A method for lumbosacral plexus nerve segmentation and three-dimensional visualization

Abstract

The present invention discloses a method for lumbar plexus nerve segmentation and three-dimensional visualization, comprising the following steps: (1) Data processing; collecting the MRI imaging data of a patient and storing it in DICOM format, converting the obtained data into a grayscale image, storing it in.png format, and performing normalization processing on the image; (2) Using a deep neural network to perform segmentation processing on the normalized image obtained in step (1); (3) Three-dimensional reconstruction; stacking the segmented two-dimensional plane images, and given the original voxel space, using the VTK package to perform three-dimensional reconstruction by the marching cubes algorithm, and using isosurface construction to display the internal details of the reconstructed object to the user. The present invention can intuitively display the relative position of the three-dimensional anatomical structure of the nerve in the spine, and can also assist in the clinical diagnosis of diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional visualization of medical images, and particularly to a method for lumbosacral plexus nerve segmentation and three-dimensional visualization. Background Art

[0002] With the development of the field of artificial intelligence, the application of artificial intelligence methods in the medical field has become increasingly widespread. At the same time, the cross - penetration of computer algorithms and medicine has penetrated the entire medical field, and there are numerous systems developed using artificial intelligence. During the process of clinical diagnosis and surgery, medical safety is very important. Only the development combined with clinical applications is a valuable artificial intelligence product. Currently, most spine - related surgeries are performed in a lying position. While it provides convenience for the surgery, it also increases the difficulty of the surgery. For patients with spinal stenosis, doctors usually perform spinal canal decompression and internal fixation surgery to reduce nerve compression and thus relieve the patient's symptoms. For patients with vertebral compression fractures, doctors usually perform percutaneous vertebroplasty to enhance the hardness of the vertebrae and prevent secondary fractures. For severe patients, doctors will also improve the fractured vertebrae by applying internal fixation screws. All of the above surgeries have a common feature that their surgical processes are also carried out through posterior approach. This surgical method will largely touch the nerves in the spinal canal, and in severe cases, it will cause problems such as lower limb paralysis and urinary and fecal incontinence in patients, which not only brings unacceptable facts to the patients but also leads to medical disputes. Therefore, how to use a visualization technology to enable doctors to have a sufficiently clear understanding of the relative positions of nerves and other tissues before surgery is particularly important, especially for the advance planning of some rare cases, which can greatly improve the doctor's surgical efficiency and reduce the risk of surgical failure. For most current spine surgeons, for such patients, based on their existing experience and combined with single - sided X - rays, CT, and MRI to determine the location of the lesion and then perform the surgery. Such planar imaging methods are difficult to fully display their relative positions. Even during the operation, only the C - arm machine is used for intraoperative fluoroscopy to present, but this fluoroscopy method can not only not present relevant soft tissues such as nerves, but also greatly increase the radiation dose of doctors and patients due to multiple fluoroscopies. In addition, performing surgery solely based on experience will inevitably have a certain degree of contingency, greatly increasing the risk of surgery. With the application of deep learning in the medical field becoming a key tool for medical image segmentation. In recent years, there has been a renewed interest in using MR images for spine segmentation. Many studies have focused on the localization, identification, and segmentation of vertebrae. However, only a few studies have focused on lumbosacral plexus nerve segmentation. Lumbosacral plexus nerve segmentation is a key step in finding abnormal structures and possible pathogenic factors of lumbosacral plexus nerve diseases. However, segmenting the lumbosacral plexus from MR images is still a challenging task. Usually, there is low contrast around the lumbosacral plexus. Secondly, the structure of the lumbosacral nerve plexus is extremely complex, and different patients have different nerve morphologies. In addition, manually marking the lumbosacral plexus from MR images is not only time - consuming but also prone to errors even for surgeons. Therefore, in clinical spine surgeries, there is an urgent need for a method that can perform preoperative three - dimensional visualization to reduce the surgical risk and thus increase the probability of surgical success. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for lumbosacral plexus nerve segmentation and three-dimensional visualization, which can intuitively display the relative position of the three-dimensional anatomical structure of the nerve in the spine and can also assist in clinical disease diagnosis.

[0004] To solve the above technical problem, the present invention provides a method for lumbosacral plexus nerve segmentation and three-dimensional visualization, including the following steps:

[0005] (1) Data processing; collecting the MRI imaging data of the patient and storing it in DICOM format, converting the obtained data into a grayscale image, storing it in.png format, and performing normalization processing on the image;

[0006] (2) Using a deep neural network to perform segmentation processing on the normalized image obtained in step (1);

[0007] (3) Three-dimensional reconstruction; stacking the segmented two-dimensional plane images and given the original voxel space, using the marching cubes algorithm in the VTK package for three-dimensional reconstruction, and using isosurface construction to display the internal details of the reconstructed object to the user.

[0008] Preferably, in step (2), the deep neural network includes an encoder, a decoder, a spatial domain convolutional encoding module, a residual skip connection module, and a scale attention module; the preprocessed neuroimage first enters the encoder, and the encoder contains four spatial domain convolutional encoding modules of different sizes. The input image obtains feature maps of different scales. Subsequently, the feature map of the last layer enters the decoder, and the feature maps of different scales in the encoder are respectively fused with the corresponding scale feature maps in the decoder. Finally, the decoders of different scales are processed by bilinear interpolation to obtain feature maps with the same dimension as the input image, and the concatenated feature maps after cascading them are processed by the scale attention module and the residual skip connection module to obtain the final segmentation image.

[0009] Preferably, in step (2), each porous encoder module first passes through a convolutional layer with a kernel size of 3×3 to obtain a feature map, and the feature map uses different convolution rates and convolution kernels to calculate the receptive field, and the receptive field is defined as follows:

[0010]

[0011] where r n represents the receptive field of the current layer, r n-1 represents the upper layer of the receptive field, s i represents the stride of the i-th convolution, and k represents the size of the convolution kernel;

[0012] The spatial domain convolution function F(x) has 4 cascaded branches with different convolution rates and convolution kernels, including convolution rates from 1 to 1, 3, and 5, and the convolution kernels include two types: 3×3 and 1×1;

[0013] [r3:r7:r9:r 19 = F(x)

[0014] where [:] represents concatenation;

[0015] Then, the receptive field of each branch will be set to 3, 7, 9, 19, and the concatenated feature maps [r3:r7:r9:r 19 are obtained. The concatenated feature maps are applied with 1×1 convolution and a linear activation function; finally, the input feature map is added to the previous feature map to obtain a new feature map at a new resolution level, and then passed through a convolutional layer with a kernel size of 3×3 and a pooling layer with a kernel size of 2×2 to obtain the final output. By introducing multiple convolutional layers with different convolution rates at four different resolution levels, the receptive field of the feature map can be well expanded, the parameters of the feature map at each resolution level are greatly reduced, and more semantic information is learned in the lumbosacral plexus nerve MRI images.

[0016] Preferably, in step (2), in the residual skip connection module, feature maps of different dimensions are mapped to a convolutional layer with 3×3 kernels, a batch normalization layer, and a ReLU as the activation layer; at the same time, the feature maps of different dimensions also become a convolutional layer with 3 three-kernel and a layer; the feature map H(x) and the original feature map X are added together in a shortcut manner to form a new feature map, which is put into the ReLU activation function to obtain a new feature map β:

[0017] β = ReLU(H(X) + E(X))

[0018] where ReLU() represents the activation function;

[0019] In addition, a 1×1 convolutional layer is added in the shortcut, which can provide some additional spatial features E(X).

[0020] Preferably, in step (2), in the scale attention module, bilinear interpolation is used to resample the feature maps of different scales obtained by the decoder to the size of the predicted image, and the feature maps of three different scales are concatenated as the input After concatenation represents a feature map with an input size of C×H×W, where C represents the input channel, and H and W represent the height and width of the feature map respectively. The input passes through a global average pooling GAP layer and a global max pooling GMP layer respectively, aiming to simplify the feature map parameters and obtain the weight information of each channel. The outputs are represented as P GAP (X) ∈ R 1×1×C and P GMP (X) ∈ R 1 ×1×C, the multi-layer perceptron MLP is implemented by two fully connected layers to obtain the scale attention coefficient and share it between X1 and X2. The scale attention coefficient α ∈ [0,1] is:

[0021] α = Sigmoid(X1 + X2)

[0022] In the scale attention module, a spatial attention block is additionally used to obtain spatial information. It consists of a 3×3 convolutional layer and a 1×1 convolutional layer, and then a Dropout layer is applied to enhance the generalization ability. The output feature map is multiplied by α and passed through the Sigmoid function to obtain a new feature map. This feature map is added to the original feature map in the spatial attention module to obtain the output feature map γ. Finally, the feature map γ is added to the original input feature map to obtain the output of the multi-scale attention block as:

[0023]

[0024] The feature map after scale attention finally passes through a 1D convolution with a kernel size of 1×1 and the Sigmoid function to obtain the final output segmentation image with the same shape as the input image.

[0025] Preferably, in step (3), the segmented two-dimensional plane images are stacked, and the original voxel space is given. The VTK package is used to perform three-dimensional reconstruction using the marching cubes algorithm. Using isosurface construction, the internal details of the reconstructed object are displayed to the user specifically as follows: First, all the imaging sequence data of this case number are read in a layered manner, two layers of data are scanned, and voxels are constructed one by one. The 8 corner points in the voxel are taken from the four corresponding pixels on the upper and lower two slices. These two slices form a cube. The definition of the isosurface is as follows:

[0026] F i,j,k = F i,j,k (x i ,y j ,z k )

[0027] {(x,y,z)|F(x,y,z) = N}

[0028] where F is the value of each pixel, x i ,y j ,z kLet \((x,y,z)\) be the coordinates of a pixel point. The set of points that satisfy the above formula is the isosurface. If the function value \(F\) of a vertex is \(F\geq N\), then the vertex is inside the isosurface and we mark it as \(+\); if the function value \(F\) of the vertex is \(F < N\), then the vertex is outside the isosurface and is marked as \(-\). Then, construct the state table of the voxel based on the result of comparing the corner function values of the voxel with the isosurface. From the obtained state table, the boundary voxels that intersect with the isosurface can be obtained. Then, by using the method of linear interpolation, calculate the intersection position coordinates \((x,y,z)\) of the isosurface and the cube:

[0029]

[0030]

[0031]

[0032] In order to use the graphics hardware to display the isosurface image and generate the normal components of each triangular patch of the isosurface. However, for each point on the isosurface, the gradient component of the tangent along the triangular surface direction is zero. Then, the gradient vector at that point represents the normal vector of the isosurface at that point. Assume the voxel vertex is \((i,j,k)\), so the representation of the gradient value is:

[0033]

[0034]

[0035]

[0036] where \(G\) x 、\(G\) y 、\(G\) z respectively represent the gradient values of the voxel vertex in the \(x\), \(y\), and \(z\) directions;

[0037] Then, normalize the obtained \(G\) x 、\(G\) y 、\(G\) z :

[0038]

[0039] The result after normalization is used as the unit normal vector of the voxel vertex \((i,j,k)\). Then, according to the linear interpolation function, the normal vectors of each vertex of the triangular patch can be calculated. Finally, based on the coordinates and normal vectors of each vertex on the triangular patch, render the isosurface image and use the lighting renderer for rendering to obtain the final surface model and realize the three-dimensional rendering of the final image.

[0040] The beneficial effects of the present invention are as follows: The segmentation algorithm of the present invention can more quickly and effectively segment the lumbosacral plexus nerve from spinal MRI medical images, and the tissue segmentation accuracy is higher, so the accuracy of the three-dimensional structure is also greatly increased; This lumbosacral plexus nerve segmentation and three-dimensional visualization algorithm can enable doctors to have a sufficiently clear understanding of the relative position of the lumbosacral plexus nerve before surgery, especially for the advance planning of some rare cases, which can greatly improve the surgical efficiency of doctors and reduce the risk of surgical failure; The present invention can not only intuitively display the relative position of the three-dimensional anatomical structure of the nerve in the spine, but also assist in clinical disease diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the overall framework diagram for implementing the method of the present invention.

[0042] Figure 2 It is the structural schematic diagram of the spatial domain convolutional coding module of the present invention.

[0043] Figure 3 It is the structural schematic diagram of the residual skip connection of the present invention.

[0044] Figure 4 It is the structural schematic diagram of the scale attention module of the present invention.

[0045] Figure 5 It is the schematic diagram of the three-dimensional reconstruction result of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0046] A method for lumbosacral plexus nerve segmentation and three-dimensional visualization includes the following steps:

[0047] (1) Data processing; Collect the MRI imaging data of the patient and store it in DICOM format, convert the obtained data into a grayscale image, store it in.png format, and perform normalization processing on the image;

[0048] (2) Use a deep neural network to perform segmentation processing on the normalized image obtained in step (1);

[0049] (3) Three-dimensional reconstruction; Stack the segmented two-dimensional plane images, and given the original voxel space, use the marching cubes algorithm in the VTK package for three-dimensional reconstruction, and use isosurface construction to display the internal details of the reconstructed object to the user.

[0050] In the data processing section, first, the MRI imaging data of the patient is collected and stored in DICOM format. The obtained data is converted into a grayscale image and stored in.png format, and the image is normalized for use in subsequent processes. In the segmentation section, we adopt the popular deep neural network in the field of artificial intelligence for the segmentation method. In this neural network, we use the classic U-net biomedical image segmentation network as the basic framework and combine it with a spatial domain convolutional coding module, a residual jump link module, and a scale attention module to complete the lumbosacral plexus nerve segmentation. The U-shaped structure segmentation network is an encoder-decoder network that can extract features from the original image and restore it to the original image size. Spatial domain convolution is essential for large-scale semantic segmentation in the field of image processing and has achieved effective improvements in feature extraction. Therefore, in this algorithm, we introduce spatial domain convolution into the encoder. However, due to consecutive pooling between lumbosacral nerve plexus segmentations, a large amount of segmentation information may be lost at different feature levels. In addition, for lumbosacral plexus nerve MR images, the nerves only account for a very small part of the image. To overcome this limitation, we use dilated convolutions with different dilation rates in the feature encoder to expand the receptive fields of the feature maps at four different resolution levels, as Figure 1 shown. Each porous encoder module first passes through a convolutional layer with a kernel size of 3×3 to obtain the feature map X. The feature map uses different dilation rates and convolutional kernels to calculate the receptive fields. Specifically, the receptive field is defined as follows:

[0051]

[0052] where, r n represents the receptive field of the current layer, r n-1 represents the upper layer of the receptive field, s i represents the stride of the i-th convolution, and k represents the size of the convolutional kernel.

[0053] The spatial domain convolution function F(x) has four cascaded branches with different dilation rates and convolutional kernels, including dilation rates of 1 to 1, 3, 5, and convolutional kernels of 3×3 and 1×1.

[0054] [r3:r7:r9:r 19 = F(x)

[0055] where [:] represents concatenation.

[0056] Then, the receptive fields of each branch will be set to 3, 7, 9, 19, and the concatenated feature map [r3:r7:r9:r 19, the cascaded feature maps apply 1×1 convolution and linear activation functions. Finally, we add the input feature map to the previous feature map to obtain a new feature map at a new resolution level, which then passes through a convolutional layer with a kernel size of 3×3 and a pooling layer with a kernel size of 2×2 to get the final output. By introducing multiple convolutional layers with different convolution rates at four different resolution levels, the receptive field of the feature map can be well expanded, the parameters of the feature map at each resolution level are greatly reduced, and more semantic information is learned in the lumbosacral plexus nerve MRI images.

[0057] Residual networks can, to a certain extent, solve the problem of network degradation and alleviate the problem of gradient dispersion to some extent. The skip connections used in residual networks are aimed at increasing feature diversity and accelerating training. Residual skip connections can introduce feature information at corresponding scales into upsampling or transposed convolution to improve the accuracy of segmentation. Therefore, in our method, we introduce residual skip connection operations between the encoder and the decoder to integrate the spatial features lost in the pooling operation in the encoder and the high-level feature maps in the decoder. Residual skip connections, in addition, can also recover detailed information from the original image and can provide multi-scale and multi-level information for image segmentation. The architecture of the residual skip connection is as Figure 2 shown. Feature maps of different dimensions are mapped to a convolutional layer with 3×3 kernels, a batch normalization layer, and a ReLU as the activation layer. At the same time, the feature maps of different dimensions also become a convolutional layer with 3 3×3 kernels and a layer. The feature map H(x) and the original feature map X are added together through shortcut connection to form a new feature map, which is put into the ReLU activation function to obtain a new feature map β:

[0058] β = ReLU(H(X) + E(X))

[0059] In addition, they also added a 1×1 convolutional layer in the shortcut connection, and E(X) can provide some additional spatial features.

[0060] To achieve the best segmentation effect, in computer vision models, a scale attention model is usually introduced. In our method, we introduce a scale attention module to better process the lumbosacral plexus nerve to obtain segmentation results at different scales. The multi-scale attention block we proposed is as Figure 3 shown. We first use bilinear interpolation to resample the feature maps at different scales obtained by the decoder to the size of the predicted image. We first concatenate three feature maps at different scales as the input After concatenation represents a feature map with an input size of C×H×W, where C represents the input channels, and H and W represent the height and width of the feature map respectively. The input Through the global average pooling (GAP) layer and the global max pooling (GMP) layer respectively. The purpose is to simplify the feature map parameters and obtain the weight information of each channel, and the outputs are denoted as P GAP (X) ∈ R 1 ×1×* and P GMP (X) ∈ R 1×1×C . The multi-layer perceptron (MLP) is implemented by two fully connected layers to obtain the scale attention coefficient and share it between X1 and X2. The scale attention coefficient α ∈ [0, 1] is:

[0061] α = Sigmoid(X1 + X2)

[0062] In the scale attention block, we additionally use the spatial attention block to obtain spatial information. It consists of a 3×3 convolutional layer and a 1×1 convolutional layer, and then a Dropout layer is applied to enhance the generalization ability. The output feature map is multiplied by α and passed through the Sigmoid function to obtain a new feature map. This feature map is added to the original feature map in the spatial attention module to obtain the output feature map γ. Finally, the feature map γ is added to the original input feature map to obtain the output of the multi-scale attention block which is:

[0063]

[0064] The feature map after scale attention finally passes through a 1D convolution with a kernel size of 1×1 and the Sigmoid function to obtain the final output segmentation image with the same shape as the input image.

[0065] The 3D technology can give a visual impact and enable people to more intuitively perceive objective objects. In this system, the segmented 2D plane images are stacked, and the original voxel space is given. The marching cubes algorithm is used for 3D reconstruction by using the VTK package. This method uses isosurface construction to display the internal details of the reconstructed object to the user, enabling doctors to more intuitively observe the patient's nerve condition. The basic idea of this algorithm is to process each voxel in the data field, classify the voxel elements where the isosurface intersects, and then use the linear interpolation method to calculate the intersection points and normal vectors of the isosurface. Finally, the isosurface is drawn using relevant graphics software. We first read all the imaging sequence data of this case number in a hierarchical manner, scan two layers of data, and construct voxels one by one. The 8 corner points in the voxel are taken from the four corresponding pixels on the upper and lower slices of two slices. These two slices form a cube, and the definition of the isosurface is as follows:

[0066] F i,j,k = F i,j,k (x i ,y j ,zk )

[0067] {(x, y, z)|F(x, y, z) = N}

[0068] where F is the value of each pixel, x i , y j , z k are the coordinates of the pixel point. The set of points satisfying the above formula is the isosurface. If the function value F of the vertex is F ≥ N, then the vertex is inside the isosurface and we mark it as +; if the function value F of the vertex is F < N, then the vertex is outside the isosurface and is marked as -. Then, the state table of the voxel is constructed based on the result of comparing the corner function values of the voxel with the isosurface. From the obtained state table, the boundary voxels intersecting with the isosurface can be obtained, and then the intersection position coordinates (x, y, z) of the isosurface and the cube can be calculated by linear interpolation:

[0069]

[0070]

[0071]

[0072] To display our isosurface image using the graphics hardware, we must generate the normal components of each triangular patch of the isosurface. However, for each point on the isosurface, the gradient component of the tangent along the triangular surface direction is zero, so the gradient vector at that point represents the normal vector of the isosurface at that point. Assuming the voxel vertex is (i, j, k), the representation of the gradient value is:

[0073]

[0074]

[0075]

[0076] where G x , G y , G z represent the gradient values of the voxel vertex in the x, y, and z directions respectively;

[0077] Then, the obtained G x , G y , G z are normalized:

[0078]

[0079] The normalized result serves as the unit normal vector of the voxel vertex (i, j, k). Then, according to the linear interpolation function, the normal vectors of each vertex of the triangular patch can be calculated. Finally, based on the coordinates and normal vectors of each vertex on the triangular patch, the isosurface image is drawn and rendered using a lighting renderer to obtain the final surface model and achieve the 3D rendering of the final image, as Figure 5 shown. This further facilitates the doctor's preoperative judgment and enables the patient to better understand their lesion site.

Claims

1. A method for lumbosacral plexus nerve segmentation and three-dimensional visualization, characterized in that The steps include: (1) Data processing: Collect the patient's MRI imaging data and store it in DICOM format, convert the obtained data into grayscale images, store them in .png format, and normalize the images; (2) using a deep neural network to segment the normalized image obtained in step (1); the deep neural network includes an encoder, a decoder, a spatial convolutional coding module, a residual skip link module, and a scale attention module; The preprocessed neural image first enters the encoder, which contains four spatial convolutional coding modules of different sizes. The input image obtains feature maps of different scales. Then the feature maps of the last layer enter the decoder. The feature maps of different scales in the encoder are fused with the feature maps of the corresponding scales in the decoder. Finally, the decoders of different scales undergo bilinear interpolation to obtain feature maps of the same dimension as the input image. The cascaded feature maps are processed by the scale attention module and the residual skip link module to obtain the final segmented image. Each porous encoder module first passes through a convolutional layer with a kernel size of 3×3 to obtain a feature map X. The feature map uses different convolution rates and convolution kernels to calculate the receptive field. The receptive field is defined as follows: Among them, r n represents the receptive field of the current layer, and r n-1 represents the upper layer of the receptive field. s i represents the stride of the i-th convolution, and k represents the size of the convolution kernel; The spatial convolution function F(x) has four cascade branches with different convolution rates and convolution kernels, including convolution rates of 1 to 1, 3, and 5, and convolution kernels of 3×3 and 1×1. [r3:r7:r9:r 19 = F(x) Where [:] represents cascade; Then, the receptive fields of each branch will be set to 3, 7, 9, 19, and the cascaded feature maps [r3:r7:r9:r 19 will be obtained. The cascaded feature maps are applied with 1×1 convolution and a linear activation function; finally, the input feature map is added to the previous feature map to obtain a new resolution-level feature map, which is then passed through a convolutional layer with a kernel size of 3×3 and a pooling layer with a kernel size of 2×2 to obtain the final output; (3) Three-dimensional reconstruction: The segmented two-dimensional plane images are stacked and the original voxel space is given. The VTK package is used to use the marching cube algorithm to perform three-dimensional reconstruction. The internal details of the reconstructed object are displayed to the user using the isosurface construction.

2. The method for lumbosacral plexus nerve segmentation and three-dimensional visualization according to claim 1, wherein In step (2), in the residual skip link module, the features of different dimensions are mapped to a convolutional layer with 3×3 kernels, a batch normalization layer and a ReLU as an activation layer; at the same time, the feature maps of different dimensions are also transformed into a convolutional layer with 3 kernels and a layer; the feature map H(x) and the original feature map X are added by short-circuiting to form a new feature map, which is put into the ReLU activation function to obtain the new feature map β: β=ReLU(H(X)+E(X)) Where ReLU() represents the activation function; In addition, a 1×1 convolutional layer is added in the short circuit to provide some additional spatial features E(X).

3. The lumbosacral plexus nerve segmentation and three-dimensional visualization method according to claim 1, wherein In step (2), in the scale attention module, the feature maps of different scales obtained by the decoder are resampled to the size of the predicted image using bilinear interpolation, and the feature maps of three different scales are concatenated as the input After concatenation represents a feature map with an input size of C×H×W, where C represents the input channels, and H and W represent the height and width of the feature map, respectively. The input passes through the global average pooling GAP layer and the global maximum pooling GMP layer respectively, aiming to simplify the feature map parameters and obtain the weight information of each channel. The outputs are denoted as P GAP (X)∈R 1×1×c and P GMP (X)∈R 1×1×C , and the multi-layer perceptron MLP is implemented by two fully connected layers to obtain the scale attention coefficient and share it between X1 and X2. The scale attention coefficient α∈[0,1] is as follows: α=Sigmoid(X1+X2) In the scale attention module, a spatial attention block is additionally used to obtain spatial information. It consists of a 3×3 convolutional layer and a 1×1 convolutional layer. Then, a Dropout layer is applied to enhance the generalization ability. The output feature map is multiplied by α and passed through the Sigmoid function to obtain a new feature map. This feature map is added to the original feature map in the spatial attention module to obtain the output feature map γ. Finally, the feature map γ is added to the original input feature map to obtain the output of the multi-scale attention block The result is: The feature map after scale attention is finally passed through a one-dimensional convolution with a kernel size of 1×1 and a Sigmoid function to obtain an output segmentation image with the same shape as the input image.

4. The method for lumbosacral plexus nerve segmentation and three-dimensional visualization according to claim 1, wherein In step (3), the segmented two-dimensional planar images are stacked, and the original voxel space is given. The marching cubes algorithm is used for 3D reconstruction by means of the VTK package. The internal details of the reconstructed object are displayed to the user by means of isosurface construction, specifically as follows: First, all the imaging sequence data of the patient are read in a layered manner. Two layers of data are scanned, and voxels are constructed one by one. The 8 corner points in the voxel are taken from the four corresponding pixels on the upper and lower slices of two slices. These two slices form a cube. The definition of the isosurface is as follows: F i,j,k = F i,j,k (x i , y j , z k ) {(x,y,z)|F(x,y,z) = N} where F is the value of each pixel, x i , y j , z k are the coordinates of the pixel points. The set of points satisfying the above formula is the isosurface. If the function value F of a vertex is ≥ N, then the vertex is inside the isosurface and we mark it as +; if the function value F of the vertex is < N, then the vertex is outside the isosurface and is marked as -. Then, the state table of the voxel is constructed from the result of comparing the corner function values of the voxel with the isosurface. From the obtained state table, the boundary voxels intersecting with the isosurface are obtained. Then, the position coordinates (x, y, z) of the intersection points of the isosurface and the cube are calculated by the method of linear interpolation: In order to display the isosurface image using the graphics hardware and generate the normal components of each triangular patch of the isosurface, however, for each point on the isosurface, the gradient component of the tangent along the triangular face direction is zero. Then the gradient vector at this point represents the normal vector of the isosurface at this point. Assuming that the voxel vertex is (i,j,k), the representation of the gradient value is as follows: Among them, G x , G y , G z respectively represent the gradient values of the voxel vertices in the x, y, and z directions; Then, the obtained G x , G y , G z are normalized as follows: The normalized result is used as the unit normal vector of the voxel vertex (i,j,k). Then, the normal vectors at each vertex of the triangular patch are calculated according to the linear interpolation function. Finally, the isosurface image is drawn based on the coordinates and normal vectors of each vertex on the triangular patch, and is rendered using a lighting renderer to obtain the final surface model to achieve the 3D rendering of the final image.

Images