Registration method for nuclear magnetic images of any size

Abstract

A registration method for nuclear magnetic images of any size comprises the steps that a brain magnetic resonance data set is scanned, and data are preprocessed; constructing a nuclear magnetic image registration model based on sparse attention, taking paired images as floating images and fixed images respectively, and inputting the images into the designed nuclear magnetic image registration model for training; inputting the floating image and the up-sampled deformation field into a spatial transformation network to generate a registered image; calculating the difference between the fixed image and the registered image by using the designed objective function; performing iterative optimization training on the nuclear magnetic image registration model until model training convergence, and then storing an optimal weight parameter; and inputting test data into the nuclear magnetic image registration model, and loading the optimal weight parameter obtained in the training stage to obtain a registration result. According to the method, algorithm complexity optimization is carried out on the attention mechanism, the overall time complexity of the algorithm is low, meanwhile, the needed calculation memory is less, and the weight is light.

Description

Technical Field

[0001] The invention belongs to the technical field of medical image processing, and relates to an image registration method, and in particular to a nuclear magnetic resonance image registration method of any size. Background Art

[0002] In the field of medical imaging, MRI registration technology is a vital tool that can align images from different imaging devices or at different time points to assist in diagnosis and treatment planning. With the development of medical imaging technology, the use of MRI data has become increasingly important, such as combining preoperative / intraoperative MRI (magnetic resonance imaging) taken at different times to provide doctors with more comprehensive information. However, due to significant differences between images, such as resolution, contrast, and noise levels, the registration process faces many challenges. Traditional registration methods often rely on dense feature matching and complex optimization strategies, which are difficult to apply when computing resources are limited. For example: dense feature matching refers to finding the correspondence between images or data through artificially designed features (such as SIFT, SURF, etc.), and it is also necessary to design complex mathematical models. The design of the model and the implementation of the optimization strategy are relatively complex.

[0003] In summary, traditional registration methods face problems such as high computational complexity, poor real-time performance, and inability to handle input data of arbitrary sizes when processing MRI medical images. How to solve the above problems is also a task that technicians in this field need to solve urgently. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a registration method for MRI images of arbitrary sizes, which is based on a lightweight design of sparse attention, combined with deformation field super-resolution upsampling technology, feature sparse representation, attention mechanism and multi-layer perceptron technology, to provide a more accurate and efficient tool for clinical diagnosis and treatment.

[0005] A method for registering nuclear magnetic resonance images of arbitrary size comprises the following steps:

[0006] S1, scan brain magnetic resonance imaging data set, preprocess the data and divide it into training set, validation set and test set;

[0007] S2. Construct a sparse attention-based MRI image registration model, treat paired images as floating images and fixed images respectively, perform downsampling on the input images, and input the processed images into the designed MRI image registration model for training;

[0008] S3, performing super-resolution upsampling processing on the deformation field representing the difference between images obtained through training, and inputting the floating image and the upsampled deformation field into the spatial transformation network to generate the registered image;

[0009] S4, using the designed objective function to calculate the difference between the fixed image and the registered image, and performing back propagation optimization on the nuclear magnetic resonance image registration model;

[0010] S5, performing iterative optimization training on the nuclear magnetic resonance image registration model until the model training converges, terminating the iterative training, and then saving the optimal weight parameters;

[0011] S6. Input the test data into the MRI image registration model, and load the optimal weight parameters obtained in the training phase to obtain the registration result.

[0012] Furthermore, the design process of the MRI registration model in step S3 is as follows:

[0013] S31. Construct a feature extraction network, perform four-fold downsampling using a sparse Transformer structure, and implement network encoding of the input data.

[0014] The formula of the designed sparse Transformer structure is defined as follows:

[0015] SHMA(X l-1 )=Concat(SHMA1,…,SHMA k )W o

[0016]

[0017] Among them, SHMA Hi represents the sparse multi-head attention mechanism, Hi∈[1,k] represents the i-th head, MLP represents the multi-layer perceptron, LN represents the layer normalization (Laver Normalization), X l-1 and X l Represent the input and output of sparse attention respectively, and the mapping matrix W o Used to project the self-attention result to the target output dimension;

[0018] The structural formula of the sparse multi-head attention mechanism is defined as follows:

[0019]

[0020] Q, K, and V represent the feature maps after mapping, respectively. i ,k i ,v i denotes the i-th row of Q, K, and V respectively. The probability form of the attention of the i-th query to all keys is defined as Attn(q i ,k j )=p(k j |q i ),q(k j |qi ) represents a uniform distribution, when k(q i ,k j ) / Σ l k(q i ,k l )=1 / L K , the similarity is measured by Kullback-Leibler divergence, which is defined as follows:

[0021]

[0022] Remove the constant term lnL k After that, the sparsity of the i-th query is defined as M(q i ,K),

[0023]

[0024] is M(q i ,K) is a simplified form, a sparse matrix and q have the same size,

[0025]

[0026] SHMA Hi (Q,K,V)=Attn(Q,K)

[0027] SHMA(X l-1 )=Concat(SHMA1,…,SHMA k )W o

[0028] Among them, Attn represents sparse attention processing, represents the scale factor, L k represents the sequence length, d k represents the sequence dimension;

[0029] S32, using deconvolution upsampling to gradually restore the size of the feature map to achieve network decoding of the output data; the corresponding layers of the network encoding and decoding are connected using a skip connection structure;

[0030] S33, outputting a deformation field representing the difference between images.

[0031] Furthermore, in step S3, the deformation field is processed using a super-resolution network structure to increase the image resolution by N times, so that the network can process nuclear magnetic resonance images of any size;

[0032] The input of the super-resolution network structure includes low-resolution data and high-resolution gold standard data In the encoder network, low-resolution data is input It is then converted into a feature map

[0033] The feature map is obtained by linear interpolation Among them, k i is the upsampling scale of the i∈[1,M]th image, M is the number of training samples;

[0034] The eigenvector is obtained by dimensionality reduction and high-resolution images The pixel coordinates C = (x, y, x) are concatenated and then input into the decoding network to obtain the predicted pixel intensity. and compared with the pixel intensity in the true label Perform loss calculation; the loss function is defined as follows:

[0035]

[0036] Where D is the number of input images, represents the coordinate sampling in each high-resolution gold standard image, is the predicted pixel intensity, is the true pixel intensity in the high-resolution gold standard data.

[0037] Furthermore, the objective function designed in step S4 is defined as follows:

[0038]

[0039] in, It is composed of mean square error and is used to represent the fixed image I f and distorted images The similarity between m represents the floating image, φ represents the registered deformation field, represents the smoothness loss of the deformation field, represents the weak supervision loss, λ represents the hyperparameter of the deformation field smoothing loss, and μ represents the hyperparameter of the weak supervision loss;

[0040] Similarity loss consisting of mean square error The definition is as follows:

[0041]

[0042] Among them, I f Represents a fixed image, represents the distorted image, Ω represents the n-dimensional feature space, and p represents the pixel in the image;

[0043] The weak supervision loss is defined as follows:

[0044]

[0045] in, They represent the segmentation labels of the fixed image and the segmentation labels of the deformed floating image, respectively. K represents the total number of regions. represents the kth sub-region of the floating segmentation label;

[0046] Deformation field smoothing loss The definition is as follows:

[0047]

[0048] in, represents the image volume neighborhood, represents the first-order derivative of the deformation field φ(·), represents the second-order derivative of the deformation field (·).

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This application processes the nuclear magnetic resonance imaging data through a lightweight feature extraction network with sparse attention to obtain a sparse feature representation of the image, which can significantly reduce the computational complexity and storage requirements while ensuring good registration performance. Among them, the use of a sparse attention mechanism to focus on key areas in the image, thereby reducing the interference of irrelevant information on the registration process, and improving the efficiency and robustness of the algorithm is the key to the design of the feature extraction network. For resource-constrained environments such as mobile devices or server platforms, lightweight registration methods aim to reduce model parameters and computational complexity to achieve rapid response, so sparse ideas are incorporated into the design of attention. In addition, the sampling strategy on the design of the deformation field enables small-size data to train a model suitable for registration of any large-size data.

[0051] The following is a further description of the scheme of the application in conjunction with the accompanying drawings and embodiments: BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of the registration method for MRI images of any size;

[0053] Figure 2 Aligning network structure graphs for sparse attention;

[0054] Figure 3 Schematic diagram of the sparse Transformer structure;

[0055] Figure 4 Schematic diagram of the sparse multi-head attention structure;

[0056] Figure 5This is the flow chart of the super-resolution network. DETAILED DESCRIPTION

[0057] The embodiments of the technical solution of the present application will be described in detail below in conjunction with the accompanying drawings. Unless otherwise specified, the technical terms or scientific terms used in the present application are generally understood by those skilled in the art.

[0058] Reference Figure 1 The present invention provides a method for registering MRI images of any size, comprising the following steps:

[0059] S1, scan brain magnetic resonance imaging data set, preprocess the data and divide it into training set, validation set and test set according to 7:1:2;

[0060] S2. Construct a sparse attention-based MRI image registration model, and treat paired images as floating images and fixed images respectively. The function of the MRI image registration model is to register floating images to fixed images. There are two situations during the training process. Increase the diversity of training, and the training data will be used as floating and fixed images respectively. Each training of the model requires the input of two images. When input into the network, one image will be used as floating and the other image will be used as fixed. The next training may be reversed. Downsampling the input image can reduce the size of the input image while retaining key information, effectively reducing the computational complexity of the network and improving training and inference efficiency. Input the processed image into the designed MRI image registration model for training;

[0061] S3, super-sampling the deformation field obtained by training to represent the difference between images to improve the resolution of the deformation field, realize the registration of input images of any size, and input the floating image and the upsampled deformation field into the spatial transformation network to generate the registered image;

[0062] S4, using the designed objective function to calculate the difference between the fixed image and the registered image, and performing back propagation optimization on the nuclear magnetic resonance image registration model to update the weights to minimize the loss function;

[0063] S5, performing iterative optimization training on the nuclear magnetic resonance image registration model until the model training converges, terminating the iterative training, and then saving the optimal weight parameters;

[0064] S6. Input the test data into the MRI image registration model, and load the optimal weight parameters obtained in the training phase to obtain the registration result.

[0065] This implementation method designs sparse attention to reduce the computational complexity of the model, while allowing the model to maintain good long-distance dependent feature extraction capabilities. Design an image downsampling module to reduce network computational complexity by reducing the data size. In addition, the super-resolution upsampling technology is integrated to increase the deformation field size, so that the small parameter model can effectively cope with the registration of data of any size. Through the above-mentioned key technical structure design, the nuclear magnetic image registration method is lightweight, and the application range of the algorithm is made wider, and accurate registration results can be inferred in real time at a lower computational cost.

[0066] In step S1, the nuclear magnetic resonance images need to be preprocessed, including normalization and affine transformation; normalization is conducive to network feature extraction and accelerates network convergence. The normalization process is defined as follows:

[0067]

[0068] Among them, x represents the pixel in the image, and after mapping, the characteristic value of the pixel is mapped to Eliminate the dimensionality effect between the data. Afterwards, use affine transformation to roughly register the floating image to the fixed image, which is beneficial to the convergence of the subsequent registration network.

[0069] In step S2, the input image is subjected to nearest neighbor downsampling, and the input image size is reduced to the original size using nearest neighbor downsampling. N represents the multiple of image size reduction. It can reduce the size of the input image while retaining key information, effectively reduce the computational complexity of the network, and improve training and reasoning efficiency.

[0070] In step S3, a sparse attention-based MRI registration model is constructed, such as Figure 2 As shown, the design process of the nuclear magnetic resonance image registration model is as follows:

[0071] S31. Construct a feature extraction network, perform four-fold downsampling using a sparse Transformer structure, and implement network encoding of the input data.

[0072] The designed sparse Transformer structure is as follows Figure 3 As shown, the formula is defined as follows:

[0073] SHMA(X l-1 )=Concat(SHMA1,…,SHMA k )W o

[0074]

[0075] Among them, SHMA Hirepresents the sparse multi-head attention mechanism, Hi∈[1,k] represents the i-th head, MLP represents the multi-layer perceptron, LN represents the layer normalization, and X l-1 and X l Represent the input and output of sparse attention respectively, and the mapping matrix W o Usually used to project the self-attention result to the target output dimension;

[0076] The structure of sparse multi-head attention mechanism SHMA is as follows Figure 4 As shown, the formula is defined as follows:

[0077]

[0078] Q, K, and V represent the feature maps after mapping, respectively. i ,k i ,v i denotes the i-th row of Q, K, and V respectively. The probability form of the attention of the i-th query to all keys is defined as Attn(q i ,k j )=p(k j |q i ),q(k j |q i ) represents a uniform distribution, when k(q i ,k j ) / ∑ l k(q i ,k l )=1 / L K , the similarity is measured by Kullback-Leibler divergence, which is defined as follows:

[0079]

[0080] Remove the constant term lnL k After that, the sparsity of the i-th query is defined as M(q i ,k),

[0081]

[0082] is M(q i ,k) is a simplified form, a sparse matrix and q have the same size,

[0083]

[0084]

[0085] SHMA Hi(Q,K,V)=Attn(Q,K)V

[0086] SHMA(X l-1 )=Concat(SHMA1,…,SHMA k )W o

[0087] Among them, Attn represents sparse attention processing, represents the scale factor, L k represents the sequence length, d k represents the sequence dimension;

[0088] S32, using deconvolution upsampling to gradually restore the size of the feature map to achieve network decoding of the output data; the corresponding layers of the network encoding and decoding are connected using a skip connection structure;

[0089] S33. Output a displacement vector field (DVF) representing the difference between images.

[0090] This implementation optimizes the algorithm complexity of the attention mechanism, and the overall time complexity of the algorithm is low. At the same time, it requires less computing memory and is lightweight.

[0091] In step S3, the deformation field is upsampled using a super-resolution network structure to increase the image resolution by N times, so that the network can process nuclear magnetic resonance images of any size; the super-resolution network structure is as follows: Figure 5 As shown in Figure 2, the 3D super-resolution network is used to improve the resolution of the deformation field and realize the registration of input images of any size.

[0092] The input of the super-resolution network structure includes low-resolution data and high-resolution gold standard data In the encoder network, low-resolution data is input It is then converted into a feature map

[0093] The feature map is obtained by linear interpolation Among them, k i is the upsampling scale of the i∈[1,M]th image, M is the number of training samples;

[0094] The eigenvector is obtained by dimensionality reduction and high-resolution images The pixel coordinates X = (x, y, z) are concatenated and input into the decoding network to obtain the predicted pixel intensity. and compared with the pixel intensity in the true label Perform loss calculation; the loss function is defined as follows:

[0095]

[0096] Where D is the number of input images, represents the coordinate sampling in each high-resolution gold standard image, is the predicted pixel intensity, is the true pixel intensity in the high-resolution gold standard data. After being processed by the super-resolution network, the resolution of the deformation field becomes N times the original.

[0097] In step S3, the floating image and the upsampled deformation field are input into a spatial transformation network (STN) to generate a registered image. The spatial transformation network is implemented using linear interpolation. The spatial transformation network is defined as follows:

[0098]

[0099] Among them, I m and I w Represent floating images and distorted images respectively, and x represents I m , y represents the voxel in the 8-pixel cube neighborhood x+φ(x), and d∈{i,j,k} represents the three dimensions of the image. represents the neighborhood of pixel x, p represents the pixel in the image, and φ represents the registered deformation field.

[0100] The objective function designed in step S4 is defined as follows:

[0101]

[0102] in, It is composed of mean square error and is used to represent the fixed image I f and distorted images The similarity between m represents the floating image, φ represents the registered deformation field, represents the smoothness loss of the deformation field, represents the weak supervision loss, λ represents the hyperparameter of the deformation field smoothing loss, and μ represents the hyperparameter of the weak supervision loss;

[0103] Similarity loss consisting of mean square error The definition is as follows:

[0104]

[0105] Among them, I f Represents a fixed image, represents the distorted image, Ω represents the n-dimensional feature space, and p represents the pixel in the image;

[0106] The weak supervision loss is defined as follows:

[0107]

[0108] in, They represent the segmentation labels of the fixed image and the segmentation labels of the deformed floating image, respectively. K represents the total number of regions. Represents the lth sub-region of the floating segmentation label;

[0109] Deformation field smoothing loss The definition is as follows:

[0110]

[0111]

[0112] in, represents the image volume neighborhood, represents the first-order derivative of the deformation field φ(·), represents the second-order derivative of the deformation field (·).

[0113] In step S5, the designed sparse attention-based MRI registration model is used for training, and the Adam optimizer is used to drive the objective function of step S4 for optimization; the optimal training model is obtained by a five-fold cross validation method, and the MRI registration model is iteratively trained for 500 cycles, the initial learning rate is set to 1e-4, and the dynamic learning rate is defined as follows:

[0114]

[0115] Among them, power = 0.9 is the decay exponent, epoch is the number of cycles of the current iteration, and max_epoch = 500 is the iteration period.

[0116] The above-mentioned arbitrary-size MRI image registration method also includes step S7 of evaluating the registration result of step S6 using evaluation indicators; testing the trained MRI image registration model, inputting the test data and the optimal weight parameters obtained through training into the MRI image registration model, and using multiple evaluation indicators to evaluate the registration performance of the method.

[0117] The accuracy of the registration result image is evaluated using the coincidence coefficient indicator. The coincidence coefficient ∈ [0, 1] is used to evaluate the accuracy of the registration result. The closer the coincidence coefficient value is to 1, the more similar the fixed image and the distorted image are. The coincidence coefficient DSC is defined as follows:

[0118]

[0119] Among them, I f and represent the fixed image and the distorted image respectively.

[0120] Any of the above implementations is based on a lightweight design of sparse attention, combined with multiple advanced technologies such as deformation field super-resolution upsampling technology, feature sparse representation, attention mechanism and multi-layer perceptron, aiming to solve the problems faced by traditional registration methods when processing magnetic resonance imaging and medical images, such as high computational complexity, poor real-time performance, and inability to cope with input data of arbitrary size, and provide more accurate and efficient tools for clinical diagnosis and treatment.

[0121] The present application has been disclosed as above with preferred implementation cases, but it is not intended to limit the present application. Any technician familiar with the profession can make slight changes or modifications to the above-disclosed structures and technical contents without departing from the scope of the technical solution of the present application, and the equivalent implementation cases with equivalent changes are still within the scope of the technical solution of the present application.

Claims

1. A method for registering nuclear magnetic resonance images of any size, characterized by: The following steps are included: S1, scan brain magnetic resonance imaging data set, preprocess the data and divide it into training set, validation set and test set; S2. Construct a sparse attention-based MRI image registration model, treat paired images as floating images and fixed images respectively, perform downsampling on the input images, and input the processed images into the designed MRI image registration model for training; S3, performing super-resolution upsampling processing on the deformation field representing the difference between images obtained through training, and inputting the floating image and the upsampled deformation field into the spatial transformation network to generate the registered image; S4, using the designed objective function to calculate the difference between the fixed image and the registered image, and performing back propagation optimization on the nuclear magnetic resonance image registration model; S5, performing iterative optimization training on the nuclear magnetic resonance image registration model until the model training converges, terminating the iterative training, and then saving the optimal weight parameters; S6. Input the test data into the MRI image registration model, and load the optimal weight parameters obtained in the training phase to obtain the registration result.

2. The method for registering nuclear magnetic resonance images of any size according to claim 1, characterized in that: In step S1, the nuclear magnetic resonance image needs to be preprocessed, including normalization and affine transformation; wherein the normalization process is defined as follows: Among them, x represents the pixel of the image, and after mapping, the characteristic value of the pixel is mapped to Afterwards, the floating image is coarsely registered to the fixed image using an affine transformation.

3. The method for registering nuclear magnetic resonance images of any size according to claim 1, characterized in that: The design process of the MRI registration model in step S3 is as follows: S31. Construct a feature extraction network, perform four-fold downsampling using a sparse Transformer structure, and implement network encoding of the input data. The formula of the designed sparse Transformer structure is defined as follows: SHMA(X l-1 )=Concat(US1,…,US k )W o Among them, SHMA Hi represents the sparse multi-head attention mechanism, Hi∈[1,k] represents the i-th head, MLP represents the multi-layer perceptron, LN represents the layer normalization, and X l-1 and X l Represent the input and output of sparse attention respectively, and the mapping matrix W o Used to project the self-attention result to the target output dimension; The structural formula of the sparse multi-head attention mechanism is defined as follows: Q, K, and V represent the feature maps after mapping, respectively. i ,k i ,v i denotes the i-th row of Q, K, and V respectively. The probability form of the attention of the i-th query to all keys is defined as Attn(q i ,k j )=p(k j |q i ),q(k j |q i ) represents a uniform distribution, when k(q i ,k j ) / ∑ l k(q i ,k l )=1 / L K , the similarity is measured by Kullback-Leibler divergence, which is defined as follows: Remove the constant term ln L k After that, the sparsity of the i-th query is defined as M(q i ,K), is M(q i ,K) is a simplified form, a sparse matrix and q have the same size, SHMA Hi (Q,K,V)=Attn(Q,K)V SHMA(X l-1 )=Concat(US1,…,US k )W o Among them, Attn represents sparse attention processing, represents the scale factor, L k represents the sequence length, d k represents the sequence dimension; S32, using deconvolution upsampling to gradually restore the size of the feature map to achieve network decoding of the output data; the corresponding layers of the network encoding and decoding are connected using a skip connection structure; S33, output a deformation field representing the difference between the images.

4. A method for registering nuclear magnetic resonance images of any size according to claim 1 or 3, characterized in that: In step S2, the input image is subjected to nearest neighbor downsampling, and the input image size is reduced to the original size using nearest neighbor downsampling. N represents the factor by which the image size is reduced.

5. According to claim 1, a method for registering nuclear magnetic resonance images of any size, characterized in that: In step S3, the deformation field is processed using a super-resolution network structure to increase the image resolution by N times, so that the network can process nuclear magnetic resonance images of any size; The input of the super-resolution network structure includes low-resolution data and high-resolution gold standard data In the encoder network, low-resolution data is input It is then converted into a feature map The feature map is obtained by linear interpolation Among them, k i is the upsampling scale of the i∈[1,M]th image, M is the number of training samples; The eigenvector is obtained by dimensionality reduction and high-resolution images The pixel coordinates X = (x, y, z) are concatenated and input into the decoding network to obtain the predicted pixel intensity. and compared with the pixel intensity in the true label Perform loss calculation; the loss function is defined as follows: Where D is the number of input images, represents the coordinate sampling in each high-resolution gold standard image, is the predicted pixel intensity, is the true pixel intensity in the high-resolution gold standard data.

6. The method for registering nuclear magnetic resonance images of any size according to claim 1, characterized in that: In step S3, the spatial transformation network is implemented using linear interpolation, and the spatial transformation network is defined as follows: Among them, I m and I w Represent floating images and distorted images respectively, and x represents I m , y represents the voxel in the 8-pixel cube neighborhood x+φ(x), and d∈{i,j,k} represents the three dimensions of the image. represents the neighborhood of pixel x, p represents the pixel in the image, and φ represents the registered deformation field.

7. The method for registering MRI images of any size according to claim 1, characterized in that: The objective function designed in step S4 is defined as follows: in, It is composed of mean square error and is used to represent the fixed image I f and distorted images The similarity between m represents the floating image, φ represents the registered deformation field, represents the smoothness loss of the deformation field, represents the weak supervision loss, λ represents the hyperparameter of the deformation field smoothing loss, and μ represents the hyperparameter of the weak supervision loss; Similarity loss consisting of mean square error The definition is as follows: Among them, I f Represents a fixed image, represents the distorted image, Ω represents the n-dimensional feature space, and p represents the pixel in the image; The weak supervision loss is defined as follows: in, They represent the segmentation labels of the fixed image and the segmentation labels of the deformed floating image, respectively. K represents the total number of regions. represents the kth sub-region of the floating segmentation label; Deformation field smoothing loss The definition is as follows: in, represents the image volume neighborhood, represents the first-order derivative of the deformation field φ(·), represents the second-order derivative of the deformation field (·).

8. The method for registering nuclear magnetic resonance images of any size according to claim 1, characterized in that: In step S5, the nuclear magnetic resonance image registration model is trained, and the objective function of step S4 is optimized by using the Adam optimizer; the optimal training model is obtained by a five-fold cross validation method, and the nuclear magnetic resonance image registration model is iteratively trained for several cycles, and the initial learning rate is set to 1e-4. The dynamic learning rate is defined as follows: Among them, power = 0.9 is the decay exponent, epoch is the number of cycles of the current iteration, and max_epoch is the iteration period.

9. The method for registering MRI images of any size according to claim 1, characterized in that: It also includes step S7 of evaluating the registration result of step S6 using an evaluation index; The accuracy of the registration result image is evaluated using the coincidence coefficient indicator. The coincidence coefficient ∈ [0, 1] is used to evaluate the accuracy of the registration result. The closer the coincidence coefficient value is to 1, the more similar the fixed image and the distorted image are. The coincidence coefficient DSC is defined as follows: Among them, I f and represent the fixed image and the distorted image respectively.

Images