A method and system for synthesizing CT images from MRI images based on deep learning

Abstract

The application discloses a kind of based on deep learning's MRI image synthesis CT image method and system, method includes: S1.MRI image data and CT image data are acquired and preprocessed;S2.K-means algorithm is used to mark the bone position of preprocessed CT image, and the corresponding binary mask image is obtained;S3.the binary mask image is extracted using to obtain the bone, non-skeletal part image of CT image;S4.MRI is established based on GAN's depth learning network model of CT conversion;S5.the loss function of the network model is constructed, the loss function includes generation confrontation loss L adv , pixel level L1 loss L pix , binary cross-entropy loss L BCE ;S6.preprocessed MRI image data is input into the depth learning network model based on MRI of GAN CT conversion, and output CT image is output.The application fully utilizes the pixel distribution difference of different parts of CT image to synthesize CT image, and improves the synthesis quality of CT image.

Description

Technical Field

[0001] The present invention relates to the field of multimodal medical imaging technology, and more specifically, to a method and system for synthesizing CT images from MRI images based on deep learning. Background Art

[0002] Multimodal imaging can ideally provide a comprehensive examination of the target tissue, including the precise location and metabolic activity of the target tissue, tissue flow and functional changes within the surrounding tissue, and characteristic changes that lead to the final disease. Computed tomography (CT) imaging is crucial for various applications, such as radiotherapy planning and positron emission tomography (PET) attenuation correction. This is because CT images provide an accurate representation of the patient's geometry, and more importantly, the CT value in Hounsfield units (HU), which measures the tissue attenuation coefficient, can be directly converted to electron density for radiation dose calculation. However, CT images have limited soft tissue contrast, and the process also exposes the patient to additional radiation. Recently, magnetic resonance imaging (MRI) has developed rapidly due to the lack of ionizing radiation, excellent soft tissue contrast, and the ability to perform multi-parameter imaging through various MRI image sequences, and people have become interested in using MRI to replace CT.

[0003] There is a large amount of research on medical image synthesis in the literature, ranging from traditional physical model-based methods [E.Hodneland, Physical models for simulation and reconstruction of human tissue deformation fields in dynamic MRI, 2016] to learning-based data-driven methods [M.Chen, Cross contrast multi-channel image registration using image synthesis for MR brain images, 2017]. Recent research mainly uses generative adversarial networks (GANs), which use adversarial losses to increase the capture of detailed tissue structures [SUDar, Image synthesis in multi-contrast MRI with conditional generative adversarial networks, 2019]. However, existing deep learning-based MRI-CT synthesis methods generally synthesize the entire part of the CT image at the same time, without fully utilizing the pixel differences between bones and other parts of the CT image, and their synthesis is often poor at the junction of bones and tissues. Summary of the Invention

[0004] In order to overcome the defects of the above-mentioned prior art that CT image synthesis does not fully utilize the pixel distribution differences in different parts of the CT image and the image synthesis quality is poor, the present invention provides a method and system for synthesizing CT images from MRI images based on deep learning.

[0005] A first aspect of the present invention provides a method for synthesizing CT images from MRI images based on deep learning, characterized by comprising the following steps:

[0006] S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and CT image, registration of the MRI image and CT image, and slice processing;

[0007] S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image;

[0008] S3. Extract the bone and non-bone portions of the CT image using the binary mask image.

[0009] S4. Establish a deep learning network model for MRI-CT conversion based on GAN;

[0010] S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ;

[0011] S6. The preprocessed MRI image data is input into a GAN-based MRI-CT deep learning network model to output a CT image.

[0012] In this solution, the pre-processing process described in step S1 is:

[0013] First, the N4 bias field correction is performed on the collected MRI image data that displays unevenness. Then, the corrected MRI image data and the collected CT image data are denoised to prevent the noise generated during the scanning process from affecting subsequent work. The denoised MRI image data and CT image data belonging to the same patient, the same period, and the same part are rigidly aligned to ensure that the size of the MRI image data and the CT image data are consistent. Finally, slicing is performed to slice the 3D registered image data according to a preset ratio, and slices with less information than the set value are removed to obtain 2D image data sets of MRI and CT.

[0014] In this solution, the processing flow described in step S2 is:

[0015] First, a K-means clustering model was constructed, using the pixel values ​​of the CT image as feature values. The distance between each feature was calculated using the L1 distance calculation method. The model parameters were then initialized, with the value of K set to 3 and the features of the centers of the initial clusters of each type initialized to 0, 127, and 255. The algorithm was iterated until the classification of each pixel no longer changed.

[0016] The K-mean clustering model marks the CT image into black background, gray tissue, and white bone classes. After clustering, the class with the largest eigenvalue is taken as the binary mask image m of the bone in the CT image, where the value 1 indicates that the CT pixel at that position belongs to the bone part, and 0 indicates other parts.

[0017] In this solution, the processing flow described in step S3 is:

[0018] The CT image is converted into matrix data with a numerical range of 0 to 255, and the partial data at the same position in the CT image as the binary mask image m described in S2 with a value of 1 is extracted as the CT bone image b. This process can be expressed as: b = m × y, where y represents the CT image data after S1 preprocessing; the rest of the CT image is extracted as the CT tissue image t. This process can be expressed as: t = yb.

[0019] In this scheme, the deep learning network model of MRI to CT conversion based on GAN includes a generator network G and a discriminator network D, wherein the generator network G includes a CT bone segmentation sub-model Bone_segm, a bone synthesis sub-model Bone_syn, a tissue synthesis sub-model Tissue_syn, and an integration model Integ;

[0020] The skeleton segmentation sub-model model function expression is:

[0021]

[0022] Among them, x i ∈R N is the MRI image preprocessed by S1, the MRI image may be a T1 type MRI image or a T2 type MRI image, and the MRI image is the input of the segmentation model, is the bone segmentation image of the MRI image. The bone segmentation sub-model segments the MRI image and obtains the corresponding bone mask, which serves as the prior information for other networks. Bone_segm(·) represents the deep learning network model that maps the MRI image to the bone mask.

[0023] In this solution, the skeleton synthesis sub-model function expression is:

[0024]

[0025] Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Cat[·] represents the splicing operation in the channel dimension, is the CT bone image synthesized by the model, and Bone_syn(·) represents the deep learning network model that maps MRI images to CT bones.

[0026] In this solution, the functional expression of the tissue synthesis sub-model described in step S4 is:

[0027]

[0028] Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Indicates the mask inversion operation, that is Cat[·] represents the concatenation operation in the channel dimension, is the CT tissue image synthesized by the model, and Tissue_syn(·) represents the deep learning network model that maps MRI images to CT tissue.

[0029] In this solution, the integrated model function expression is:

[0030]

[0031] Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Synthesize bone images for the bone synthesis model, Tissue images synthesized by the tissue synthesis model, is the final synthesized CT image, and Integ(·) represents the model that integrates the synthesized bone image and tissue image into the final CT image.

[0032] The generator G expression is:

[0033] G=Integ(Bone_syn(·),Tissue_syn(·),Bone_segm(·))

[0034] Among them, Bone_syn is the bone synthesis sub-model, Tissue_syn is the tissue synthesis sub-model, and Bone_segm is the bone segmentation sub-model.

[0035] In this solution, the generation adversarial loss L described in step S5 is adv Expressed as: L adv =E x [log(1-D(x,G(x)))], where x represents the MRI image input image, E x It is obtained from the data distribution in the dataset;

[0036] The pixel level L1 loss L pix Expressed as:

[0037]

[0038] Among them, b i is the CT bone image extracted in step S3, is the skeleton image synthesized by the skeleton synthesis network, t i is the CT tissue image extracted in step S3, is the tissue image synthesized by the tissue synthesis network, y i is the CT image obtained by preprocessing in step S2, is the final synthetic CT image synthesized by the generator G, θ1, θ2 and θ3 are L pix_b , L pix_t and L pix_y The corresponding weight coefficient.

[0039] The binary cross entropy loss L BCE Expressed as:

[0040]

[0041] Among them, m i The binary mask data of the CT bones extracted in step S4, Mask image segmented by the skeleton segmentation network.

[0042] A second aspect of the present invention provides a system for synthesizing CT images from MRI images based on deep learning, comprising a memory and a processor. The memory includes a program for synthesizing CT images from MRI images based on deep learning. When the program is executed by the processor, the following steps are implemented:

[0043] S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and CT image, registration of the MRI image and CT image, and slice processing;

[0044] S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image;

[0045] S3. Extract the bone and non-bone portions of the CT image using the binary mask image.

[0046] S4. Establish a deep learning network model for MRI image conversion to CT based on GAN;

[0047] S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ;

[0048] S6. The preprocessed MRI image data is input into a GAN-based deep learning network model for MRI image to CT conversion, and a CT image is output.

[0049] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0050] The present invention proposes a method and system for synthesizing CT images from MRI images based on deep learning. Based on the fitting ability of deep learning, a deep learning neural network model including different sub-models is constructed to learn the distribution of bones and tissues respectively, and then fuse and integrate them. At the same time, the characteristics of CT images are utilized and the K-means clustering method is adopted for adaptive segmentation. In the original MR-CT data set, it is easy to obtain the required bone mask data in batches. After the bones and tissues are separated, their distribution information is clearer, which simplifies the synthesis task of each synthesis sub-network. At the same time, different sub-models can be selected for synthesis according to the characteristics of different parts. A more targeted network can greatly reduce the complexity of the network. Compared with the existing CT image synthesis method, the quality of the synthesized image is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method for synthesizing CT images from MRI images based on deep learning in the present invention.

[0052] Figure 2 This is a flow chart of the preprocessing of MRI image data and CT image data of the present invention.

[0053] Figure 3 This is a flow chart for obtaining a binary mask image in the present invention.

[0054] Figure 4 Schematic diagram of the deep learning network model framework for MRI image conversion to CT based on GAN in the present invention.

[0055] Figure 5 This is a system block diagram of the present invention for synthesizing CT images from MRI images based on deep learning. DETAILED DESCRIPTION

[0056] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0057] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0058] Example

[0059] The present invention proposes a method for synthesizing CT images from MRI images based on deep learning, and its implementation steps are as follows: Figure 1 As shown, the detailed steps are as follows:

[0060] S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and the CT image, registration of the MRI image and the CT image, and slice processing;

[0061] It's important to note that raw MRI and CT image data often contain noise. Furthermore, due to varying scanning mechanisms and parameter settings, scans of the same patient and the same area during the same period are often misregistered. Therefore, before using these data as training data for the model, they require preprocessing. This preprocessing step includes N4 bias correction for MRI images, denoising of MRI and CT images, registration of the two images, and slice processing.

[0062] Figure 2 The preprocessing flow of MRI image data and CT image data is shown.

[0063] According to an embodiment of the present invention, Figure 2 The pre-processing process described in step S1 is as follows:

[0064] First, the N4 bias field correction is performed on the collected MRI image data that displays unevenness. Then, the corrected MRI image data and the collected CT image data are denoised to prevent the noise generated during the scanning process from affecting subsequent work. The denoised MRI image data and CT image data belonging to the same patient, the same period, and the same part are rigidly aligned to ensure that the size of the MRI image data and the CT image data are consistent. Finally, slicing is performed to slice the 3D registered image data according to a preset ratio, and slices with less information than the set value are removed to obtain 2D image data sets of MRI images and CT.

[0065] S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image;

[0066] It should be noted that the model described in this embodiment requires a dataset that includes not only complete MRI and CT images, but also skeletal mask data, CT bone images, and CT tissue images. Taking advantage of the fact that pixel values ​​in skeletal areas of CT images differ significantly from those in other areas, the K-means method is used to cluster CT image pixels.

[0067] Figure 3 The process of obtaining a binary mask image is shown.

[0068] According to an embodiment of the present invention, Figure 3 The processing flow of step S2 is shown as follows:

[0069] First, a K-means clustering model was constructed, using the pixel values ​​of the CT image as feature values. The distance between each feature was calculated using the L1 distance calculation method. The model parameters were then initialized, with the value of K set to 3 and the features of the centers of the initial clusters of each type initialized to 0, 127, and 255. The algorithm was iterated until the classification of each pixel no longer changed.

[0070] The K-mean clustering model marks the CT image into black background, gray tissue, and white bone classes. After clustering, the class with the largest eigenvalue is taken as the binary mask image m of the bone in the CT image, where the value 1 indicates that the CT pixel at that position belongs to the bone part, and 0 indicates other parts.

[0071] S3. Extract the bone and non-bone portions of the CT image using the binary mask image.

[0072] According to an embodiment of the present invention, the processing flow described in step S3 is as follows:

[0073] The CT image is converted into matrix data with a numerical range of 0 to 255, and the partial data at the same position in the CT image as the binary mask image m described in S2 with a value of 1 is extracted as the CT bone image b. This process can be expressed as: b = m × y, where y represents the CT image data after S1 preprocessing; the rest of the CT image is extracted as the CT tissue image t. This process can be expressed as: t = yb.

[0074] S4. Establish a deep learning network model for MRI image conversion to CT based on GAN;

[0075] Figure 4 The deep learning network model framework for converting MRI images to CT based on GAN is shown.

[0076] According to an embodiment of the present invention, Figure 4 As shown, the deep learning network model of MRI image to CT conversion based on GAN includes a generator network G and a discriminator network D, wherein the generator network G includes a CT bone segmentation sub-model Bone_segm, a bone synthesis sub-model Bone_syn, a tissue synthesis sub-model Tissue_syn, and an integration model Integ;

[0077] The skeleton segmentation sub-model model function expression is:

[0078]

[0079] Among them, x i ∈R N is the MRI image preprocessed by S1, the MRI image may be a T1 type MRI image or a T2 type MRI image, and the MRI image is the input of the segmentation model, is the bone segmentation image of the MRI image. The bone segmentation sub-model segments the MRI image and obtains the corresponding bone mask, which serves as the prior information for other networks. Bone_segm(·) represents the deep learning network model that maps the MRI image to the bone mask.

[0080] In this solution, the skeleton synthesis sub-model function expression is:

[0081]

[0082] Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Cat[·] represents the splicing operation in the channel dimension, Bone_syn(·) represents a deep learning network model for mapping the MRI image to the CT bone.

[0083] In this scheme, the tissue synthesis sub-model function expression in step S4 is:

[0084]

[0085] wherein x i ∈R N is the MRI image after preprocessing S1, is the bone mask obtained by segmenting the same MRI image using a bone segmentation network, represents a mask inversion operation, i.e. Cat[·] represents a channel dimension concatenation operation, Tissue_syn(·) represents a deep learning network model for mapping the MRI image to the CT tissue.

[0086] In this scheme, the integrated model function expression is:

[0087]

[0088] wherein x i ∈R N is the MRI image after preprocessing S1, is the bone mask obtained by segmenting the same MRI image using a bone segmentation network, is a bone image synthesized by the bone synthesis model, is a tissue image synthesized by the tissue synthesis model, Integ(·) represents a model for integrating the synthesized bone image and tissue image into a final CT image.

[0089] In this scheme, the generator G expression is:

[0090] G=Integ(Bone_syn(·),Tissue_syn(·),Bone_segm(·))

[0091] wherein Bone_syn is the bone synthesis sub-model, Tissue_syn is the tissue synthesis sub-model, and Bone_segm is the bone segmentation sub-model.

[0092] It should be noted that the execution flow of the generator is as follows: first, the bone segmentation network is used to segment the MRI image to obtain the bone mask This mask is then used as prior information, along with the MRI image, as input to other sub-networks, providing explicit bone and non-bone information. Next, the bone generation network and tissue generation network generate corresponding CT image portions. Finally, the integration sub-network adaptively fuses the bone and tissue synthesis data to obtain the final composite CT image.

[0093] S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ;

[0094] According to an embodiment of the present invention, the adversarial loss L is generated. adv Expressed as: L adv =E x [log(1-D(x,G(x)))], where x represents the MRI image input image, E x It is obtained from the data distribution in the dataset;

[0095] The pixel level L1 loss L pix Expressed as:

[0096]

[0097] Among them, b i is the CT bone image extracted in step S3, is the skeleton image synthesized by the skeleton synthesis network, t i is the CT tissue image extracted in step S3, is the tissue image synthesized by the tissue synthesis network, y i is the CT image obtained by preprocessing in step S2, is the final synthetic CT image synthesized by the generator G, θ1, θ2 and θ3 are L pix_b , L pix_t and L pix_y The corresponding weight coefficient.

[0098] The binary cross entropy loss L BCE Expressed as:

[0099]

[0100] Among them, m i The binary mask data of the CT bones extracted in step S4, Mask image segmented by the skeleton segmentation network.

[0101] S6. Input the MRI image into the deep learning-based adversarial generative network model for synthesizing CT images from MRI images, thereby converting the preprocessed MRI image into a synthesized CT image.

[0102] Figure 5 A system block diagram of synthesizing CT images from MRI images based on deep learning is shown.

[0103] like Figure 5 As shown, a second aspect of the present invention provides a system for synthesizing CT images from MRI images based on deep learning, including a memory 51 and a processor 52. The memory includes a method program for synthesizing CT images from MRI images based on deep learning. When the method program for synthesizing CT images from MRI images based on deep learning is executed by the processor, the following steps are implemented:

[0104] S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and CT image, registration of the MRI image and CT image, and slice processing;

[0105] S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image;

[0106] S3. Extract the bone and non-bone portions of the CT image using the binary mask image.

[0107] S4. Establish a deep learning network model for MRI image conversion to CT based on GAN;

[0108] S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ;

[0109] S6. The preprocessed MRI image data is input into a GAN-based deep learning network model for MRI image to CT conversion, and a CT image is output.

[0110] The method described in the present invention includes two parts: CT bone mask, bone and tissue extraction, and image synthesis. In the subsequent CT image synthesis task, if the synthesis model based on the segmentation integration idea has been trained, the MRI image is directly input into the CT image synthesis model to obtain the CT synthesis result.

[0111] By analyzing and verifying the pixel information of the bone pixels and tissue parts of the CT image, the present invention finds that the pixel distribution of the bones and other parts of the CT image is greatly different and easy to segment. Based on the above characteristics, a method and system for synthesizing CT images from MRI images based on deep learning are proposed. Based on the fitting ability of deep learning, a deep learning neural network model including different sub-models is constructed to learn the distribution of bones and tissues respectively, and then fuse and integrate them. At the same time, utilizing the characteristics of CT images, the K-means clustering method is used for adaptive segmentation. In the original MR-CT data set, it is easy to obtain the required bone mask data in batches. After the bones are separated from the tissues, their distribution information is clearer, which simplifies the synthesis task of each synthesis sub-network. At the same time, different sub-models can be selected for synthesis according to the characteristics of different parts, resulting in a more targeted network, which can greatly reduce the complexity of the network. Compared with the existing CT image synthesis method, the quality of the synthesized image is improved.

[0112] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0113] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0114] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0115] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks or optical disks, and other media that can store program codes.

[0116] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. A method for synthesizing CT images from MRI images based on deep learning, characterized in that: The steps include: S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and CT image, registration of the MRI image and CT image, and slice processing; S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image; S3. Extract the bone and non-bone portions of the CT image using the binary mask image. S4. Establish a deep learning network model for MRI image conversion to CT based on GAN, which includes a generator network G and a discriminator network D, wherein the generator network G includes a CT bone segmentation sub-model Bone_segm, a bone synthesis sub-model Bone_syn, a tissue synthesis sub-model Tissue_syn, and an integration model Integ; The skeleton segmentation sub-model model function expression is: Among them, x i ∈R N is the MRI image preprocessed by S1, the MRI image may be a T1 type MRI image or a T2 type MRI image, and the MRI image is the input of the segmentation model, is the bone segmentation image of the MRI image. The bone segmentation sub-model segments the MRI image and obtains the corresponding bone mask, which serves as the prior information for other networks. Bone_segm(·) represents the deep learning network model that maps the MRI image to the bone mask. The skeleton synthesis sub-model function expression is: Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Cat[·] represents the splicing operation in the channel dimension, is the CT bone image synthesized by the model, and Bone_syn(·) represents the deep learning network model that maps MRI images to CT bones. S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ; S6. The preprocessed MRI image data is input into a GAN-based deep learning network model for MRI image to CT conversion, and a CT image is output.

2. The method for synthesizing CT images from MRI images based on deep learning according to claim 1, characterized in that: The pre-processing process described in step S1 is: First, the N4 bias field correction is performed on the collected MRI image data that displays unevenness. Then, the corrected MRI image data and the collected CT image data are denoised. The denoised MRI image data and CT image data belonging to the same patient, the same period, and the same part are rigidly aligned to ensure that the size of the MRI image data and the CT image data are consistent. Finally, slicing is performed to slice the 3D registered image data according to a preset ratio, and slices with less information than the set value are removed to obtain 2D image datasets of MRI images and CT.

3. The method for synthesizing CT images from MRI images based on deep learning according to claim 2, characterized in that: The processing flow described in step S2 is: First, a K-means clustering model was constructed, using the pixel values ​​of the CT image as feature values. The distance between each feature was calculated using the L1 distance calculation method. The model parameters were then initialized, with the value of K set to 3 and the features of the centers of the initial clusters of each type initialized to 0, 127, and 255. The algorithm was iterated until the classification of each pixel no longer changed. The K-mean clustering model marks the CT image into black background, gray tissue, and white bone classes. After clustering, the class with the largest eigenvalue is taken as the binary mask image m of the bone in the CT image, where the value 1 indicates that the CT pixel at that position belongs to the bone part, and 0 indicates other parts.

4. The method for synthesizing CT images from MRI images based on deep learning according to claim 3, characterized in that: The processing flow described in step S3 is: The CT image is converted into matrix data with a numerical range of 0 to 255, and the partial data at the same position in the CT image as the binary mask image m described in S2 with a value of 1 is extracted as the CT bone image b. This process can be expressed as: b = m × y, where y represents the CT image data after S1 preprocessing; the rest of the CT image is extracted as the CT tissue image t. This process can be expressed as: t = yb.

5. The method for synthesizing CT images from MRI images based on deep learning according to claim 4, characterized in that: The functional expression of the tissue synthesis sub-model in step S4 is: Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Indicates the mask inversion operation, that is Cat[·] represents the concatenation operation in the channel dimension, is the CT tissue image synthesized by the model, and Tissue_syn(·) represents the deep learning network model that maps MRI images to CT tissue.

6. The method for synthesizing MRI images into CT images based on deep learning according to claim 4, characterized in that: The integrated model function expression is: Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Synthesize bone images for the bone synthesis model, Tissue images synthesized by the tissue synthesis model, is the final synthesized CT image, Integ(·) represents the model that integrates the synthesized bone image and tissue image into the final CT image; The generator G expression is: G=Integ(Bone_syn(·), Tissue_syn(·), Bone_segm(·)) Among them, Bone_syn is the bone synthesis sub-model, Tissue_syn is the tissue synthesis sub-model, and Bone_segm is the bone segmentation sub-model.

7. The method for synthesizing MRI images into CT images based on deep learning according to claim 1, characterized in that: The generated adversarial loss L in step S5 adv Expressed as: L adv =E x [log(1-D(x, G(x)))], where x represents the MRI image input image, E x It is obtained from the data distribution in the dataset; The pixel level L1 loss L pix Expressed as: Among them, b i is the CT bone image extracted in step S3, is the skeleton image synthesized by the skeleton synthesis network, t i is the CT tissue image extracted in step S3, is the tissue image synthesized by the tissue synthesis network, y i is the CT image obtained by preprocessing in step S2, is the final synthetic CT image synthesized by the generator G, θ1, θ2 and θ3 are L pix_b 、L pix_t and L pix_y The corresponding weight coefficient; The binary cross entropy loss L BCE Expressed as: Among them, m i The binary mask data of the CT bones extracted in step S4, Mask image segmented by the skeleton segmentation network.

8. A system for synthesizing MRI images and CT images based on deep learning, comprising a memory and a processor, characterized in that: The memory includes a method program for synthesizing MRI images into CT images based on deep learning. When the method program for synthesizing MRI images into CT images based on deep learning is executed by the processor, the following steps are implemented: S1. Acquire MRI image data and CT image data and perform preprocessing, wherein the preprocessing steps include N4 bias correction of the MRI image, denoising of the MRI image and CT image, registration of the MRI image and CT image, and slice processing; S2. Use the K-means algorithm to mark the bone positions of the preprocessed CT image and obtain the corresponding binary mask image; S3. Extract the bone and non-bone portions of the CT image using the binary mask image. S4. Establish a deep learning network model for MRI image conversion to CT based on GAN, which includes a generator network G and a discriminator network D, wherein the generator network G includes a CT bone segmentation sub-model Bone_segm, a bone synthesis sub-model Bone_syn, a tissue synthesis sub-model Tissue_syn, and an integration model Integ; The skeleton segmentation sub-model model function expression is: Among them, x i ∈R N is the MRI image preprocessed by S1, the MRI image may be a T1 type MRI image or a T2 type MRI image, and the MRI image is the input of the segmentation model, is the bone segmentation image of the MRI image. The bone segmentation sub-model segments the MRI image and obtains the corresponding bone mask, which serves as the prior information for other networks. Bone_segm(·) represents the deep learning network model that maps the MRI image to the bone mask. The skeleton synthesis sub-model function expression is: Among them, x i ∈R N is the MRI image after S1 preprocessing, is the bone mask obtained by segmenting the same MRI image through the bone segmentation network, Cat[·] represents the splicing operation in the channel dimension, is the CT bone image synthesized by the model, and Bone_syn(·) represents the deep learning network model that maps MRI images to CT bones. S5. Construct the loss function of the network model, which includes generating adversarial loss L adv , pixel-level L1 loss L pix , binary cross entropy loss L BCE ; S6. The preprocessed MRI image data is input into a GAN-based deep learning network model for MRI image to CT conversion, and a CT image is output.

Images