Method and system for synthesizing CT images into MR images based on similar reference cases

Abstract

The invention belongs to the technical field of image processing, and provides a method and system for synthesizing a CT image into an MR image based on similar reference cases in order to solve the problem that the accuracy of synthesizing the CT image into the MR image is poor. The method for synthesizing the MR image based on the CT image of the similar reference case comprises the steps of obtaining a 2D CT slice image, high-frequency information corresponding to the 2D CT slice image and a 3D MRI image of the similar reference case; processing the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of the similar reference case by using a generative adversarial neural network model to obtain a corresponding MR image; wherein the generative adversarial neural network model comprises a feature encoder, a resampling and multi-layer perceptron module, a mask Transform module and a feature decoder, more accurate feature matching and image synthesis are realized, and image distortion is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and system for synthesizing MR images based on CT images of similar reference cases. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Empirical medicine shows that after reading medical images of similar conditions, radiologists can calibrate their judgments on the same type of condition, and can transfer image information of similar conditions to the medical images of the current patient for condition comparison during the reading process. If the generative adversarial neural network is compared to the doctor's brain, the empirical medicine guides the introduction of data of similar conditions into the generative adversarial neural network, which can better realize the conversion of CT data to MR data. However, the problem faced by introducing similar reference cases is that the body position of the current patient and the reference patient is moved when the images are taken. At the same time, although the conditions of the two patients are similar, individual differences will still lead to differences in medical images, which will affect the accuracy of the synthesis of MR images from CT images and reduce their actual reference value. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method and system for synthesizing MR images based on CT images of similar reference cases, which can achieve more accurate feature matching and image synthesis.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A first aspect of the present invention provides a method for synthesizing an MR image based on a CT image of a similar reference case.

[0006] In one or more embodiments, a method for synthesizing an MR image based on a CT image of a similar reference case is provided, comprising: Obtain 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; Using a generative adversarial neural network model, the 2D CT slice image, high-frequency information corresponding to the 2D CT slice image, and a 3D MRI image of a similar reference case are processed to obtain a corresponding MR image; The generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

[0007] A second aspect of the present invention provides a system for synthesizing MR images based on CT images of similar reference cases.

[0008] In one or more embodiments, a system for synthesizing an MR image based on a CT image of a similar reference case includes: An information acquisition unit, which is used to acquire 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; An image generation unit, which is used to process the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image, and the 3D MRI image of a similar reference case by using a generative adversarial neural network model to obtain a corresponding MR image; The generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

[0009] A third aspect of the present invention provides a computer-readable storage medium.

[0010] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the method for synthesizing MR images based on CT images of similar reference cases as described above.

[0011] A fourth aspect of the present invention provides a computer program product.

[0012] A computer program product includes a memory and a processor, wherein the memory stores a computer program, and when the program is executed by the processor, the steps in the method for synthesizing MR images based on CT images of similar reference cases as described above are implemented.

[0013] A fifth aspect of the present invention provides an electronic device.

[0014] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for synthesizing MR images based on CT images of similar reference cases as described above are implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention adopts a generative adversarial neural network (GAN) model and combines the high-frequency information of 3D MRI images and CT images of similar reference cases to solve the problem of synthesizing high-quality MR images from CT images, generating a synthesis effect that is closer to the real MR image, and improving the reliability of medical image analysis.

[0016] (2) The present invention solves the problem of insufficient feature alignment and fusion of CT and MRI images through the multi-module collaboration of feature encoder, resampling and multi-layer perceptron module, mask Transformer module and feature decoder, achieves more accurate feature matching and image synthesis, and avoids image distortion.

[0017] (3) The present invention uses the mask Transformer module to perform a secondary fusion of the primary fusion features and the CT high-frequency information, which solves the problem of insufficient fusion of high-frequency information and global features, generates richer details and natural image transition effects, and improves the quality of the synthesized MR images.

[0018] (4) The present invention introduces 3D MRI images of similar reference cases, which solves the problem of insufficient information in a single CT image and generates synthetic MR images that are more consistent with anatomical structures and pathological characteristics, providing more comprehensive imaging support for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 is a flow chart of a method for synthesizing MR images based on CT images of similar reference cases according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a generative adversarial neural network model according to an embodiment of the present invention; Figure 3 is the characteristic attention effect of the embodiment of the present invention; Figure 4 is a schematic diagram of resampling and a multi-layer perceptron according to an embodiment of the present invention; Figure 5 is a schematic diagram of a mask Transformer module of an embodiment of the present invention; Figure 6 is a gate control module according to an embodiment of the present invention; Figure 7 is a schematic diagram of feature supervision based on contrastive learning in an embodiment of the present invention; Figure 8 The embodiment of the present invention is a spinal sagittal and transverse CT, MR and generation of MR slices; Fig. 9 Schematic diagram of the system structure for synthesizing MR images based on CT images of similar reference cases according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0024] like Figure 1 As shown, the method for synthesizing an MR image based on a CT image of a similar reference case in this embodiment may include: Obtain 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; Using a generative adversarial neural network model, the 2D CT slice image, high-frequency information corresponding to the 2D CT slice image, and a 3D MRI image of a similar reference case are processed to obtain a corresponding MR image; Among them, Figure 2 As shown, the generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask Transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

[0025] In this embodiment, the input data of the feature encoder includes 2D CT slices, high-frequency information corresponding to the 2D CT slices, and 3D MRI of similar reference cases. Since the features corresponding to the high-frequency information corresponding to the 2D CT slices have a complementary effect on the features of the 2D CT slices, such as Figure 3 As shown, the feature encoder consists of three parts, as shown in Table 1. The parameters in the parameter Downsample in the table are the convolution kernel size, the number of convolution kernels, and the step size. The 2D CT slice feature is obtained through the above encoder , 3DMRI features , 2D CT slice high frequency features .

[0026] Table 1 Input data of feature encoder

[0027] The 2D CT slice features and 3D MRI features are input into the resampling and multi-layer perceptron module. It is hoped that after passing through the module, the useful features of the 2D CT slice features and 3D MRI features are retained, while the redundant features are removed. The above process can be described as the 2D CT slice features and 3D MRI features are: , whose probability is calculated by the sigmoid function , from which we obtain k A useful feature is represented by In practice, we take the logarithm of p(f) and add random Gumbel noise to construct the x sample: ; Where G is an independent and identically distributed random variable with full Gumbel distribution Its probability density function is ; Where m is the value of the random variable, μ is the location parameter, β is the scale parameter (determines the degree of dispersion of the distribution), and z is a standardized variable used to simplify the expression; ; The inverse function is: ; Where n is the probability value; Subsequent use =0, = 1 can be obtained by sampling from a uniform distribution, that is: ; in is a uniformly distributed sample value, is the sample value of Gumbel distribution; The probability distribution of T is: ; Where t is the threshold, x is an element in the feature set f, lnp(x) is the logarithmic probability of feature x, and G is a random variable with Gumbel distribution; This can be written as a product operation: ; When G =0, =1, the above expression becomes: ; and correspond The distribution of is expressed as , the above expression becomes ; As long as from By collecting enough samples, we can restore distribution, and thus obtain Therefore, from Filter The problem changes from the probability perspective to solving the problem of maximum probability of k positions, let , is the maximum probability value, we can get: ; The probability that the kth position feature is the largest is that for all Integrate and get ; in )and ) represent the probability values ​​of the kth and i-th position features respectively, and K is the total number of features.

[0028] All positions are probabilistically calculated as described above, and then onehot encoded to form a mask, which is Multiply them together to get Then the dimension is adjusted through the multi-layer perceptron to obtain the final resampled features ,like Figure 4 shown.

[0029] In order to take into account the complementary effect of high-frequency feature information on feature attention, and resample the Gumbel-Softmax According to the structure corresponding to the high-frequency information of 2D CT slices, the present invention designs a mask Transformer module to and The overall structure is shown in the figure below. Figure 5 shown.

[0030] right and Perform position encoding and obtain the encoding vector through the following operations , , : ; in and are the input features, , and is the learnable weight matrix.

[0031] and The initial correspondence is achieved through the cosine attention mechanism: ; in and yes and The average value of yes and The initial matching value reflects and Matching effect, if If there are redundant features in , it will lead to a low matching score, which will be eliminated after the subsequent activation function operation. This will involve unreliable matching scores. To overcome this limitation, we determine whether the initial matching scores are reliable or not. For reliable matches, we calculate them as follows: ; in Set to 100 to control the sharpness of the softmax function. is the reliable matching score after filtering by the ReLU activation function, The attention weights normalized by the softmax function, V is the value matrix, A feature vector representing the matched part.

[0032] For the calculation of unreliable matching, the original image with high-frequency information is first convolved to obtain the modulation parameters and , corresponding to the scale factor tensor and the bias factor tensor, through these two tensors Modulate and get ; in and is the mean and variance in the second dimension, in order to To calculate the unreliable matching, we also sum the second dimension and multiply the corresponding elements between tensors through the broadcast mechanism to obtain : ; and It is learned from the high-frequency information of 2D CT slices, so The modulated features correspond to its semantic information, which is consistent with Multiplying them together gives the portion that is not a reliable match. The feature vector representing the mismatched part.

[0033] In theory, the matching part and the non-matching part should be independent of each other, and a position normalization method is used to normalize the , , Perform position normalization.

[0034] ; in, is the position normalization function; is the feature vector obtained after position normalization.

[0035] Different from the original Transformer structure, this paper does not use FFN, but uses CBAM attention mechanism and layer normalization ( ), Gaussian error linear unit ( )、2d convolution( ),right Processing is performed to obtain the final fusion result: ; The fusion feature output after Transformer is expressed as .

[0036] To prevent Excessive forgetting of features occurs. In the decoder, each upsampling is done by and To supplement the features, and The distribution of the MR slices to be generated is inconsistent, so an adaptive instance normalization-gating module is designed to supplement the features each time the image is upsampled. Adaptive instance normalization is often used in image style transfer. The distribution of is consistent with the final MR slice global feature distribution (this operation is achieved by adding contrastive learning supervision). layer, Upsampling of layer features , Through a fully connected layer , through the adaptive instance normalization operation, The mean and variance of The mean and variance of are aligned as follows: ; Among them and Indicates the operation of taking the mean and variance. Adaptive instance normalization operation result , and are sent to the gating module for feature supplementation, where the gating module is as follows: Figure 6 Shown for and , perform the following calculation: ; for and , perform the following calculation: ; in , , , , , is the training parameter matrix, The sigmoid function acts as a gate, generating values ​​between 0 and 1, and tanh processes the input features. By stacking , , get.

[0037] ; in, It is the stacking operation function.

[0038] According to the above operation, the final decoded MR slice image is obtained by four upsamplings. .

[0039] The present invention selects features through Gumbel-Softmax resampling, and then performs secondary screening on the features through mask Transformer. The obtained features need to be effectively supervised to ensure the quality of the generated results. Here we introduce the contrastive learning method as a supervision method. The overall structure diagram is shown in Figure 7 shown.

[0040] The current patient CT data, the current patient CT high-frequency information, and the similar reference case 3D MRI are sent as a group to the encoder, Gumbel-Softmax resampling and multi-layer perceptron module, and the mask Transformer module to obtain the features as positive samples represented as follows: The current patient CT data, the current patient CT high-frequency information, and the current patient 3D MRI are fed as a set to the encoder, Gumbel-Softmax resampling and multi-layer perceptron module, and the features obtained by the mask Transformer module are used as anchor points and are expressed as The current patient CT data, the current patient CT high-frequency information, and the non-similar reference case 3DMRI are fed as a set to the encoder, Gumbel-Softmax resampling and multi-layer perceptron module, and the mask Transformer module to obtain the features as negative samples represented as .

[0041] Generator loss function: The total loss function of the generator includes contrast loss, perceptual loss, context loss and structural loss, and its formula is as follows: ; The contrast loss is: ; in, The value is 0.07. It is set to 128, that is, each positive sample is equipped with 128 negative samples for contrastive learning.

[0042] The perceived loss is estimated by and real It is sent to a pre-trained VGG19 for feature extraction, and then the high-level features of the two are compared. The formula is as follows: ; in Represents operations for extracting high-level features.

[0043] Context loss is used to ensure the consistency of local texture features, and its formula is as follows ; in is a coefficient used to balance the continuity effect of multi-scale hierarchical features.

[0044] The structural loss is calculated using the Learning Perceptual Patch Similarity (LPIPS) method, and the results are as follows: ; in, is the total loss function for the generative adversarial neural network model; is the contrast loss; For perceived loss; is the context loss; For structural losses; and are the estimated MR image and the real MR image; Represents the operation of extracting high-level features; is the coefficient; LPIPS is the learning perceptual image block similarity function; It represents negative samples; For anchor point; is the positive sample representation; is a constant coefficient; represents the number of negative samples; To select multi-level feature functions.

[0045] The discriminator and discriminator loss functions used in this example are consistent with the discriminator and discriminator loss functions of pix2pix. Figure 8 As shown, the leftmost side is a CT image, the middle side is a corresponding MR image, and the edge side is an MR image generated by the network of an embodiment of the present invention.

[0046] The present invention establishes for the first time a spine data set of CT images and MR images with similar conditions, which includes the patient's registered C and MRI 2D slices, and 3DMRI of similar patients, with a total of XX groups.

[0047] This paper proposes a new generative adversarial neural network, which can transfer information from 3D MRI of similar patients based on the structural features of 2D CT slices, and designs three modules to achieve stable synthesis of MR images from spinal CT images: The Gumbel-Softmax resampling and multi-layer perceptron modules designed in this paper realize the initial screening of 3D MRI features and 2D CT slice features of similar cases, identify redundant features, and realize the alignment of 2D features and 3D features. The features output by the Gumbel-Softmax resampling and multi-layer perceptron modules of the present invention and the high-frequency information features of 2D CT slices are secondary screened and aligned through a Transformer with a masking effect to achieve accurate multimodal feature fusion.

[0048] In the process of the decoder, the present invention puts the 2D CT slice features, the 2D CT slice high-frequency information features and the features output by the Transformer into a gated normalization module to achieve feature supplementation and prevent excessive feature forgetting.

[0049] In terms of network supervision, the present invention designs a loss function that can optimize multiple factors such as structure, texture, and details. In particular, a comparative learning method that maximizes the difference between similar reference cases and non-similar reference cases is added to improve the quality of generated MR slices. In one or more embodiments, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for synthesizing MR images based on CT images of similar reference cases as described above are implemented.

[0050] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer readable medium, the computer program including a computer program for executing Figure 1In such an embodiment, the computer program can be downloaded and installed from a network through the communication part, and / or installed from a removable medium. When the computer program is executed by the central processing unit 401, various functions defined in the device of the present application are executed.

[0051] in, Figure 1 The computer program instructions corresponding to the method shown may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0052] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0053] Fig. 9 is a schematic diagram of a system structure for synthesizing MR images based on CT images of similar reference cases in an embodiment of the present invention. Figure 1 The method of synthesizing MR images based on CT images of similar reference cases corresponds to Fig. 9 As shown, the system for synthesizing MR images based on CT images of similar reference cases in this embodiment may include: An information acquisition unit 901 is used to acquire 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; An image generation unit 902 is used to process the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image, and the 3D MRI image of a similar reference case using a generative adversarial neural network model to obtain a corresponding MR image; The generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

[0054] It should be noted here that Fig. 9 The various units in the system for synthesizing MR images based on CT images of similar reference cases are Figure 1 The various steps in the method of synthesizing MR images based on CT images of similar reference cases correspond one to one, and the specific implementation process is the same, which will not be repeated here.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for synthesizing MR images based on CT images of similar reference cases, characterized in that: include: Obtain 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; Using a generative adversarial neural network model, the 2D CT slice image, high-frequency information corresponding to the 2D CT slice image, and a 3D MRI image of a similar reference case are processed to obtain a corresponding MR image; The generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment and fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

2. The method for synthesizing MR images based on CT images of similar reference cases according to claim 1, characterized in that: The feature encoder includes a 3D encoder and two 2D encoders.

3. The method for synthesizing MR images based on CT images of similar reference cases according to claim 1, characterized in that: The resampling and multi-layer perceptron module first stacks the 3D MRI features and the 2D CT slice features to obtain stacked features, and then performs a primary screening on the stacked features through Gumbel distribution.

4. The method for synthesizing MR images based on CT images of similar reference cases as claimed in claim 1, characterized in that: In the feature decoder, an adaptive instance normalization operation is used to align the mean and variance of the upper layer upsampled from the initial fusion feature with the feature mean and variance obtained from the initial fusion feature through the fully connected layer.

5. The method for synthesizing MR images based on CT images of similar reference cases according to claim 1, characterized in that: The total loss function of the generative adversarial neural network model contains contrast loss, perceptual loss, contextual loss and structural loss.

6. The method for synthesizing MR images based on CT images of similar reference cases according to claim 1, characterized in that: The total loss function expression of the generative adversarial neural network model is: ； ； ； ； ； in, is the total loss function for the generative adversarial neural network model; is the contrast loss; For perceived loss; is the context loss; For structural losses; and are the estimated MR image and the real MR image; Represents the operation of extracting high-level features; is the coefficient; LPIPS is the learning perceptual image block similarity function; It represents negative samples; For anchor point; is the positive sample representation; is a constant coefficient; represents the number of negative samples; To select multi-level feature functions.

7. A system for synthesizing MR images based on CT images of similar reference cases, characterized in that: include: An information acquisition unit, which is used to acquire 2D CT slice images, high-frequency information corresponding to the 2D CT slice images, and 3D MRI images of similar reference cases; An image generation unit, which is used to process the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image, and the 3D MRI image of a similar reference case by using a generative adversarial neural network model to obtain a corresponding MR image; The generative adversarial neural network model includes a feature encoder, a resampling and multi-layer perceptron module, a mask transformer module and a feature decoder; The feature encoder is used to extract 2D CT slice features, 2D CT slice high-frequency features and 3D MRI features from the 2D CT slice image, the high-frequency information corresponding to the 2D CT slice image and the 3D MRI image of a similar reference case respectively; The resampling and multi-layer perceptron module is used to perform a primary screening of 2D CT slice features and 3D MRI features, and then perform alignment fusion to obtain primary fusion features; The mask transformer module is used to match and fuse the primary fusion features with the 2D CT slice high-frequency features to obtain secondary fusion features; The feature decoder is used to decode both the secondary fusion feature and the primary fusion feature to obtain a corresponding MR image.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for synthesizing an MR image based on a CT image of a similar reference case as described in any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the program is executed by a processor, the steps in the method for synthesizing an MR image based on a CT image of a similar reference case as described in any one of claims 1 to 6 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for synthesizing an MR image based on a CT image of a similar reference case as described in any one of claims 1 to 6 are implemented.

Images