CN114519770A - Lumbosacral plexus nerve segmentation and three-dimensional visualization method - Google Patents

Abstract

The invention discloses a lumbosacral plexus nerve segmentation and three-dimensional visualization method, which comprises the following steps: (1) processing data; acquiring MRI (magnetic resonance imaging) image data of a patient, storing the MRI image data in a DICOM (digital imaging and communications in medicine) format, converting the acquired data into a gray image, storing the gray image in a png format, and performing normalization processing on the image; (2) adopting a deep neural network to perform segmentation processing on the normalized image obtained in the step (1); (3) three-dimensional reconstruction; stacking the divided two-dimensional plane images, giving an original voxel space, performing three-dimensional reconstruction by using a VTK package and adopting a moving cube algorithm, and displaying the internal details of a reconstructed object to a user by using an isosurface structure. The invention can visually display the relative position of the three-dimensional anatomical structure of the nerve in the spine and can assist in clinical disease diagnosis.

Description

Lumbosacral plexus nerve segmentation and three-dimensional visualization method

Technical Field

The invention relates to the technical field of three-dimensional visualization of medical images, in particular to a lumbosacral plexus nerve segmentation and three-dimensional visualization method.

Background

With the development of the field of artificial intelligence, the application of the artificial intelligence method in the medical field is gradually widened. Meanwhile, the computer algorithm and the medicine are crossed and penetrated into the whole medical field, and the number of the system developed by utilizing the artificial intelligence is not as large as possible. Medical safety is very important in the process of clinical diagnosis and surgery, and only in combination with the development of clinical application is a valuable artificial intelligence product. Most of the current spinal related procedures are performed in a recumbent position. The operation difficulty is increased while the convenience is provided for the operation. For patients with spinal stenosis, physicians often perform spinal decompression internal fixation procedures to reduce nerve compression and thereby relieve the patient's symptoms. For patients with compression fracture of the vertebral body, a doctor usually performs percutaneous vertebroplasty to enhance the hardness of the vertebral body and prevent secondary fracture, and for serious patients, the doctor improves the fractured vertebral body by applying internal fixing screws. The above operations have a common characteristic, the operation process is also performed through posterior approach, the operation mode can reach nerves in the vertebral canal to a great extent, the problems of lower limb paralysis, incontinence of urine and stool and the like of a patient can be caused under severe conditions, the fact that the patient is difficult to accept is brought, and medical disputes can also be caused. Therefore, how to utilize a visualization technology is important to enable a doctor to have a sufficiently clear understanding of the relative position of the nerve and other tissues before an operation, and especially for the advanced planning of some rare cases, the operation efficiency of the doctor can be greatly improved, and the risk of operation failure can be reduced. For most spine surgeons, the position of the lesion is determined by combining unilateral X-ray film, CT and MRI on an experienced basis for the patients to operate, and the relative position between the patients is difficult to completely show by the plane imaging mode. Even though the X-ray fluoroscopy is only performed by using a C-arm machine during the operation, the X-ray fluoroscopy cannot present related soft tissues such as nerves, and the radiation quantity of doctors and patients is greatly increased due to multiple X-ray fluoroscopy. In addition, the operation is inevitably performed by experience only, so that certain contingency exists, and the operation risk is greatly increased. The application of deep learning in the medical field becomes a key tool for medical image segmentation. In recent years, interest in spine segmentation using MR images has been renewed. Much research has focused on the location, identification, and segmentation of vertebral bodies. However, only a few studies have focused on the segmentation of the lumbosacral plexus. The segmentation of the lumbosacral plexus nerve is a key step for searching abnormal structures and possible pathogenic factors of the lumbosacral plexus nerve diseases. However, segmenting the lumbosacral plexus from MR images remains a challenging task. There is generally low contrast around the lumbosacral plexus, and secondly, the lumbosacral nerve plexus is extremely complex in structure, with different patient neuromorphies. Furthermore, manually marking the lumbosacral plexus from the MR image is not only time consuming, but is also prone to error even for the surgeon. Therefore, there is a need for a method capable of performing preoperative three-dimensional visualization to reduce the risk of surgery, thereby improving the success rate of surgery in clinical spinal surgery.

Disclosure of Invention

The invention aims to provide a lumbosacral plexus nerve segmentation and three-dimensional visualization method, which can intuitively display the relative position of a three-dimensional anatomical structure of a nerve in a spine and can assist in clinical disease diagnosis.

In order to solve the technical problem, the invention provides a lumbosacral plexus nerve segmentation and three-dimensional visualization method, which comprises the following steps:

(1) processing data; acquiring MRI (magnetic resonance imaging) image data of a patient, storing the MRI image data in a DICOM (digital imaging and communications in medicine) format, converting the acquired data into a gray image, storing the gray image in a png format, and performing normalization processing on the image;

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

(3) three-dimensional reconstruction; stacking the divided two-dimensional plane images, giving an original voxel space, performing three-dimensional reconstruction by using a VTK package and adopting a moving cube algorithm, and displaying the internal details of a reconstructed object to a user by using an isosurface structure.

Preferably, in the step (2), the deep neural network comprises an encoder, a decoder, a spatial convolutional coding module, a residual jump linking module and a scale attention module; the preprocessed neural image firstly enters an encoder, the encoder comprises four spatial domain convolutional coding modules with different sizes, an input image obtains feature maps with different scales, then a last layer of feature map enters a decoder, the feature maps with different scales in the encoder are respectively fused with the feature maps with corresponding scales in the decoder, finally the decoder with different scales obtains the feature map with the same dimension as the input image through bilinear interpolation processing, and the feature maps after the cascade connection are processed through a scale attention module and a residual jump link module to obtain a final segmentation image.

Preferably, in step (2), each multi-aperture encoder module first obtains a feature map through a convolution layer with kernel size of 3 × 3, the feature map uses different convolution rates and convolution kernels to calculate the receptive field, which is defined as follows:

wherein r is_nRepresenting the receptive field of the current layer, r_n-1Upper layer, s, representing the receptive field_iDenotes the step size of the i-th convolution, k denotes the size of the convolution kernel;

the spatial convolution function F (x) has 4 cascade branches with different convolution rates and convolution kernels, wherein the convolution rates are 1 to 1, 3 and 5, and the convolution kernels comprise 3 x 3 and 1 x 1;

[r₃:r₇:r₉:r₁₉]＝F(x)

wherein [: represents cascade;

then, the receptive fields for each branch would be set to 3,7,9,19 and a cascade profile r is obtained₃:r₇:r₉:r₁₉]The cascade signature applies a 1 × 1 convolution and a linear activation function; finally, the input feature map is added with the previous feature map to obtain a new resolution level feature map, and the final output is obtained through a convolution layer with the kernel size of 3 × 3 and a pooling layer with the kernel size of 2 × 2. By introducing a plurality of convolution layers with different convolution rates into 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 MRI image.

Preferably, in the step (2), in the residual jump linking module, the features with different dimensions are mapped to a convolution layer with 3 × 3 kernels, a batch normalization layer and a ReLU as an active layer; meanwhile, feature mapping of different dimensions also becomes a convolution layer with 3 cores and a layer; adding the characteristic diagram H (X) and the original characteristic diagram X in a short circuit mode to form a new characteristic diagram, and putting the new characteristic diagram into a ReLU activation function to obtain a new characteristic diagram beta:

β＝ReLU(H(X)+E(X))

wherein ReLU () represents an activation function;

in addition, a 1 × 1 convolutional layer is added in the short circuit, which can provide some additional spatial features e (x).

Preferably, in the step (2), in the scale attention module, the bilinear interpolation method is used to resample the feature maps of different scales obtained by the decoder to the size of the predicted image, and the three feature maps of different scales are concatenated as input

After cascade connection

Representing a feature map with an input size of C H W, where C represents the input channel, H and W represent the height and width, respectively, of the feature map, and the input

Respectively passing through a global average pooling GAP layer and a global maximum pooling GMP layer, in order to simplify the characteristic diagram parameters and obtain the weight information of each channel, and outputting P_GAP(X)∈R^1×1×CAnd P_GMP(X)∈R^{1 ×1×C}Multilayer perceptual MLP is implemented with two fully connected layers to obtain a scale attention coefficient and X₁And X₂Are shared between them. The scale attention coefficient alpha is belonged to [0,1 ]]Comprises the following steps:

α＝Sigmoid(X₁+X₂)

in the scale attention module, spatial attention is additionally usedObtaining space information by a block, wherein the space information comprises a 3 multiplied by 3 convolution layer and a 1 multiplied by 1 convolution layer, then applying a Dropout layer to enhance generalization capability, multiplying an output characteristic diagram by alpha to obtain a new characteristic diagram through a Sigmoid function, adding the characteristic diagram and an original characteristic diagram in a space attention module to obtain an output characteristic diagram gamma, and finally adding the characteristic diagram gamma and an original input characteristic diagram to obtain the output characteristic diagram of the multi-scale attention block

Comprises the following steps:

finally, the feature map subjected to scale attention is subjected to a one-dimensional convolution with the kernel size of 1 × 1 and a Sigmoid function to obtain an output segmentation image which has the same shape as the input image.

Preferably, in step (3), stacking the segmented two-dimensional plane images, giving an original voxel space, performing three-dimensional reconstruction by using a VTK package and a marching cube algorithm, and displaying internal details of the reconstructed object to a user by using an isosurface structure specifically includes: firstly, reading all the iconography sequence data of the patient in a layered reading mode, scanning two layers of data, constructing voxels one by one, wherein 8 corner points in the voxels are taken from four pixel points which correspond to the upper part and the lower part of two slices, the two slices form a cube, and the isosurface is defined 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_kIf the function value F of the vertex is more than or equal to N, the vertex is positioned in the equivalent surface and is marked as plus; if the function value F of the vertex<N, then the vertex is outside the iso-surface, marked-and then the corner function value of the voxel is compared with the junction after the iso-surfaceIf the state table of the voxel is constructed, the boundary voxel with the intersection point with the isosurface can be obtained from the obtained state table, and then the position coordinates (x, y, z) of the intersection point can be obtained by calculating the isosurface and the cube through a linear interpolation method:

to display the iso-surface image using graphics hardware, the normal components of the triangular patches of the iso-surface are generated, however for each point on the iso-surface, the gradient component of the tangent in the direction of the triangular surface is zero, then the gradient vector for that point represents the normal vector of the iso-surface at that point, assuming the voxel vertex is (i, j, k) and thus the gradient values are represented as:

wherein G is_x、G_y、G_zRespectively representing gradient values of voxel vertexes in x, y and z directions;

then the obtained G is_x、G_y、G_zCarrying out normalization treatment:

and finally, drawing an isosurface image according to each vertex coordinate and the normal vector on the triangular surface patch, and rendering by using a light renderer to obtain a final surface model so as to realize the three-dimensional drawing of the final image.

The beneficial effects of the invention are as follows: the segmentation algorithm can more quickly and effectively segment the lumbosacral plexus nerve from the spine MRI medical image, and the accuracy of the segmented tissue is higher, so that the accuracy of the three-dimensional structure is greatly increased; the lumbosacral plexus nerve segmentation and three-dimensional visualization algorithm can enable a doctor to have a sufficiently clear understanding of the relative position of the lumbosacral plexus before an operation, and especially can greatly improve the operation efficiency of the doctor and reduce the risk of operation failure for the advance planning of some rare cases; the invention not only can intuitively display the relative position of the three-dimensional anatomical structure of the nerve in the spine, but also can assist in clinical disease diagnosis.

Drawings

FIG. 1 is a general block diagram of a method implementation of the present invention.

FIG. 2 is a schematic structural diagram of a spatial convolutional coding module according to the present invention.

Fig. 3 is a schematic diagram of a residual jump connection structure according to the present invention.

Fig. 4 is a structural diagram of a scale attention module according to the present invention.

Fig. 5 is a schematic diagram of a three-dimensional reconstruction result according to the present invention.

Detailed Description

A lumbosacral plexus nerve segmentation and three-dimensional visualization method comprises the following steps:

(1) processing data; acquiring MRI (magnetic resonance imaging) image data of a patient, storing the MRI image data in a DICOM (digital imaging and communications in medicine) format, converting the acquired data into a gray image, storing the gray image in a png format, and performing normalization processing on the image;

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

(3) three-dimensional reconstruction; stacking the divided two-dimensional plane images, giving an original voxel space, performing three-dimensional reconstruction by using a VTK package and adopting a moving cube algorithm, and displaying the internal details of a reconstructed object to a user by using an isosurface structure.

In the data processing part, firstly, MRI image data of a patient is collected and stored in a DICOM format, the obtained data is converted into a gray image and stored in a png format, and the image is normalized for use in the subsequent process. In the segmentation part, a deep neural network popular in the field of artificial intelligence is adopted in the segmentation method, and in the neural network, a U-net classical biomedical image segmentation network is used as a basic frame and is combined with a spatial domain convolution coding module, a residual jump link module and a scale attention module to complete lumbosacral plexus nerve segmentation. The U-shaped structure segmentation network is a coding-decoding network, can perform feature extraction on an original image and restore the original image to the size. The spatial convolution is essential to the wide-range semantic segmentation in the image processing field, and the effective improvement is achieved in the aspect of feature extraction. So in this algorithm we introduce spatial convolution into the encoder. However, the different feature levels may lose a large amount of segmentation information due to the continuous pooling between lumbosacral nerve plexus segmentations. Furthermore, for a lumbosacral plexus MR image, nerves account for only a small portion of the image. To overcome this limitation, we use hole convolution at different convolution rates in the feature encoder to extend the receptive field of the feature map at four different resolution levels, as shown in FIG. 1. Each multi-aperture encoder module first obtains a feature map X over a convolutional layer of kernel size 3 × 3. The feature maps use different convolution rates and convolution kernels to calculate the receptive field. In particular, the receptive fields are defined as follows:

wherein,r_nrepresenting the receptive field of the current layer, r_n-1Upper layer, s, representing the receptive field_iDenotes the step size of the i-th convolution and k denotes the size of the convolution kernel.

The spatial convolution function f (x) has 4 cascaded branches of different convolution rates and convolution kernels, including convolution rates of 1 to 1, 3, 5, and convolution kernels including both 3 x 3 and 1 x 1.

[r₃:r₇:r₉:r₁₉]＝F(x)

Wherein [: represents cascade.

Then, the receptive field for each branch will be set to 3,7,9,19, and a cascade profile r is obtained₃:r₇:r₉:r₁₉]The cascade signature applies a 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, and the final output is obtained through a convolution layer with a kernel size of 3 × 3 and a pooling layer with a kernel size of 2 × 2. By introducing a plurality of convolution layers with different convolution rates into 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 MRI image.

The residual error network can solve the problem of network degradation to a certain extent and relieve the problem of gradient dispersion to a certain extent, and jump layer connection used in the residual error network achieves the purposes of increasing feature diversity and accelerating training. Residual skipping concatenations can introduce feature information at the corresponding scale into the upsampling or deconvolution to improve the accuracy of the segmentation. Thus, in our approach, we introduce a residual skip concatenation operation between the encoder and decoder to integrate the spatial features lost to pooling in the encoder and the high-level feature maps in the decoder. The residual skips the connection, besides, the detailed information can be recovered from the original image, and multi-scale and multi-level information can be provided for image segmentation. The architecture of the residual jump connection is shown in fig. 2. Features of different dimensions are mapped to a convolutional layer with 3 x 3 kernels, a bulk normalization layer, and a ReLU as the active layer. Meanwhile, feature mapping of different dimensions also becomes a convolution layer with 3 cores and a layer. Adding the characteristic diagram H (X) and the original characteristic diagram X in a short circuit mode to form a new characteristic diagram, and putting the new characteristic diagram into a ReLU activation function to obtain a new characteristic diagram beta:

β＝ReLU(H(X)+E(X))

in addition, they have added 1 × 1 convolutional layers in the shorts, and e (x) may provide some additional spatial features.

In order to achieve the best segmentation effect, a scale attention model is usually introduced in the computer vision model. In our approach, we introduce a scale attention module to better process the lumbosacral plexus to obtain segmentation results at different scales. The multi-scale attention block we propose is shown in fig. 3. We first resample the different scales of feature maps obtained by the decoder to the predicted image size using bilinear interpolation. We first concatenate three feature maps of different scales as input

After cascade connection

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. Input device

Respectively through a Global Average Pooling (GAP) layer and a Global Maximum Pooling (GMP) layer. The purpose is to simplify the characteristic diagram parameters and obtain the weight information of each channel, and the output is respectively expressed as P_GAP(X)∈R^{1 ×1×*}And P_GMP(X)∈R^1×1×C. Multi-layer perception (MLP) is implemented by two fully-connected layers to obtain the scale attention coefficient and at X₁And X₂Are shared between them. The scale attention coefficient alpha is belonged to [0,1 ]]Comprises the following steps:

α＝Sigmoid(X₁+X₂)

in the scale attention Block, IThey additionally use the spatial attention block to obtain spatial information. It consists of one 3 x 3 convolutional layer and one 1 x 1 convolutional layer, and then a Dropout layer is applied to enhance the generalization capability. And multiplying the output feature map by alpha to obtain a new feature map through a Sigmoid function. The feature map and the original feature map in the spatial attention module are added to obtain an output feature map gamma, and finally the feature map gamma is added with the original input feature map to finally obtain the output of the multi-scale attention block

Comprises the following steps:

finally, the feature map subjected to scale attention is subjected to a one-dimensional convolution with the kernel size of 1 × 1 and a Sigmoid function to obtain an output segmentation image which has the same shape as the input image.

The three-dimensional technology can give visual impact to people, people can feel objective objects more visually, segmented two-dimensional plane images are stacked in the system, an original voxel space is given, three-dimensional reconstruction is carried out by using a VTK package and a moving cube algorithm, the method utilizes an isosurface structure, internal details of the reconstructed object can be displayed like a user, and therefore doctors can observe the nerve condition of patients more visually. The basic idea of the algorithm is to process each voxel in a data field, classify the body elements intersected with the isosurface, calculate the intersection point and normal vector with the isosurface by using a linear difference method, and finally draw the isosurface by using related graphic software. Firstly, reading all the iconography sequence data of the patient in a layered reading mode, scanning two layers of data, and constructing voxels one by one, wherein 8 corner points in the voxels are taken from four pixel points which correspond to the upper part and the lower part of two slices. The two slices form a cube, and the iso-surface is defined 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_kIf the function value F of the vertex is more than or equal to N, the vertex is positioned in the equivalent surface and is marked as plus; if the function value F of the vertex<N, then the vertex is outside the iso-surface, labeled-and then the result of comparing the corner function value of the voxel with the iso-surface constructs the state table for that voxel. The boundary voxel with the intersection point with the isosurface can be obtained from the obtained state table, and then the intersection point position coordinate (x, y, z) can be obtained by calculating the isosurface and the cube through a linear interpolation method:

in order to display our iso-surface image using graphics hardware, we must generate the normal components of the triangular patches of the iso-surface, however, for each point on the iso-surface, the gradient component of the tangent in the direction of the triangular surface is zero, then the gradient vector for that point represents the normal vector of the iso-surface at that point, assuming the voxel vertex is (i, j, k) and thus the gradient values are represented as:

wherein G is_x、G_y、G_zRespectively representing gradient values of voxel vertexes in x, y and z directions;

then the obtained G_x、G_y、G_zCarrying out normalization treatment:

the normalized result is taken as the unit normal vector of the voxel vertex (i, j, k). And then normal vectors on all vertexes of the triangular patch can be calculated according to the linear difference function. And finally, drawing an isosurface image according to each vertex coordinate and normal vector on the triangular surface patch, and rendering by using a light renderer to obtain a final surface model to realize three-dimensional drawing of the final image, as shown in fig. 5. Thereby facilitating the preoperative judgment of doctors and leading patients to better know the focus parts of themselves.

Claims (6)

1. A lumbosacral plexus nerve segmentation and three-dimensional visualization method is characterized by comprising the following steps:

(1) processing data; acquiring MRI (magnetic resonance imaging) image data of a patient, storing the MRI image data in a DICOM (digital imaging and communications in medicine) format, converting the acquired data into a gray image, storing the gray image in a png format, and performing normalization processing on the image;

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

(3) three-dimensional reconstruction; stacking the divided two-dimensional plane images, giving an original voxel space, performing three-dimensional reconstruction by using a VTK package and adopting a moving cube algorithm, and displaying the internal details of a reconstructed object to a user by using an isosurface structure.

2. The lumbosacral plexus nerve segmentation and three-dimensional visualization method of claim 1, wherein in the step (2), the deep neural network comprises an encoder, a decoder, a spatial convolutional coding module, a residual jump linking module, and a scale attention module; the preprocessed neural image firstly enters an encoder, the encoder comprises four spatial domain convolutional coding modules with different sizes, an input image obtains feature maps with different scales, then a last layer of feature map enters a decoder, the feature maps with different scales in the encoder are respectively fused with the feature maps with corresponding scales in the decoder, finally the decoder with different scales obtains the feature map with the same dimension as the input image through bilinear interpolation processing, and the feature maps after the cascade connection are processed through a scale attention module and a residual jump link module to obtain a final segmentation image.

3. The lumbosacral plexus nerve segmentation and three-dimensional visualization method as claimed in claim 2, wherein in step (2), each multi-aperture encoder module first obtains a feature map X through a convolution layer with kernel size of 3X 3, the feature map using different convolution rates and convolution kernels to calculate the receptive field, the receptive field being defined as follows:

wherein r is_nShowing the receptive field of the current layer, r_n-1Upper layer, s, representing the receptive field_iDenotes the step size of the i-th convolution, k denotes the size of the convolution kernel;

the spatial convolution function F (x) has 4 cascade branches with different convolution rates and convolution kernels, wherein the convolution rates are 1 to 1, 3 and 5, and the convolution kernels comprise 3 x 3 and 1 x 1;

[r₃:r₇:r₉:r₁₉]＝F(x)

wherein [: represents cascade;

then, the receptive fields for each branch would be set to 3,7,9,19 and a cascade profile r is obtained₃:r₇:r₉:r₁₉]The cascade signature applies a 1 × 1 convolution and a linear activation function; finally, the input feature map is added to the previous feature map,obtaining a new resolution level characteristic diagram, and obtaining the final output through a convolution layer with the kernel size of 3 multiplied by 3 and a pooling layer with the kernel size of 2 multiplied by 2.

4. The lumbosacral plexus nerve segmentation and three-dimensional visualization method as claimed in claim 2, wherein in step (2), in the residual jump linking module, features of different dimensions are mapped to a convolution layer having 3 x 3 kernels, a batch normalization layer and a ReLU as an activation layer; meanwhile, feature mapping of different dimensions also becomes a convolution layer with 3 cores and a layer; adding the characteristic diagram H (X) and the original characteristic diagram X in a short circuit mode to form a new characteristic diagram, and putting the new characteristic diagram into a ReLU activation function to obtain a new characteristic diagram beta:

β＝ReLU(H(X)+E(X))

wherein ReLU () represents an activation function;

in addition, a 1 × 1 convolutional layer is added in the short circuit, which can provide some additional spatial features e (x).

5. The method for lumbosacral plexus segmentation and three-dimensional visualization as claimed in claim 2, wherein in the step (2), the scale attention module resamples the feature maps of different scales obtained from the decoder to the size of the predicted image by using bilinear interpolation, and concatenates the three feature maps of different scales as input

After cascade connection

Representing a feature map with an input size of C H W, where C represents the input channel, H and W represent the height and width, respectively, of the feature map, and the input

Respectively pass through a global average pooling GAP layer and a global maximum pooling GMP layer in order to simplify the feature map parameters and obtainObtaining the weight information of each channel, and outputting the weight information which is respectively expressed as P_GAP(X)∈R^1×1×CAnd P_GMP(X)∈R^1×1×CMultilayer perceptual MLP is implemented with two fully connected layers to obtain a scale attention coefficient and X₁And X₂Share between them, the scale attention coefficient alpha is in [0,1 ]]Comprises the following steps:

α＝Sigmoid(X₁+X₂)

in the scale attention module, a space attention block is additionally used for acquiring space information, the space information consists of a 3 x 3 convolution layer and a 1 x 1 convolution layer, then a Dropout layer is applied to enhance the generalization capability, an output feature map is multiplied by alpha to obtain a new feature map through a Sigmoid function, the feature map is added with an original feature map in the space attention module to obtain an output feature map gamma, and finally the feature map gamma is added with an original input feature map to obtain the output feature map of the multi-scale attention block

Comprises the following steps:

finally, the feature map subjected to scale attention is subjected to a one-dimensional convolution with the kernel size of 1 × 1 and a Sigmoid function to obtain an output segmentation image which has the same shape as the input image.

6. The lumbosacral plexus nerve segmentation and three-dimensional visualization method as claimed in claim 1, wherein in step (3), the segmented two-dimensional plane images are stacked, an original voxel space is given, three-dimensional reconstruction is performed by using a moving cube algorithm using a VTK package, and internal details of a reconstructed object are displayed to a user by using an isosurface structure, specifically: firstly, reading all the iconography sequence data of the patient in a layered reading mode, scanning two layers of data, constructing voxels one by one, wherein 8 corner points in the voxels are taken from four pixel points which correspond to the upper part and the lower part of two slices, the two slices form a cube, and the isosurface is defined 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_kIf the function value F of the vertex is more than or equal to N, the vertex is positioned in the equivalent surface and is marked as plus; if the function value F of the vertex<N, the vertex is positioned outside the isosurface and is marked as-, then a state table of the voxel is constructed by comparing the corner point function value of the voxel with the isosurface, the boundary voxel with an intersection point with the isosurface can be obtained by the obtained state table, and then the position coordinate (x, y, z) of the intersection point is obtained by calculating the isosurface and the cube through a linear interpolation method:

to display the iso-surface image using graphics hardware, the normal components of the triangular patches of the iso-surface are generated, however for each point on the iso-surface, the gradient component of the tangent in the direction of the triangular surface is zero, then the gradient vector for that point represents the normal vector of the iso-surface at that point, assuming the voxel vertex is (i, j, k) and thus the gradient values are represented as:

wherein G is_x、G_y、G_zRespectively representing gradient values of voxel vertexes in x, y and z directions;

then the obtained G_x、G_y、G_zCarrying out normalization treatment:

and finally, drawing an isosurface image according to each vertex coordinate and the normal vector on the triangular surface patch, and rendering by using a light renderer to obtain a final surface model so as to realize the three-dimensional drawing of the final image.

Images