A method for stitching and fusing based on segmented MRI images of a digital human body

Abstract

The application discloses a kind of based on digital human segmentation MRI image splicing and fusion method, comprising: the MRI data of different parts of human body obtained by nuclear magnetic resonance scanning is first registered after being preprocessed such as noise reduction, using the resolution, orientation etc. of reference image and floating image after registration, further realize the transformation of floating image between image coordinate system and world coordinate system, simultaneously based on trilinear interpolation interpolation and operation in the above spatial transformation process.The application realizes the seamless fusion of the three-dimensional image to be spliced in high and low frequency direction, effectively utilizes image high and low frequency component information.The application not only simplifies the two-stage image splicing and fusion process of traditional splicing method, i.e.fusing overlapping area first and then splicing non-overlapping area, and greatly improves the overall splicing and fusion accuracy of segmented MRI data.

Description

Technical Field

[0001] This invention relates to an algorithm for stitching and fusing segmented MRI data of the human body obtained by magnetic resonance scanning, belonging to the field of computer image processing. Background Technology

[0002] Magnetic Resonance Imaging (MRI) utilizes the principle of nuclear magnetic resonance. Based on the varying attenuation of emitted energy in different structural environments within matter, and by detecting the emitted electromagnetic waves through an applied gradient magnetic field, it acquires images of the internal structures of an object. MRI is a commonly used clinical diagnostic technique in the field of medical imaging. Compared to imaging methods such as CT and X-rays, MRI offers more comprehensive imaging parameters and higher soft tissue resolution, typically providing more accurate soft tissue contrast. This is of great significance for studying diseased organs in different clinical locations. However, in clinical medical research, MRI scans are usually performed on only localized areas of the human body, and there is very little whole-body MRI data available in publicly available datasets.

[0003] Currently, in the field of medical image processing, rigid registration technology for 3D images is relatively mature. However, since the registered reference image and the floating image often have different orientation information, the integrated stitching and fusion of the two segments of 3D data after registration remains a problem to be solved. Traditional image fusion algorithms can be mainly divided into pixel-level image fusion, feature-level image fusion, and decision-level image fusion strategies. Pixel-level image fusion algorithms directly integrate the data information of each source image; feature-level image fusion methods process the feature information extracted from each source image; decision-level image fusion is an image fusion method that makes the optimal decision based on certain criteria and the confidence level of each decision. Feature-level and decision-level image fusion methods require extensive reasoning and analysis by an expert decision system before registration and fusion. In current traditional image fusion algorithms, pixel-level image fusion strategies are more commonly used. Common traditional image fusion methods include:

[0004] (1) Image fusion based on single scale:

[0005] Single-scale image fusion methods directly fuse source images, employing either maximum or mean fusion strategies for overlapping regions. This approach is a relatively simple and direct pixel-level image fusion method.

[0006] (2) Multi-scale image fusion based on pyramid transform:

[0007] Multi-scale image fusion methods are a pyramidal decomposition-based image fusion strategy. This requires dividing the source image into sub-images in different frequency bands, and then selecting appropriate fusion operators for reconstruction based on the image features of each frequency band. Pyramid-based image fusion algorithms first perform pyramidal decomposition on the source images to be fused, such as Gaussian pyramid decomposition or Laplacian pyramid decomposition, to obtain image pyramids at different resolutions. Then, corresponding fusion operators are selected for the data at different frequency levels obtained after decomposition to obtain the corresponding fusion pyramids at each frequency level. Finally, the fused result image corresponding to this set of source images is obtained through pyramidal reconstruction.

[0008] (3) Non-subsampled Contourlet Transform (NSCT Transform) based on multi-directional and multi-resolution:

[0009] The NSCT transform is also a multi-scale image fusion strategy, mainly composed of a non-subsampled pyramid structure (NSP structure) and a non-subsampled directional filter bank (NSDFB structure). The NSP structure performs pyramid decomposition on the input source image, dividing it into high-frequency and low-frequency sub-bands. The NSDFB structure further subdivides the high-frequency sub-band into multiple directional sub-bands, obtaining high-frequency information in different directions. The low-frequency sub-band can be further decomposed upwards. After decomposition, corresponding fusion operators are used to reconstruct the image based on the data features at different resolutions, yielding the final fusion result.

[0010] Generally, multi-scale image fusion methods can achieve better fusion results than single-scale image fusion methods. However, these fusion algorithms are usually only applicable to source images with similar structural information. In the problem of stitching three-dimensional segmented human MRI data, the two data segments to be stitched usually have only about 10 to 20 overlapping layers. Although these traditional fusion algorithms can solve the fusion problem between structurally similar overlapping regions, there are usually obvious stitching seams between the fused result of the overlapping region and the volume data segments outside the overlapping region that did not participate in the fusion. This requires the introduction of other stitching seam elimination algorithms for optimization, such as the stitching seam elimination strategy based on curve fitting. The main idea of ​​this algorithm is as follows, and the algorithm flowchart can be seen in the appendix. Figure 2 :

[0011] (1) Calculate the grayscale mean difference Δg between the two volumes to be spliced, and calculate the correction width W based on the grayscale mean difference;

[0012] (2) Adjustments were made using the "forced correction method" approach;

[0013] 2.1 Calculate the average grayscale value (avg) of the volume data voxels within the correction width W;

[0014] 2.2 Obtain and save the texture information of voxels within the range of W based on the mean avg;

[0015] 2.3 Perform cubic curve fitting on the gray values ​​of these 2*W individual pixels;

[0016] 2.4 Based on the new fitted value obtained from curve fitting and combined with texture information, the corrected new grayscale value grey1 is obtained;

[0017] 2.5 It is determined that the gray value of each voxel within the W range is a linear combination of the original voxel's gray value grey0 and the new gray value grey1;

[0018] (3) Further fine-tune the results obtained by the forced correction method, that is, make another fine-tune the voxel points within W / 5 range on both sides of the splice seam.

[0019] Applying this curve fitting-based seam elimination algorithm to the image stitching and fusion problem can reduce the seam problem between overlapping and non-overlapping regions to a certain extent. However, this two-stage image stitching and fusion method is not easy to operate, and the strategy of first fusing overlapping regions and then stitching non-overlapping regions has an unavoidable seam problem in theory. Summary of the Invention

[0020] To address the seam problem caused by traditional algorithms that first fuse overlapping regions and then stitch non-overlapping regions, and to simplify the integrated fusion and stitching process of 3D data, this invention provides a stitching and fusion method based on segmented MRI images of the digital human body. This method achieves seamless fusion of the 3D data to be stitched in both low-frequency and high-frequency components, improves the stitching and fusion accuracy of segmented MRI data, and simplifies the operation process.

[0021] To achieve the above objectives, the present invention provides the following technical solution:

[0022] A method for stitching and fusing segmented digital human MRI images includes the following steps:

[0023] Step 1: Acquire two segments of human MRI data obtained from clinical scans, and preprocess the two segments of human MRI data respectively;

[0024] Step 2: Identify the reference image and the floating image from the two preprocessed MRI data segments and rigidly register them;

[0025] Step 3: Based on the resolution and spatial information of the registered reference image and the floating image, the transformation of the floating image between the IJK coordinate system and the RAS coordinate system is realized using the trilinear interpolation algorithm.

[0026] Step 4: Resample the two 3D data segments to be stitched together and crop out invalid regions;

[0027] Step 5: Using the DWT transform based on the db2 wavelet basis, the high-frequency and low-frequency components of the two segments of three-dimensional data to be stitched are separated.

[0028] Step 6: Based on the data distribution characteristics of the separated high and low frequency components, perform image normalization respectively;

[0029] Step 7: Using a 3DUnet-based network framework, seamless fusion of the low-frequency and high-frequency components separated from the two segments of data to be spliced ​​is achieved.

[0030] Step 8: Obtain the low-frequency and high-frequency components in each direction after fusion from the splicing and fusion network in Step 7, and use the inverse discrete wavelet transform based on db2 to integrate the high-frequency fusion components and the low-frequency fusion components to obtain the final splicing and fusion result.

[0031] Furthermore, in step 1, a median filter algorithm is used for preprocessing, and the preprocessing includes at least noise reduction processing.

[0032] Furthermore, in step 2, a rigid registration process from the floating image to the reference image is implemented using the registration method provided by the SimpleITK library. Specifically, the process includes the following steps: selecting mutual information as the image similarity measure, using the stochastic gradient descent algorithm as the optimizer, using the trilinear interpolation algorithm as the interpolator, and using the output geometric transformation parameters to update and iterate the entire registration algorithm model through the rigid transformation model.

[0033] Furthermore, step 3 specifically includes the following process:

[0034] Using a fixed image as a reference, spatial initialization preprocessing is performed on the floating image:

[0035] ① Based on the z-axis resolution of the fixed image and the floating image, obtain the maximum resolution in the z-axis direction after stitching the two segments of 3D data. Combine this with the resolution of the data in the x-axis and y-axis to initialize a stitched result image.

[0036] ② The floating image undergoes spatial transformations in the IJK and RAS coordinate systems to determine its final relative position with respect to the reference image in the stitching result. This process includes the following steps: First, based on the number of overlapping layers in the two segments of data to be stitched, the starting layer number of the floating image in the final stitching result is roughly determined. Then, combining the origin, voxel spacing, and orientation information of the reference image, each layer of data in the initial stitching result starting from that layer undergoes a transformation from the image coordinate system to the world coordinate system (IJKtoRAS), as shown in Formula 1. Points transformed to the world coordinate system are then transformed again from the world coordinate system (IJKtoRAS) to the image coordinate system (RAStoIJK), combining the origin, voxel spacing, and orientation information of the floating image, as shown in Formula 2. Data mapped from the reference image to the floating image that is not on grid points is interpolated and calculated using a trilinear interpolation algorithm.

[0037]

[0038]

[0039] in, Represents the coordinate values ​​of the image in the world coordinate system. This represents the subscript value in the image coordinate system, where D represents the orientation matrix parameter. This represents the spatial resolution parameter of the image. Represents the coordinates of the origin of the image.

[0040] Furthermore, step 4 specifically includes the following process:

[0041] The voxel spacing between the reference image and the floating image after the above spatial transformation is resampled to (1,1,1), and the resampling formula is shown in (3). Then, the invalid background area in the two segments of data to be stitched is cropped.

[0042]

[0043] in, This indicates the size of the 3D data before resampling. This represents the voxel spacing in the X, Y, and Z axes of the 3D data before resampling. This indicates the size of the 3D data after resampling. This represents the voxel spacing of the 3D data after resampling.

[0044] Furthermore, step 5 specifically includes the following process: performing three-dimensional DWT transformation based on the db2 wavelet basis on the two segments of MRI human data to be spliced; separating a low-frequency component and high-frequency components in seven other directions from the three-dimensional data.

[0045] Furthermore, in step 6, the low-frequency components are normalized using the minimum-maximum normalization method, as shown in formulas 4 and 5:

[0046]

[0047]

[0048] Among them, low freq and high freq Let min(low) represent the low-frequency and high-frequency components of the image separated from the original volume data, respectively. freq ) and max(low freq The numbers ) represent the minimum and maximum values ​​of the low-frequency component, respectively, and mean(high) represents the minimum and maximum values ​​of the low-frequency component. freq ) and max(abs(high freq )) represent the mean and maximum values ​​of the high-frequency components, respectively.

[0049] Furthermore, in step 7, the normalized low-frequency and high-frequency components are trained using neural networks based on 3DUnet. The last convolutional layer of the neural network corresponding to the low-frequency component uses the sigmoid function as the activation function, and the last convolutional layer of the neural network corresponding to the high-frequency component uses the hyperbolic tangent function as the activation function.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] 1. This invention combines the Discrete Wavelet Transform (DWT Transform) based on the db2 wavelet basis with a 3DUnet-like network based on deep learning to achieve seamless fusion of the three-dimensional images to be stitched in the high and low frequency directions, effectively utilizing the high and low frequency component information of the images.

[0052] 2. This invention trains a grayscale adaptive 3DUnet network to achieve an integrated stitching and fusion process of human segmented MRI data obtained from magnetic resonance scanning. This not only simplifies the two-stage image stitching and fusion process in traditional stitching methods, which involves first fusing overlapping areas and then stitching non-overlapping areas, but also greatly improves the overall stitching and fusion accuracy of segmented MRI data.

[0053] 3. The image stitching and fusion algorithm based on deep learning implemented in this invention not only solves the problem of the traditional two-stage stitching algorithm being difficult to operate, but also greatly optimizes the effect of stitching human segmented MRI data. The wavelet transform based on the db2 wavelet basis fully preserves the high and low frequency information of the image. Compared with the traditional downsampling method, the image processing method adopted in this invention reduces the image resolution while still preserving the high and low frequency information components in the image to a great extent, effectively ensuring the resolution of the final stitching result.

[0054] 4. This invention stitches and fuses segmented MRI data acquired from magnetic resonance imaging (MRI) scans, achieving integrated stitching and fusion of human segmented MRI data and obtaining superior three-dimensional image fusion results compared to traditional algorithms. Based on this, subsequent fusion and display of various physiological parameters allows for clinical research on various human organs with higher soft tissue resolution. Attached Figure Description

[0055] Figure 1 The overall flowchart of stitching and fusing based on segmented MRI images of the human body provided by the present invention;

[0056] Figure 2 This is a flowchart of a seam elimination algorithm based on curve fitting.

[0057] Figure 3 This is a schematic diagram of the neural network structure used in this invention;

[0058] Figure 4 The diagram shows the components of the reference image and the floating image to be stitched after wavelet transform. The upper part of the diagram shows the original image of the reference image to be stitched, the schematic diagram of the 3D model, and the corresponding low-frequency component and high-frequency components in 7 directions after wavelet transform. The lower part of the diagram shows the original image of the floating image to be stitched, the schematic diagram of the 3D model, and the results of each component after wavelet transform.

[0059] Figure 5 This is a schematic diagram of the result after the components to be spliced ​​from the DWT transform are spliced ​​and fused by a neural network.

[0060] Figure 6 This diagram illustrates the comparison between the stitching results of different human body parts obtained using traditional algorithms and the results obtained by this invention. The reference image is labeled "fixed," and the floating image is labeled "moving."

[0061] (6-a1), (6-a2), and (6-a3) are two-dimensional cross-sectional images of the coronal plane obtained by directly stitching together three-dimensional segmented human MRI data in the chest, waist, and pelvic regions using traditional algorithms.

[0062] (6-b1), (6-b2), and (6-b3) are two-dimensional cross-sectional images of the coronal plane obtained by stitching and fusing three-dimensional segmented human MRI data in the chest, waist, and pelvic regions using the present invention. Detailed Implementation

[0063] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0064] This invention provides a method for stitching and fusing segmented MRI images of the human body. First, a reference image and a floating image are determined from human MRI data acquired through clinical scans and rigidly registered. The origin, voxel spacing, and orientation information of the registered data are used to achieve spatial transformation of the floating image between the IJK and RAS coordinate systems. The data to be input to the network is resampled for voxel spacing and invalid background regions are cropped to accelerate the convergence speed of subsequent networks. Simultaneously, to avoid loss of effective information due to downsampling and to better fuse high and low frequency information, discrete wavelet transform based on the db2 wavelet basis is used to separate the high and low frequency signal components of the image to be stitched. Based on the data distribution characteristics of the separated high and low frequency signal components, appropriate data normalization methods are selected. The processed data is fed into a pre-trained 3DUnet-like network to achieve seamless fusion of various frequency components. The fusion results of each component output by the network are then integrated using inverse wavelet transform. Finally, the integrated stitching and fusion result of this deep learning-based MRI segmented data is compared with the stitching result obtained using traditional algorithms to evaluate the performance of the network model.

[0065] Specifically, the overall process of this invention is as follows: Figure 1 As shown, it includes the following steps:

[0066] Step 1: Acquire two segments of human MRI data obtained from clinical scans, and use the median filter algorithm to perform noise reduction and other preprocessing operations on the two segments of MRI data to be stitched together.

[0067] Step 2: Determine the reference image and the floating image from the two preprocessed MRI data segments. Use the registration method provided by the SimpleITK library to realize the rigid registration process from the floating image to the reference image. Select mutual information as the image similarity measure, use the stochastic gradient descent algorithm as the optimizer, and use the trilinear interpolation algorithm as the interpolator. The output geometric transformation parameters are updated and iterated through the rigid transformation model to realize the entire registration algorithm model.

[0068] Step 3: Using the fixed image as a reference, perform spatial initialization preprocessing on the floating image:

[0069] ① Based on the z-axis resolution of the fixed and floating images, obtain the maximum resolution in the z-axis direction after stitching the two 3D data segments. Combine this with the resolution of the data in the x and y axes to initialize a stitched result image.

[0070] ② Since the registered reference image and the floating image usually have different orientation information, the floating image needs to undergo spatial transformation between the IJK coordinate system and the RAS coordinate system to determine its final relative position with respect to the reference image in the stitching result. Specifically: First, based on the number of overlapping layers of the two segments of data to be stitched, the starting layer number of the floating image in the final stitching result is roughly determined; combining the origin, voxel spacing, and orientation information of the reference image, each layer of data in the initial stitching result starting from this layer is transformed from the image coordinate system to the world coordinate system (referred to as IJK to RAS transformation), as shown in Formula 1; the points transformed to the world coordinate system are then transformed back to the image coordinate system (referred to as RAS to IJK transformation) by combining the origin, voxel spacing, and orientation information of the floating image, as shown in Formula 2; data mapped from the reference image to the floating image that are not on grid points are interpolated and calculated using a trilinear interpolation algorithm.

[0071]

[0072]

[0073] in, Represents the coordinate values ​​of the image in the world coordinate system. This represents the subscript value in the image coordinate system, where D represents the orientation matrix parameter. This represents the spatial resolution parameter of the image. Represents the coordinates of the origin of the image.

[0074] Step 4: Resample the voxel spacing between the spatially transformed reference image and the floating image to (1,1,1), using the resampling formula shown in (3), and then crop the invalid background regions in the two segments of data to be stitched together. Indicates the dimensions of the 3D data before resampling. This represents the voxel spacing in the X, Y, and Z axes of the 3D data before resampling. This indicates the size of the 3D data after resampling. This represents the voxel spacing of the 3D data after resampling.

[0075]

[0076] Step 5: To fully utilize the effective information of the original data at each frequency component while avoiding data loss due to downsampling, a three-dimensional DWT transformation based on the db2 wavelet basis is performed on the two MRI human body data segments to be stitched together. From each three-dimensional data segment, one corresponding low-frequency component and seven other high-frequency components in the other directions are separated. The data size of each separated component is half the size of the original data, as shown in the attached figure. Figure 4 As shown.

[0077] Step 6: Based on the data distribution characteristics of the high and low frequency components obtained from the DWT transform, the low-frequency and high-frequency components separated from the two segments of human MRI data to be stitched are normalized respectively. The low-frequency component is normalized using the minimum-maximum normalization method, and the high-frequency component is normalized, as shown in formulas 4 and 5. Wherein, low freq and high freq Let min(low) represent the low-frequency and high-frequency components of the image separated from the original volume data, respectively. freq ) and max(low freq The numbers ) represent the minimum and maximum values ​​of the low-frequency component, respectively, and mean(high) represents the minimum and maximum values ​​of the low-frequency component. freq ) and max(abs(high freq )) represent the mean and maximum values ​​of the high-frequency components, respectively.

[0078]

[0079]

[0080] Step 7, Train the Neural Network: Train the model using a 3DUnet-based neural network for the normalized low-frequency and high-frequency components respectively. The specific network model structure can be found in the appendix. Figure 3 In this case, the last convolutional layer of the neural network corresponding to the low-frequency components uses the sigmoid function as the activation function, while the last convolutional layer of the neural network corresponding to the high-frequency components uses the hyperbolic tangent function as the activation function.

[0081] Step 8: The low-frequency and high-frequency components of the two human MRI data segments to be stitched are processed by the trained neural network to obtain the fusion results in the corresponding directions. The fusion results of each component are shown in the attached figure. Figure 5 As shown, the fusion results in various directions are integrated using the inverse DWT transform based on the db2 wavelet basis to obtain the final stitched fusion result of the segmented MRI data. The result is compared with the effect of direct stitching by traditional algorithms, and the performance of the fusion network model is evaluated.

[0082] To verify the effectiveness of the human segmented MRI data stitching and fusion algorithm disclosed in this invention, a set of human MRI datasets acquired from clinical scans was used to demonstrate the significant improvement of the method disclosed in this invention compared to traditional algorithms. Comparative experiments were conducted using MRI data from the chest, lumbar region, and pelvis as examples. Figure 6The images show a comparison of the coronal and sagittal plane effects of the stitching and fusion results of three-dimensional MRI data obtained using the traditional algorithm and the stitching algorithm of this invention, respectively. In comparison, the method provided by this invention can achieve seamless stitching of human segmented MRI data while ensuring data resolution, and effectively solves the stitching seam problem that occurs in traditional stitching algorithms. The results are significantly improved and have profound significance for conducting clinical medical research.

[0083] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for stitching and fusing segmented digital human MRI images, characterized in that, Includes the following steps: Step 1: Acquire two segments of human MRI data obtained from clinical scans, and preprocess the two segments of human MRI data respectively; Step 2: Identify the reference image and the floating image from the two preprocessed MRI data segments and rigidly register them; Step 3: Based on the resolution and spatial information of the registered reference image and the floating image, the transformation of the floating image between the IJK coordinate system and the RAS coordinate system is realized using the trilinear interpolation algorithm. Step 4: Resample the two 3D data segments to be stitched together and crop out invalid regions; Step 5: Using the DWT transform based on the db2 wavelet basis, the high-frequency and low-frequency components of the two segments of three-dimensional data to be stitched are separated. Step 6: Based on the data distribution characteristics of the separated high and low frequency components, perform image normalization respectively; Step 7: Using a 3DUnet-based network framework, seamless fusion of the low-frequency and high-frequency components separated from the two segments of data to be spliced ​​is achieved. Step 8: Obtain the low-frequency and high-frequency components in each direction after fusion from the splicing and fusion network in Step 7, and use the inverse discrete wavelet transform based on db2 to integrate the high-frequency fusion components and the low-frequency fusion components to obtain the final splicing and fusion result.

2. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, In step 1, a median filter algorithm is used for preprocessing, which includes at least noise reduction.

3. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, In step 2, a rigid registration process from a floating image to a reference image is implemented using the registration method provided by the SimpleITK library. Specifically, the process includes the following steps: mutual information is selected as the image similarity measure, stochastic gradient descent algorithm is used as the optimizer, trilinear interpolation algorithm is used as the interpolator, and the output geometric transformation parameters are updated and iterated through the rigid transformation model to realize the entire registration algorithm model.

4. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, Step 3 specifically includes the following process: Using a fixed image as a reference, spatial initialization preprocessing is performed on the floating image: ① Based on the z-axis resolution of the fixed image and the floating image, obtain the maximum resolution in the z-axis direction after stitching the two segments of 3D data. Combine this with the resolution of the data in the x-axis and y-axis to initialize a stitched result image. ② The floating image undergoes spatial transformations in the IJK and RAS coordinate systems to determine its final relative position with respect to the reference image in the stitching result. This process includes the following steps: First, based on the number of overlapping layers in the two segments of data to be stitched, the starting layer number of the floating image in the final stitching result is roughly determined. Then, combining the origin, voxel spacing, and orientation information of the reference image, each layer of data in the initial stitching result starting from that layer undergoes a transformation from the image coordinate system to the world coordinate system (IJKtoRAS), as shown in Formula 1. Points transformed to the world coordinate system are then transformed again from the world coordinate system (IJKtoRAS) to the image coordinate system (RAStoIJK), combining the origin, voxel spacing, and orientation information of the floating image, as shown in Formula 2. Data mapped from the reference image to the floating image that is not on grid points is interpolated and calculated using a trilinear interpolation algorithm. in, Represents the coordinate values ​​of the image in the world coordinate system. This represents the subscript value in the image coordinate system, where D represents the orientation matrix parameter. This represents the spatial resolution parameter of the image. Represents the coordinates of the origin of the image.

5. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, Step 4 specifically includes the following process: The voxel spacing between the reference image and the floating image after the above spatial transformation is resampled to (1,1,1), and the resampling formula is shown in (3). Then, the invalid background area in the two segments of data to be stitched is cropped. in, This indicates the size of the 3D data before resampling. This represents the voxel spacing in the X, Y, and Z axes of the 3D data before resampling. This indicates the size of the 3D data after resampling. This represents the voxel spacing of the 3D data after resampling.

6. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, Step 5 specifically includes the following process: performing three-dimensional DWT transformation based on the db2 wavelet basis on the two segments of MRI human data to be stitched; separating a low-frequency component and high-frequency components in seven other directions from the three-dimensional data.

7. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, In step 6, the low-frequency components are normalized using the minimum-maximum normalization method, as shown in formulas 4 and 5: Among them, low freq and high freq Let min(low) represent the low-frequency and high-frequency components of the image separated from the original volume data, respectively. freq ) and max(low freq The numbers ) represent the minimum and maximum values ​​of the low-frequency component, respectively, and mean(high) represents the minimum and maximum values ​​of the low-frequency component. freq ) and max(abs(high freq )) represent the mean and maximum values ​​of the high-frequency components, respectively.

8. The method for stitching and fusing segmented MRI images of the human body according to claim 1, characterized in that, In step 7, the normalized low-frequency and high-frequency components are trained using neural networks based on 3DUnet. The last convolutional layer of the neural network corresponding to the low-frequency component uses the sigmoid function as the activation function, and the last convolutional layer of the neural network corresponding to the high-frequency component uses the hyperbolic tangent function as the activation function.

Images