Image registration method, system, device and medium based on partial differential adversarial learning
Through the method based on partial differential adversarial learning, the differential term alternative library is constructed using generative adversarial networks and prior knowledge, which solves the problems of high dependence on gold standards and low computational efficiency of deep learning image registration methods, and achieves high efficiency and reliability image registration effect.
Patent Information
- Application Number
- CN202210273643.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-19
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-19
AI Technical Summary
Existing deep learning image registration methods rely on high-quality gold standard data sets, which are difficult to obtain and have high accuracy dependence. Dual-supervised and weak-supervised methods cannot completely eliminate the dependence on supervised information, and the unsupervised methods are inefficient in computing, which can easily lead to Deformation field distortion.
Using a method based on partial differential adversarial learning, a differential term alternative library is constructed by generating adversarial networks and prior knowledge, learning differential term coefficients, increasing gradient penalty terms and regular terms, realizing sparse selection, obtaining the target partial differential equation, and applying the image Perform spatial deformation to complete registration.
Output a true and reliable smooth deformation field with higher computing efficiency, reducing network training parameters and computing power costs, improving image registration accuracy and efficiency, reducing dependence on gold standards, and avoiding the problem of deformation field distortion.
Smart Images

Figure CN114612535B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to an image registration method, system, device and medium based on partial differential adversarial learning. Background Art
[0002] The image registration problem is the process of finding a specific optimal geometric transformation to transform two or more images of the same scene in different coordinate systems into the same coordinate system. In the field of medical imaging, different imaging technologies obtain certain differences in morphological and functional information on the same tissue of the human body, but the information they provide is complementary to each other. In order to make full use of the information provided by these imaging methods, clinicians are eager to register and fuse different image information of the same patient and combine the imaging results, which requires registration of different images and information fusion technology. For multimodal images, registration is a necessary condition for image fusion. Registration is the premise of fusion and is also a key technology that determines the development of image fusion technology.
[0003] At present, medical image registration, as a key technology for image-assisted diagnosis and treatment, has been applied to multiple clinical scenarios such as disease diagnosis, surgical guidance, and radiotherapy. Deep learning has become a breakthrough in improving the performance of medical image registration. Its end-to-end application method and high-precision registration results are currently a research hotspot in the field of image registration.
[0004] According to the supervision information used in the network training process as the classification standard, deep medical image registration can be divided into fully supervised, double-supervised, weakly supervised and unsupervised registration methods.
[0005] There are several problems and defects in the existing deep learning-based image registration methods: (1) Fully supervised registration uses the true transformation (gold standard) between image pairs as supervision information, and uses the error loss between the true transformation parameters and the estimated transformation parameters to supervise network training. The registration accuracy is highly dependent on the quality of the gold standard. (2) Dual-supervised registration adds image similarity loss on top of the parameter error loss to reduce dependence on the gold standard. Weak supervision uses label information such as segmentation masks and biological key points to replace the gold standard, but dual supervision and weak supervision cannot completely eliminate the registration network's dependence on supervision information. (3) Unsupervised registration no longer requires any labeled data, and only uses image similarity loss to supervise network training. It is easy to estimate non-smooth deformation fields such as folding and voxel drift, resulting in deformation field distortion.
[0006] Through the above analysis, the problems and defects of the prior art are as follows:
[0007] (1) The registration accuracy that can be obtained by traditional fully supervised registration methods is highly dependent on the quality of the gold standard. However, it is extremely difficult to obtain such datasets with gold standards, and the setting of the gold standard is very inaccurate and cannot be used as a learning indicator.
[0008] (2) Dual supervision and weak supervision cannot completely eliminate the registration network’s dependence on supervisory information, and face the same dilemma as the fully supervised configuration method, namely, it is very difficult to obtain high-quality datasets.
[0009] (3) Most existing unsupervised registrations use 3D medical image registration networks. Compared with 2D registration networks, the trainable parameters of this type of network grow exponentially with the number of convolutional layers, and the extremely high complexity leads to low computational efficiency.
[0010] To this end, the present invention proposes an image registration method, system, device and medium based on partial differential adversarial learning. Summary of the invention
[0011] To solve the above problems, the present invention provides an image registration method, system, device and medium based on partial differential adversarial learning. The method can output a real and reliable smooth deformation field with higher computing efficiency, greatly reducing the trainable parameters in the network training process, reducing computing power costs, and further promoting the application of deep medical image registration technology in clinical diagnosis and treatment.
[0012] The present invention provides the following technical solutions.
[0013] An image registration method based on partial differential adversarial learning comprises the following steps:
[0014] Using prior knowledge and basis function representation to construct a library of candidate differential terms;
[0015] Based on the basic structure of the generative adversarial network, the corresponding coefficients of the differential terms in the differential term candidate library are learned; the generator of the generative adversarial network adopts multiple single-layer regression networks;
[0016] Add gradient penalty terms and regularization terms to the loss function of the generative adversarial network to achieve sparse selection of differential terms; obtain the target partial differential equation through the trained generative adversarial network;
[0017] The target partial differential equation is used to spatially deform the preprocessed image to be registered to obtain the registered image.
[0018] Preferably, the method further includes preprocessing the image data set to be registered in sequence, including the following steps:
[0019] Perform structural segmentation on the image to be registered and adaptively remove some structures;
[0020] The processed image to be registered is gray-scale equalized and spatially normalized to obtain the preprocessed image to be registered.
[0021] Preferably, the structural segmentation of the image to be registered comprises the following steps:
[0022] The image is resampled to a space with a DPI of 1mm. The original image size is 256×256×256. Each image is cut to remove all image layers without values and only retain layers with values so that the entire target structure is in the center of the image. After cutting, the image is scaled to a size of 64×64×64.
[0023] Preferably, the method of constructing a candidate library of differential terms using prior knowledge and basis function representation method comprises the following steps:
[0024] The partial differential equation is an equation containing partial derivatives or partial differentials of an unknown function. Its general form is expressed as:
[0025]
[0026] Where t is the time variable, x and y are space variables; the partial differential equation u t It is composed of several differential terms x, y, u, u x 、u y 、u xx 、u xy 、u yy Function of composition;
[0027] The above formula can be expressed as follows:
[0028] U lib =Θ(u,u x ,u xx ,...)W
[0029] Among them, Θ represents the candidate library matrix composed of various differential terms, and W represents the corresponding coefficient of the differential term, that is:
[0030] First, construct a candidate library containing all differential terms, then learn the coefficient matrix W composed of the coefficients corresponding to each differential term, and finally obtain the target partial differential equation;
[0031] The overall construction of the candidate library is based on the basis function representation method, which calculates u and u respectively. x 、u y 、u xx 、u xy 、u yy Basic differential terms, and for specific partial differential equation learning objectives, that is, corresponding prior knowledge, add differential terms that have a strong correlation with prior knowledge into the candidate library;
[0032] The method of constructing differential terms selected when constructing the differential term candidate library uses the Chebyshev polynomial interpolation method.
[0033] Preferably, the method of learning corresponding coefficients of differential terms in a differential term candidate library based on a basic structure of a generative adversarial network specifically includes the following steps:
[0034] The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1.
[0035] The entire optimization process follows the following loss function:
[0036]
[0037] in, and x represent the generated samples and real samples respectively, P r and P g They represent the sample distribution generated by the real samples respectively.
[0038] Preferably, the step of adding a gradient penalty term and a regularization term to the loss function of the generative adversarial network comprises the following steps:
[0039] Modify the loss function to:
[0040]
[0041] in, Indicates that in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; 1 and λ 2 L 1 and L 2 The weight coefficient, regularization loss L 1 and the regularization loss L 2 Specifically expressed as:
[0042]
[0043]
[0044] Among them, Param G Param DRepresent the network weight parameters of the generator and the discriminator respectively.
[0045] Preferably, the step of spatially deforming the preprocessed image to be registered by using the target partial differential equation comprises:
[0046] The target partial differential equation learned by the network is used to perform temporal differential calculations on the image to be registered, and the deformed image is output.
[0047] An image registration system based on partial differential adversarial learning, comprising:
[0048] A differential term candidate library construction module is used to construct a candidate library, and uses prior knowledge and basis function representation to select differential terms to construct the differential term candidate library;
[0049] The differential term coefficient learning module is used to utilize the basic structure of the generative adversarial network, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library;
[0050] Gradient penalty and regularized network convergence module, which adds gradient penalty and regularization terms to the loss function, realizes sparse selection of differential terms, and obtains the target partial differential equation;
[0051] The image deformation module performs time-direction differential calculation on the image to be registered according to the target partial differential equation, thereby obtaining the deformed image and completing the image registration.
[0052] A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the following steps:
[0053] Using prior knowledge and basis function representation to construct a library of candidate differential terms;
[0054] The basic structure of the generative adversarial network is used, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library;
[0055] Add gradient penalty terms and regularization terms to the loss function to achieve sparse selection of differential terms and obtain the target partial differential equation;
[0056] The target partial differential equation is used to perform spatial deformation on the image to be registered to obtain the registered image.
[0057] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the following steps:
[0058] Using prior knowledge and basis function representation to construct a library of candidate differential terms;
[0059] The basic structure of the generative adversarial network is used, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library;
[0060] Add gradient penalty terms and regularization terms to the loss function to achieve sparse selection of differential terms and obtain the target partial differential equation;
[0061] The target partial differential equation is used to perform spatial deformation on the image to be registered to obtain the registered image.
[0062] Beneficial effects of the present invention:
[0063] The present invention utilizes software to remove useless tissue structures from initial image data and performs size cropping, thereby improving the utilization rate of target tissue structures; utilizes prior knowledge and basis function representation to construct a differential term candidate library, thereby making the candidate library complete and reasonable and reducing the amount of calculation; then utilizes the basic structure of a generative adversarial network, wherein a generator adopts multiple single-layer regression networks to learn corresponding coefficients of differential terms in the candidate library; then, gradient penalty terms and regularization terms are added to the adversarial loss to help the network converge faster, thereby achieving sparse selection of differential terms and obtaining a high-precision target partial differential equation; finally, the image to be registered is differentially calculated in the time direction according to the learned target partial differential equation, thereby obtaining a deformed image and completing image registration.
[0064] Compared with other existing image registration methods, the present invention can output a true and reliable smooth deformation field at a higher computational efficiency, greatly reducing the trainable parameters in the network training process and reducing the computing power cost. The present invention designs a deep differential equation network based on adversarial learning to complete image registration, which is used to solve the problem that the traditional fully supervised and supervised image registration methods are too dependent on the setting of the gold standard, and the gold standard is easily affected by human subjective factors, and it is difficult to obtain a high-quality data set of evaluation indicators. In addition, the unsupervised method also has the problem of low computational efficiency caused by distortion of the deformation field and too many learnable parameters. The deep differential equation network based on adversarial learning can learn a rich type of differential terms of partial differential equations, and has stronger initial value robustness in processing tasks that are more sensitive to final value data. The trainable parameters in the deep differential equation network based on adversarial learning are greatly reduced, which reduces the requirements for computing power, improves the registration effect, and improves the operating efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 is a flow chart of an image registration method based on partial differential adversarial learning according to an embodiment of the present invention;
[0066] Figure 2 is a structural schematic diagram of an image registration system based on partial differential adversarial learning according to an embodiment of the present invention;
[0067] Figure 3is a flowchart of an implementation method of an image registration method based on partial differential adversarial learning according to an embodiment of the present invention;
[0068] Figure 4 It is the image registration result of the image registration method based on partial differential adversarial learning in an embodiment of the present invention. DETAILED DESCRIPTION
[0069] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0070] Example 1
[0071] The present invention provides an image registration method based on partial differential adversarial learning, such as Figure 1 and Figure 3 As shown, the following steps are included:
[0072] S1: Preprocess the image dataset to be registered in sequence:
[0073] The FreeSurfer tool was used to adaptively remove the skull structure in the brain MRI image, so that only the brain tissue was retained in the image and its structural segmentation mark was obtained. Then the grayscale equalization and spatial normalization of the entire brain image were performed, so that the grayscale value of the image was limited to a certain range and the spatial position was at the origin.
[0074] The parameters for data preprocessing were set as follows: the image was resampled to a space with a DPI of 1 mm. The original image size was 256×256×256, and each image was cut to remove all image layers without values, leaving only layers with values, so that the entire brain was in the center of the image, and the image was scaled to 64×64×64 after cutting.
[0075] S2: Construct a library of candidate differential terms using prior knowledge and basis function representation:
[0076] A partial differential equation is an equation involving partial derivatives or partial differentials of an unknown function. Its general form is:
[0077]
[0078] Where t is the time variable, x and y are space variables; the partial differential equation u t It is composed of several differential terms x, y, u, u x 、u y 、u xx 、u xy 、u yy Function of composition;
[0079] The above formula can be expressed as follows:
[0080] U lib =Θ(u,u x ,u xx ,...)W
[0081] Among them, Θ represents the candidate library matrix composed of various differential terms, and W represents the corresponding coefficient of the differential term, that is:
[0082] First, construct a candidate library containing all differential terms, then learn the coefficient matrix W composed of the coefficients corresponding to each differential term, and finally obtain the target partial differential equation;
[0083] The overall construction of the candidate library is based on the basis function representation method, which calculates u and u respectively. x 、u y 、u xx 、u xy 、u yy Basic differential terms, and according to the specific partial differential equation learning objectives, that is, the corresponding prior knowledge, the differential terms that are strongly correlated with the prior knowledge are added into the candidate library; the size of the candidate library needs to be appropriate. Too large or too small will affect the accuracy of the learned form of partial differential equations and the precision of the learned coefficients, and will also affect the efficiency of network learning.
[0084] When constructing the differential term candidate library, the methods of constructing differential terms selected include forward difference, backward difference, central difference, and Chebyshev polynomial interpolation method, and the Chebyshev polynomial interpolation method is mainly selected.
[0085] S3: Based on the basic structure of the generative adversarial network, learn the corresponding coefficients of the differential terms in the differential term candidate library; the generator of the generative adversarial network adopts multiple single-layer regression networks:
[0086] The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1.
[0087] The entire optimization process follows the following loss function:
[0088]
[0089] in, and x represent the generated samples and real samples respectively, P r and P gThey represent the sample distribution generated by the real samples respectively.
[0090] S4: Add gradient penalty terms and regularization terms to the loss function of the generative adversarial network to achieve sparse selection of differential terms; obtain the target partial differential equation through the trained generative adversarial network:
[0091] Modify the loss function to:
[0092]
[0093] in, Indicates that in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; 1 and λ 2 L 1 and L 2 The weight coefficient, regularization loss L 1 and the regularization loss L 2 Specifically expressed as:
[0094]
[0095]
[0096] Among them, Param G Param D Represent the network weight parameters of the generator and the discriminator respectively.
[0097] Specifically, L 1 Loss helps in the sparse selection of parameters. Sparsity helps the neural network remove unimportant features during the learning process and avoids too many parameters affecting our interpretation and analysis of the model. 2 The loss can shorten the training time and help unimportant features decay quickly, but its non-sparse nature avoids overfitting and underfitting of the network.
[0098] S5: Use the target partial differential equation to spatially deform the preprocessed image to be registered to obtain a registered image.
[0099] like Figure 2 As shown, the present invention provides an image registration system, comprising:
[0100] The data preprocessing module is used to remove useless tissue structures from the initial image data and cut the size;
[0101] A differential term candidate library construction module is used to construct a candidate library, and uses prior knowledge and basis function representation to select differential terms to construct the differential term candidate library;
[0102] The differential term coefficient learning module is used to utilize the basic structure of the generative adversarial network, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library;
[0103] Gradient penalty and regularized network convergence module, which adds gradient penalty and regularization terms to the loss function, realizes sparse selection of differential terms, and obtains the target partial differential equation;
[0104] The image deformation module performs time-direction differential calculation on the image to be registered according to the target partial differential equation, thereby obtaining the deformed image and completing the image registration.
[0105] In this embodiment, the technical effect of the present invention is described in detail through simulation.
[0106] Simulation conditions:
[0107] The present invention uses Python language to complete the simulation experiment of the present invention on a PC with an Intel(R) Core(TM) i7-9700 CPU, a CPU of 3.60 GHz, a RAM of 16.00 GB, and an Ubuntu 16.04 operating system.
[0108] Simulation experiment content and result analysis:
[0109] The experimental data of the present invention are data generated by various classical partial differential equations and real noisy data generated when partial differential equations are used for image processing. The classical partial differential equations include Burgers equation, KdV equation, diffusion equation and Hamilton equation. The experimental results are shown in Table 1. The image registration results are shown in Table 1. Figure 4 As shown in FIG. 1 , there are the image to be registered, the deformed image and the template image.
[0110] Table 1 Effect of the method of the present invention on learning classical partial differential equations
[0111]
[0112] According to the above results, the deep differential equation network based on adversarial learning can not only accurately complete the learning of partial differential equations, but also realize image registration based on the partial differential equations learned. Compared with other existing image registration methods, this method can output a real and reliable smooth deformation field at a higher computational efficiency, greatly reducing the trainable parameters in the network training process and reducing the computing cost. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or a dedicated hardware design. A person of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer executable instructions and / or included in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.
[0113] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. An image registration method based on partial differential adversarial learning, characterized in that: include: Using prior knowledge and basis function representation to construct a library of candidate differential terms; Based on the basic structure of the generative adversarial network, learn the corresponding coefficients of the differential terms in the differential term candidate library; The generator of the generative adversarial network adopts multiple single-layer regression networks; Add gradient penalty terms and regularization terms to the loss function of the generative adversarial network to achieve sparse selection of differential terms; obtain the target partial differential equation through the trained generative adversarial network; The target partial differential equation is used to spatially deform the preprocessed image to be registered to obtain the registered image; The method of learning the corresponding coefficients of the differential terms in the differential term candidate library based on the basic structure of the generative adversarial network specifically includes the following steps: The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1. The entire optimization process follows the following loss function: in, and x represent the generated samples and real samples respectively, Ρ r and P g They represent the sample distribution generated by real samples respectively; The step of adding a gradient penalty term and a regularization term to the loss function of the generative adversarial network comprises the following steps: Modify the loss function to: in, Indicated in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; λ1 and λ2 are the weight coefficients of L1 and L2 respectively. The regularization loss L1 and the regularization loss L2 are specifically expressed as: Among them, Param G Param D Represent the network weight parameters of the generator and the discriminator respectively.
2. The image registration method based on partial differential adversarial learning according to claim 1, characterized in that: The method also includes preprocessing the image data set to be registered in sequence, including the following steps: Perform structural segmentation on the image to be registered and adaptively remove some structures; The processed image to be registered is gray-scale equalized and spatially normalized to obtain the preprocessed image to be registered.
3. The image registration method based on partial differential adversarial learning according to claim 2, characterized in that: The method of performing structural segmentation on the image to be registered comprises the following steps: The image is resampled to a space with a DPI of 1mm. The original image size is 256×256×256. Each image is cut to remove all image layers without values and only retain layers with values so that the entire target structure is in the center of the image. After cutting, the image is scaled to a size of 64×64×64.
4. The image registration method based on partial differential adversarial learning according to claim 1, characterized in that: The method of constructing a differential term candidate library by using prior knowledge and basis function representation method comprises the following steps: The partial differential equation is an equation containing partial derivatives or partial differentials of an unknown function. Its general form is expressed as: Where t is the time variable, x and y are space variables; the partial differential equation u t It is composed of several differential terms x, y, u, u x 、u y 、u xx 、u xy 、u yy Function of composition; The above formula can be expressed as follows: IN lib =Θ(u,u x ,in xx ,...)W Among them, Θ represents the candidate library matrix composed of various differential terms, and W represents the corresponding coefficient of the differential term, that is: First, construct a candidate library containing all differential terms, then learn the coefficient matrix W composed of the coefficients corresponding to each differential term, and finally obtain the target partial differential equation; The overall construction of the candidate library is based on the basis function representation method, which calculates u and u respectively. x 、u y 、u xx 、u xy 、u yy Basic differential terms, and for specific partial differential equation learning objectives, that is, corresponding prior knowledge, add differential terms that have a strong correlation with prior knowledge into the candidate library; The method of constructing differential terms selected when constructing the differential term candidate library uses the Chebyshev polynomial interpolation method.
5. The image registration method based on partial differential adversarial learning according to claim 1, characterized in that: The method of using the target partial differential equation to spatially deform the preprocessed image to be registered includes: The target partial differential equation learned by the network is used to perform temporal differential calculations on the image to be registered, and the deformed image is output.
6. An image registration system based on partial differential adversarial learning, characterized in that: include: A differential term candidate library construction module is used to construct a candidate library, and uses prior knowledge and basis function representation to select differential terms to construct the differential term candidate library; The differential term coefficient learning module is used to utilize the basic structure of the generative adversarial network, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library; Gradient penalty and regularized network convergence module, which adds gradient penalty and regularization terms to the loss function, realizes sparse selection of differential terms, and obtains the target partial differential equation; The image deformation module performs time-dependent differential calculations on the image to be registered according to the target partial differential equation, thereby obtaining the deformed image and completing the image registration. The learning of the corresponding coefficients of the differential terms in the candidate library specifically includes the following steps: The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1. The entire optimization process follows the following loss function: in, and x represent the generated samples and real samples respectively, Ρ r and P g They represent the sample distribution generated by real samples respectively; Adding a gradient penalty term and a regularization term to the loss function includes the following steps: Modify the loss function to: in, Indicated in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; λ1 and λ2 are the weight coefficients of L1 and L2 respectively. The regularization loss L1 and the regularization loss L2 are specifically expressed as: Among them, Param G Param D Represent the network weight parameters of the generator and the discriminator respectively.
7. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Using prior knowledge and basis function representation to construct a library of candidate differential terms; The basic structure of the generative adversarial network is used, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library; Add gradient penalty terms and regularization terms to the loss function to achieve sparse selection of differential terms and obtain the target partial differential equation; The target partial differential equation is used to spatially deform the image to be registered to obtain the registered image; The learning of the corresponding coefficients of the differential terms in the candidate library specifically includes the following steps: The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1. The entire optimization process follows the following loss function: in, and x represent the generated samples and real samples respectively, Ρ r and P g They represent the sample distribution generated by real samples respectively; Adding a gradient penalty term and a regularization term to the loss function includes the following steps: Modify the loss function to: in, Indicated in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; λ1 and λ2 are the weight coefficients of L1 and L2 respectively. The regularization loss L1 and the regularization loss L2 are specifically expressed as: Among them, Param G Param D Represent the network weight parameters of the generator and the discriminator respectively.
8. A computer-readable storage medium, characterized in that: The computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following steps: Using prior knowledge and basis function representation to construct a library of candidate differential terms; The basic structure of the generative adversarial network is used, in which the generator uses multiple single-layer regression networks to learn the corresponding coefficients of the differential terms in the candidate library; Add gradient penalty terms and regularization terms to the loss function to achieve sparse selection of differential terms and obtain the target partial differential equation; The target partial differential equation is used to spatially deform the image to be registered to obtain the registered image; The learning of the corresponding coefficients of the differential terms in the candidate library specifically includes the following steps: The basic structure of the generative adversarial network is adopted, in which the generator of the generative adversarial network is stacked with n single-layer regression networks. A single sub-network continuously iterates the corresponding coefficient matrix W of the differential term in the candidate library through regression. The generator outputs the predicted sample after n multiple time steps and sends it to the discriminator of the generative adversarial network for scoring. The discriminator marks the generated predicted sample with a logical value of 0, and marks the real observed data with a logical value of 1. The entire optimization process follows the following loss function: in, and x represent the generated samples and real samples respectively, Ρ r and P g They represent the sample distribution generated by real samples respectively; Adding a gradient penalty term and a regularization term to the loss function includes the following steps: Modify the loss function to: in, Indicated in P r and P g The probability distribution of joint random sampling is specifically expressed as λ gp is the weight coefficient of the gradient penalty term; λ1 and λ2 are the weight coefficients of L1 and L2 respectively. The regularization loss L1 and the regularization loss L2 are specifically expressed as: Among them, Param G Param D Represent the network weight parameters of the generator and the discriminator respectively.
