Automatic planning method for bone cement channel based on vertebral edema area

Abstract

The invention belongs to the field of orthopedics, and provides a vertebral body edema area-based bone cement channel automatic planning method, which comprises the following steps of S1, based on MRI image segmentation identification, screening all to-be-processed vertebral sections with edema areas, and calculating the geometric center of the edema area as a first target position of a bone cement channel; s2, establishing a mapping relation between the to-be-processed vertebral segment MRI image and the CT image, and converting a first target position in an MRI image slice to a second target position corresponding to the CT image; s3, geometric centers of the vertebral pedicle areas on the two sides are calculated to serve as third target positions of the bone cement channel; the second target position on the CT image is shifted towards the two sides of the centrum by a set distance, a fourth target position is obtained, and the connecting line of the third target position and the fourth target position corresponding to the two sides is the bone cement planning channel. According to the invention, the bone cement channel can be automatically established through automatic analysis and calculation based on the multi-modal medical image.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing and relates to an auxiliary planning method based on medical images. Background Art

[0002] In orthopedic surgery, bone cement injection has become an important tool for treating vertebral fractures and vertebral repair. The effectiveness of this technique depends heavily on precise cement channel establishment and injection control. However, existing techniques present numerous challenges, particularly in cases associated with vertebral edema.

[0003] Traditionally, establishing a bone cement surgical channel relies primarily on the surgeon's experience and manual manipulation. This approach not only requires extensive clinical experience but also requires multiple intraoperative C-arm X-rays for fluoroscopic confirmation. In practice, the lack of precise positioning guidance often leads to uneven bone cement distribution and increased risk of leakage. More seriously, improper manipulation can lead to complications such as new vertebral compression fractures.

[0004] While existing medical imaging technologies, such as CT, can provide detailed images of vertebral structures, their application in planning bone cement surgeries remains limited. In particular, existing technologies have yet to provide effective solutions for identifying and addressing areas of vertebral edema. Accurately identifying edematous areas and planning optimal bone cement injection pathways accordingly presents a pressing technical challenge in this field.

[0005] In addition, the existing technology lacks automated planning methods, resulting in low surgical efficiency and cumbersome operation procedures, which to some extent limits the promotion and application of bone cement injection technology. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for automatically planning bone cement channels based on vertebral edema areas. This method uses MRI images to identify edema areas within the vertebral body and combines CT images to plan bone cement injection channels. It solves the problem of automatic planning of bone cement channels in vertebral edema areas of osteoporotic vertebral fractures, greatly shortens the doctor's planning and confirmation time, improves surgical efficiency, and has a wide range of application scenarios in related surgical fields.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The method for automatically planning bone cement channels based on vertebral edema areas includes the following steps:

[0009] Step S1: Obtain a sequence of MRI image slices of the case's vertebral segment, select a slice located at the anatomical plane at the center of the vertebral segment, input a first deep learning network model for segmenting the two-dimensional contour of the vertebral body, and obtain the segmentation result of the vertebral body in the MRI image slice;

[0010] Screening all vertebral segments to be processed that have edema areas, and calculating the geometric center of the edema area as the first target position of the bone cement channel;

[0011] Step S2: Obtain a CT image of the case's vertebral segment, input a second deep learning network model for segmenting the vertebral body three-dimensional contour, or perform manual annotation to obtain a segmentation result of the vertebral segment in the CT image;

[0012] Inputting the segmentation result of the vertebral segment to be processed into a third deep learning network model for segmenting the vertebral segment structure to obtain the vertebral body and pedicle segmentation results of the vertebral segment to be processed;

[0013] Based on the segmentation result of the vertebral segment to be processed in the MRI image slice obtained in step S1 and the segmentation result of the vertebral body and pedicle of the vertebral segment to be processed obtained in step S2, a mapping relationship between the MRI image and the CT image of the vertebral segment to be processed is established according to the matched feature information in the MRI image and the CT image, and a first target position in the MRI image slice is converted to a second target position corresponding to the CT image through the mapping relationship;

[0014] Step S3, based on the pedicle segmentation result of the vertebra to be processed obtained in step S2, calculating the geometric center of the pedicle area on both sides as the third target position of the bone cement channel;

[0015] The second target position on the CT image is offset to the two sides of the vertebral body by a set distance to obtain the fourth target position. The line connecting the third target position and the fourth target position on both sides is the bone cement planning channel.

[0016] Preferably, in step S1, the first deep learning network model is set to a U-Net network model, and the downsampling structure of the U-Net network model includes a fixed window self-attention module, a sliding window self-attention module and a multi-layer perceptron module. The fixed window self-attention module and the sliding window self-attention module perform self-attention calculations in the fixed window and sliding window of the MRI image slice in turn, and the multi-layer perceptron module performs nonlinear transformation and inter-channel information fusion on the output features of the self-attention module.

[0017] Preferably, the downsampling structure of the U-Net network model is configured as follows: the MRI image slices are processed in sequence by a fixed window self-attention module, a multi-layer perceptron module, a sliding window self-attention module and a multi-layer perceptron module, and the processing process of any module includes normalization of the input image and feature fusion of the input image and the output image.

[0018] Preferably, in step S1, the screening of the vertebral segments to be processed having edema areas in all vertebral segments includes: binarizing the segmentation results of all vertebral segments in the MRI image slices to obtain a first ratio of the foreground area and the vertebral segment area after binarization of all vertebral segments, comparing the first ratio with the second ratio of the foreground area and the vertebral segment area of ​​the normal vertebral segment, and screening out the vertebral segments to be processed having edema areas.

[0019] Preferably, the second deep learning network model includes a 3D U-Net network model or a nnUNet network model.

[0020] Preferably, the third deep learning network model includes a 3D U-Net network model or a nnUNet network model.

[0021] Preferably, in step S2, the mapping relationship between the MRI image and the CT image of the vertebra to be processed is established based on the matching feature information in the MRI image and the CT image, including: establishing a mapping relationship between the MRI image and the CT image of the vertebra to be processed based on the contour information of the upper end plate or the lower end plate of the vertebral body matched in the MRI image and the CT image.

[0022] Preferably, the obtaining of the contour information of the upper or lower endplate of the vertebral body in the MRI image and the CT image includes the following steps: based on the vertebral segmentation result of the vertebra to be processed, obtaining the geometric center of the vertebral body, and determining the vertebral body contour on the set sagittal plane passing through the geometric center, and fitting the upper or lower endplate straight line on the sagittal vertebral body contour.

[0023] Preferably, the determining of the vertebral contour on the set sagittal plane passing through the geometric center includes: obtaining a vertebral surface point cloud based on the vertebral segmentation result of the vertebral segment to be processed, calculating a set distance between the vertebral surface point cloud and the set sagittal plane, selecting a point set whose set distance is less than the pixel spacing and projecting it onto the set sagittal plane to form the vertebral contour on the set sagittal plane.

[0024] Preferably, in step S2, the mapping relationship between the MRI image and the CT image of the vertebra to be processed is established, including: establishing a coordinate system in the MRI image and the CT image, respectively, based on the straight line direction of the upper end plate or the lower end plate and the geometric center of the vertebral body: the geometric center of the vertebral body is recorded as the origin O of the coordinate system, the straight line direction of the upper end plate or the lower end plate is recorded as the x-axis direction, and the direction perpendicular to the x-axis and pointing to the other end plate is recorded as the y-axis direction, so as to obtain the coordinate system.

[0025] Preferably, the establishing of the mapping relationship between the MRI image and the CT image of the vertebra to be processed includes: determining the transformation relationship between the MRI image coordinate system and the CT image coordinate system:

[0026]

[0027] Where T is the two-dimensional mapping transformation coefficient, θ is x ct and x mri Angle; C ct is the two-dimensional coordinate system of the vertebral contour of the CT image, and O ct The x-axis coordinate and y-axis coordinate, x ct and O ct C ct The x-axis direction and origin of C mri is the two-dimensional coordinate system of the vertebral contour of the MRI image, and O mri The x-axis coordinate and y-axis coordinate, x mri and O mri C mri The x-axis direction and origin;

[0028] The second target position P ct The calculation formula is:

[0029] P ct =T·P mri

[0030] Where, P mri is the first target position.

[0031] Preferably, in step S3, offsetting the second target position on the CT image by a set distance to both sides of the vertebral body to obtain the fourth target position includes: calculating the vertical distance between the second target position and the contours of both sides of the vertebral body, and setting the distance to half of the vertical distance.

[0032] Preferably, the upper endplate or lower endplate straight line is fitted on the vertebral body contour of the set sagittal plane, and the fitting method includes a least squares algorithm and a random sampling consistency algorithm.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention segments the vertebral contour and extracts the edema area therein based on the MRI image of the case vertebral segment. By performing contour segmentation on the corresponding vertebral segment in the CT image, a registration relationship between the MRI image and the CT image is established. Based on the geometric relationship between the edema area and the vertebral body, an optimal bone cement injection channel is constructed on the CT image to ensure that the channel can smoothly penetrate the pedicle. At the same time, the final target point of the channel is positioned at the center of the edema area, which greatly improves the surgical accuracy, reduces complications and reduces the difficulty of operation.

[0035] 2. The first deep learning model used in the present invention to segment the vertebral contour of MRI images is a U-Net network model. The downsampling structure of the model includes a fixed window self-attention module, a sliding window self-attention module and a multi-layer perceptron module. By performing self-attention calculations in the fixed window and sliding window of the MRI image slice in turn, and using a multi-layer perceptron module to perform nonlinear transformation and inter-channel information fusion on the output features of the self-attention module, image segmentation with different scales and spatial resolution features can be achieved. While retaining the global features of the image, the amount of calculation is reduced, and the adaptability to different image sizes is stronger. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the method for automatically planning bone cement channels based on vertebral edema areas according to the present invention;

[0037] Figure 2 This is the MRI image of vertebral edema in the case of the present invention;

[0038] Figure 3 Schematic diagram of the traditional U-Net network structure of the present invention (the red dotted box is the downsampling module);

[0039] Figure 4 This is a schematic diagram of the multi-head attention mechanism structure of the present invention;

[0040] Figure 5 This is a schematic diagram of the vertebral segmentation results and the extraction of the vertebral edema area (yellow area) of the MRI slice image of the present invention;

[0041] Figure 6 Schematic diagram of vertebral segmentation and recognition in CT images of cases in the present invention (A is a 3D schematic diagram; B is a 2D schematic diagram);

[0042] Figure 7 Schematic diagram of the segmentation and extraction of the thoracic vertebra T12 vertebral pedicle of the present invention (A is the T12 vertebral segment; B is the segmentation result);

[0043] Figure 8 Schematic diagram of the vertebral surface voxel point cloud and sagittal section of the CT image of the case of the present invention (green is the schematic diagram of the surface voxel point cloud, and red is the schematic diagram of the sagittal section through the geometric center of the voxel point cloud);

[0044] Figure 9 The vertebral contour coordinate system and mapping transformation of MRI and CT images of the present invention (A is the vertebral contour coordinate system of MRI images, B is the vertebral contour coordinate system of CT images, and C is a schematic diagram of the mapping transformation);

[0045] Figure 10 Schematic diagram of the bone cement channel position points on both sides of the CT image of the case of the present invention (A is the target point P of the bone cement channel of the edematous vertebra on the CT image ct Sagittal diagram of the vertebral body with edema. B is the target point P of the bone cement channel on the CT image. ct cross-sectional diagram). DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is merely intended to provide a better understanding of the present application by illustrating examples of the present application.

[0047] like Figure 1 As shown, the present invention provides a method for automatically planning a bone cement channel based on a vertebral edema area, comprising the following steps:

[0048] Step S1: Obtain a sequence of MRI image slices of the case vertebra. The MRI image of the case vertebra is generally composed of a group of slices scanned in the left and right directions of the human body. A slice in the middle is selected as the slice closest to the anatomical plane of the cone center (e.g. Figure 2 As shown), the slice is input into the first deep learning network model for segmenting the two-dimensional contour of the vertebral body. Figure 2 Perform two-dimensional segmentation on each vertebral region in the MRI image slice to obtain the segmentation results of the vertebral body. Figure 5 shown.

[0049] It should be noted here that the first deep learning network model in this embodiment is an improved U-Net network model, which is obtained by:

[0050] 1) Traditional U-Net network model structure (see Figure 3 ) is called the encoder, which includes a four-layer downsampling structure, and the right part is called the decoder, which includes a four-layer upsampling structure. In order to achieve segmentation of features of different scales and spatial resolutions, this embodiment replaces the traditional four-layer downsampling structure with Figure 4The normalization operation module (LN), fixed window self-attention module (W-MSA), sliding window self-attention module (SW-MSA) and multi-layer perceptron module (MLP) shown in the figure perform self-attention calculations in the fixed window and sliding window of the MRI image in turn, and fuse the attention features with the original input to obtain an improved U-Net network model structure.

[0051] Specifically, Figure 4 The data processing flow shown is as follows: the input MRI image slice (feature map) is processed in sequence by the W-MSA module, MLP module, SW-MSA module and MLP module. When the feature map enters each module, it must first be normalized and preprocessed by the LN module, and the input value (red channel in the figure) and output value (green channel in the figure) of each module are fused.

[0052] Taking the W-MSA and SW-MSA modules as examples, the input feature map is first normalized by the LN module. The W-MSA module then independently calculates self-attention weights within each window to obtain an output feature map of the internal feature relationship of the local region. After the input and output feature maps are fused, they are used as the input feature map of the MLP module. After processing by the MLP module, the feature map is input to the SW-MSA module and first normalized by the LN module. The SW-MSA module then calculates self-attention weights within overlapping and sliding local windows to obtain an output feature map of cross-window global information interaction. After the input and output feature maps are fused, they are used as the input feature map of the second MLP module.

[0053] Each time the self-attention module outputs features, the MLP module performs nonlinear transformation and inter-channel information fusion on the self-attention module output features, which can improve the anatomical rationality of the segmentation boundary. After the second MLP module processing, the segmentation results of the vertebral body in the MRI image slice are obtained.

[0054] 2) Select at least 50 groups of centered MRI slice images (training set) for vertebral annotation. Then input them into a deep learning network with an improved U-Net network model structure for training. Use the Dice Similarity Coefficient (DICE) as the training result evaluation metric to obtain a trained improved U-Net network model.

[0055] The improved U-Net network structure uses sliding windows to overlap adjacent windows, covering different areas in the next layer. This indirectly enables cross-window information exchange, avoids global computation, and reduces the amount of computation. Furthermore, the self-attention mechanism in this network structure dynamically adjusts weights based on inter-pixel correlations, enabling direct interaction even with distant pixels. The sliding window (SW-MSA) offsets windows so that adjacent layers cover different areas, enabling global feature aggregation. Furthermore, the attention mechanism essentially models inter-pixel relationships, and the window division can be dynamically adjusted, eliminating the need for fixed-size convolution kernels or pooling layers. This allows for greater adaptability to varying image sizes.

[0056] Figure 2 After vertebral segmentation is achieved, the number of segmented vertebrae can be obtained. Each vertebra represents a vertebral segment. Since the number of vertebral segments in the human body is generally fixed, the lumbar vertebrae have five segments L1 to L5, the thoracic vertebrae have twelve segments T1 to T12, and the cervical vertebrae have seven segments C1 to C7. When the case MRI image is collected, the collection site and the vertebral segment information of the site have been determined. Therefore, according to the number of vertebral segments obtained by segmentation, it is easy to identify the vertebral segment represented by each vertebra. Figure 5 The lowest vertebra in the spine is L5.

[0057] Screen all vertebral segments with edema areas and calculate the geometric center of the edema area as the first target position of the bone cement channel, including:

[0058] The OTSU binarization method is used to process the image of each segmented vertebral region, that is, the segmented vertebral region image is binarized into two parts, the foreground and the background. The white foreground represents edema, and the black background represents normal vertebral body. A threshold K is set (defined as the second ratio of the foreground area and the vertebral area of ​​the normal vertebral segment). The threshold K is an indicator for clinicians to judge edema, which is an empirical value or a statistical value. For example, the ratio of the foreground area to the total area of ​​the vertebral region in 10,000 cases of edematous vertebrae and non-edematous vertebrae is counted as KA and KB, and the average of KA and KB is the threshold K. When the ratio of the foreground area to the total area of ​​the vertebral region is greater than K, it is represented as an edematous vertebra. This embodiment uses the OTSU binarization method to determine the following: Figure 5 The yellow area in the T12 vertebra is the foreground edema area after binarization, and the T12 vertebra is the edema vertebra; the geometric center P of the edema area is calculated. mri As the target position point of the bone cement channel (first target position).

[0059] Step S2: Obtain CT images of the case vertebral segments, input the second deep learning network model for segmenting the vertebral three-dimensional contour or perform manual annotation, segment the three-dimensional shape of each vertebral segment in the CT image and identify the vertebral segment name, and obtain the segmentation results of the vertebral segment in the CT image as shown in the figure. Figure 6 As shown, Figure 6 The figure shows the segmentation and recognition results of eight vertebrae in this case, from thoracic vertebra T10 to lumbar vertebra L5. It is known that the vertebra to be processed with edema area is thoracic vertebra T12.

[0060] The second deep learning network model is a deep learning segmentation and recognition network, which can be a 3D U-Net network model or a nnUNet network model. There are many related deep learning network segmentation and recognition technologies in the existing technology, which will not be repeated here.

[0061] The segmentation result of the vertebra to be processed (thoracic vertebra T12) is input into the third deep learning network model for segmenting the vertebral structure (such as 3D U-Net network or nnUNet network model, the present invention adopts 3D U-Net network), and the vertebral body and pedicle segmentation results of the vertebra to be processed are obtained, as shown in FIG. Figure 7 As shown, the vertebral body is Figure 7 The red area and pedicle in B are Figure 7 Green areas in B, these areas are represented by voxel point clouds with different voxel values.

[0062] Based on the segmentation results of the vertebral segment to be processed in the MRI image slice obtained in step S1 and the vertebral body and pedicle segmentation results of the vertebral segment to be processed obtained in step S2, the vertebral surface point cloud (such as Figure 8 The green point cloud shown in Figure 8 The red section is defined as the set sagittal plane, which is the section passing through the geometric center of the vertebral surface point cloud. The geometric center of the vertebral surface point cloud is obtained by averaging the voxel values ​​of the surface point cloud.

[0063] Calculate the set distance between the vertebral surface point cloud of the CT image and the set sagittal plane, select the point set in the vertebral surface point cloud whose distance to the section is less than the image pixel interval, and the perpendicular foot of these point sets on the section can generate the vertebral contour on the sagittal section, such as Figure 9 As shown in B. Use least square fitting or random sampling consensus (RANSAC) algorithms to segment the lower end plate (or upper end plate) of the vertebral contour and fit a straight line, as shown in Figure 9 As shown by the red curve in B, the direction of the fitting line is x ct , with the geometric center of the vertebral body O ct is the origin, x ct The x-axis direction is perpendicular to the x-axis and points to the y-axis of the other end plate. ct The direction is the y-axis direction, and a two-dimensional coordinate system C is established. ct .

[0064] according to Figure 5 The segmentation result of the vertebral segment to be processed with edema area in the MRI image is obtained, and the sagittal contour point cloud of the vertebral body in the MRI image is obtained, such as Figure 9As shown in A, the lower end plate (or upper end plate) of the sagittal contour is segmented using algorithms such as least squares fitting or random sampling consensus (RANSAC), and the direction of the fitting line is denoted as x. mri , with the geometric center of the vertebral body O mri is the origin, x mri The x-axis direction is perpendicular to the x-axis and points to the y-axis of the other end plate. mri The direction is the y-axis direction, and a two-dimensional coordinate system C is established. mri .

[0065] Determine the transformation relationship between the MRI image coordinate system and the CT image coordinate system:

[0066]

[0067] Where T is the two-dimensional mapping transformation coefficient, θ is x ct and x mri The angle between and O ct The x-axis coordinate and y-axis coordinate, and O mri The x-axis coordinate and y-axis coordinate of .

[0068] The transformation coefficient T can achieve the approximate alignment of the vertebral contours in MRI images and CT images (see Figure 9 Therefore, the first target position in the MRI image slice can be converted to the second target position corresponding to the CT image through the mapping relationship, that is, the geometric center P of the edema area obtained in the MRI image mri (first target position), mapped to the corresponding point P in the CT image ct (Second target position), can be calculated as follows:

[0069] P ct =T·P mri .

[0070] Step S3, Figure 10 A and Figure 10 The B in the figure are the target position points P of the bone cement channel in the case CT image. ct Schematic diagram of the sagittal and transverse sections of Figure 10 The target point P of the bone cement channel in A ct Map to Figure 10 In the B cross section, the target point P is obtained ct Located on the middle cross section shown by the red dotted line, according to Figure 10 The B in the figure calculates the relevant position information of the bone cement channels on the left and right sides, and the path P of the bone cement channels on the left and right sides can be completed. l Cl 、P r C r Automatic planning of:

[0071] 1) According to the pedicle segmentation results of the single vertebral segment to be processed obtained in step S2, calculate Figure 10 The geometric center of the left and right pedicles in B is C l 、C r , serving as the waypoint for the bone cement channels on both sides (the third target position).

[0072] 2) In Figure 10 Determine the target point P on B in ct And the vertical line perpendicular to the middle cross section, calculate the target point P ct The first distance and the second distance to the left and right vertebral sides along the vertical line are respectively taken as the first set distance and the second set distance. ct The target position point P of the bone cement channel is obtained by shifting the first set distance and the second set distance to both sides respectively. l 、P r (Fourth target position), the target position point P l 、P r It should be noted that the choice of the set distance can be adjusted according to the patient's actual vertebral structure or the surgeon's experience, and there is no specific limitation.

[0073] 3) The corresponding bone cement channel path C on both sides l 、C r and the target position point P l 、P r Connection P l C l 、P r C r This is the bone cement planning channel.

[0074] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may of course make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. An automatic bone cement channel planning method based on vertebral edema area, characterized in that: The following steps are involved: Step S1: Obtain a sequence of MRI image slices of the case's vertebral segment, select a slice located at the anatomical plane at the center of the vertebral segment, input a first deep learning network model for segmenting the two-dimensional contour of the vertebral body, and obtain the segmentation result of the vertebral body in the MRI image slice; Screening all vertebral segments to be processed that have edema areas, and calculating the geometric center of the edema area as the first target position of the bone cement channel; Step S2: Obtain a CT image of the case's vertebral segment, input a second deep learning network model for segmenting the vertebral body three-dimensional contour, or perform manual annotation to obtain a segmentation result of the vertebral segment in the CT image; Inputting the segmentation result of the vertebral segment to be processed into a third deep learning network model for segmenting the vertebral segment structure to obtain the vertebral body and pedicle segmentation results of the vertebral segment to be processed; Based on the segmentation result of the vertebral segment to be processed in the MRI image slice obtained in step S1 and the segmentation result of the vertebral body and pedicle of the vertebral segment to be processed obtained in step S2, a mapping relationship between the MRI image and the CT image of the vertebral segment to be processed is established according to the matched feature information in the MRI image and the CT image, and a first target position in the MRI image slice is converted to a second target position corresponding to the CT image through the mapping relationship; Step S3, based on the pedicle segmentation result of the vertebra to be processed obtained in step S2, calculating the geometric center of the pedicle area on both sides as the third target position of the bone cement channel; The second target position on the CT image is offset to the two sides of the vertebral body by a set distance to obtain the fourth target position. The line connecting the third target position and the fourth target position on both sides is the bone cement planning channel.

2. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: In step S1, the first deep learning network model is set to a U-Net network model. The downsampling structure of the U-Net network model includes a fixed window self-attention module, a sliding window self-attention module and a multi-layer perceptron module. The fixed window self-attention module and the sliding window self-attention module perform self-attention calculations in the fixed window and sliding window of the MRI image slice in turn. The multi-layer perceptron module performs nonlinear transformation and inter-channel information fusion on the output features of the self-attention module.

3. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 2, characterized in that: The downsampling structure of the U-Net network model is configured as follows: MRI image slices are processed sequentially by a fixed window self-attention module, a multi-layer perceptron module, a sliding window self-attention module, and a multi-layer perceptron module. The processing process of any module includes normalization of the input image and feature fusion of the input image and the output image.

4. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: In step S1, the vertebral segments to be processed with edema areas in all vertebral segments are screened, including: binarizing the segmentation results of all vertebral segments in the MRI image slices to obtain a first ratio of the foreground area and the vertebral segment area after binarization of all vertebral segments, comparing the first ratio with the second ratio of the foreground area and the vertebral segment area of ​​the normal vertebral segment, and screening out the vertebral segments to be processed with edema areas.

5. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: The second deep learning network model includes a 3D U-Net network model or a nnUNet network model.

6. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: The third deep learning network model includes a 3D U-Net network model or a nnUNet network model.

7. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: In step S2, the mapping relationship between the MRI image and the CT image of the vertebra to be processed is established based on the matching feature information in the MRI image and the CT image, including: establishing a mapping relationship between the MRI image and the CT image of the vertebra to be processed based on the contour information of the upper end plate or the lower end plate of the vertebrae matched in the MRI image and the CT image.

8. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 7, characterized in that: The method of obtaining the contour information of the upper or lower endplate of the vertebral body in the MRI image and the CT image includes the following steps: based on the vertebral body segmentation result of the vertebral segment to be processed, obtaining the geometric center of the vertebral body, determining the vertebral body contour on a set sagittal plane passing through the geometric center, and fitting the upper or lower endplate straight line on the sagittal plane vertebral body contour.

9. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 8, characterized in that: The method of determining the vertebral contour on the set sagittal plane passing through the geometric center includes: obtaining a vertebral surface point cloud based on the vertebral segmentation result of the vertebral segment to be processed, calculating a set distance between the vertebral surface point cloud and the set sagittal plane, selecting a point set whose set distance is less than the pixel spacing and projecting it onto the set sagittal plane to form the vertebral contour on the set sagittal plane.

10. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 8, characterized in that: In step S2, the mapping relationship between the MRI image and the CT image of the vertebra to be processed is established, including: establishing a coordinate system in the MRI image and the CT image respectively based on the straight line direction of the upper end plate or the lower end plate and the geometric center of the vertebra: the geometric center of the vertebra is recorded as the origin O of the coordinate system, the straight line direction of the upper end plate or the lower end plate is recorded as the x-axis direction, and the direction perpendicular to the x-axis and pointing to the other end plate is recorded as the y-axis direction, so as to obtain the coordinate system.

11. The method for automatically planning bone cement channels based on vertebral edema areas according to claim 10, characterized in that: The establishment of the mapping relationship between the MRI image and the CT image of the vertebra to be processed includes: determining the transformation relationship between the MRI image coordinate system and the CT image coordinate system: Where T is the two-dimensional mapping transformation coefficient, θ is x ct and x mri Angle; C ct is the two-dimensional coordinate system of the vertebral contour of the CT image, and O ct The x-axis coordinate and y-axis coordinate, x ct and O ct C ct The x-axis direction and origin of C mri is the two-dimensional coordinate system of the vertebral contour of the MRI image, and O mri The x-axis coordinate and y-axis coordinate, x mri and O mri C mri The x-axis direction and origin; The second target position P ct The calculation formula is: P ct =T·P mri Where, P mri is the first target position.

12. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 1, characterized in that: In step S3, the second target position on the CT image is shifted to both sides of the vertebral body by a set distance to obtain a fourth target position, including: calculating the vertical distance between the second target position and the contours of both sides of the vertebral body, and setting the distance to half of the vertical distance.

13. The method for automatic planning of bone cement channels based on vertebral edema areas according to claim 8, characterized in that: The upper end plate or lower end plate straight line is fitted on the vertebral body contour of the set sagittal plane, and the fitting method includes a least squares algorithm and a random sampling consistency algorithm.

Images