unit; The feature map output by the mth downsampling convolution layer is mapped to the input of the corresponding mth upsampling deconvolution layer, and the feature map output by the mth downsampling convolution layer and the feature map output by the m-1th upsampling deconvolution layer are connected through the channel dimension to form the mth skip connection structure.

3. The multi-criteria cancer image dose prediction method according to claim 2, characterized in that: The step 4.2 comprises: Step 4.2.

1. Restored to the weights of the corresponding layers in the dose prediction basic model , thus obtaining the Gth generation i-th dose prediction generalization model, and predicting the preprocessed CT images, target binary mask PTVs and OARs binary mask OMs to obtain the Gth generation i-th dose distribution map ; Step 4.2.2: Use equations (1) and (2) to obtain the dose distribution map of the i-th target area of ​​the G-th generation: , Gth generation, i-th organ at risk , and then according to formula (4)-formula (10) we can get Fitness .

4. The multi-criteria cancer image dose prediction method according to claim 3, characterized in that: The step 4.4 comprises: Step 4.4.1: Based on the non-dominated frontier and its crowding distance, use the binary tournament method to randomly select the i-th individual from the G-th generation population in ascending order. and the oth individual , and after randomly selecting the crossover point k, exchange and The decision variables after the kth position, thus corresponding to the generation of the i-th offspring individual and the oth offspring individual ; Step 4.4.2 Calculate using formula (11) and The variation value of each decision variable in , and in Generate random numbers in a range ,like , then use formula (12) to mutate the decision variable at the corresponding position to obtain the mutated decision variable, otherwise, do not mutate, and thus obtain the mutated i-th offspring individual and the oth offspring individual ;in, is the mutation probability; (11) In formula (11), Represents a custom non-negative real number; Indicated in Random numbers in the interval; (12) In formula (12), express Middle The decision variables at each location, express Middle The decision variables after mutation at each position.

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the multi-criteria cancer image dose prediction method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-criteria cancer image dose prediction method according to any one of claims 1 to 4 are executed.

Images