A deep learning point cloud lumbar spine registration method for minimally invasive spinal surgery navigation

Through deep learning point cloud technology, using three-dimensional V-shaped deep segmentation network model and point cloud conversion, the rapid registration of the lumbar spine during minimally invasive spine surgery is achieved, and the radiation exposure and coordination caused by multiple X-ray shooting in the existing technology is solved. Quasi-complexity issues improve the safety and efficiency of surgery.

CN115049709BActive Publication Date: 2024-11-12NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210578192.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-26
Publication Date
2024-11-12
Estimated Expiration
2042-05-26

AI Technical Summary

Technical Problem

Prior art In minimally invasive spinal surgery, especially during percutaneous puncture, X-ray images are required to achieve three-dimensional model registration of the lumbar spine, resulting in radiation exposure to doctors and patients, and the registration algorithm is complex and difficult to quickly Finish.

Method used

Using deep learning point cloud technology, lumbar image segmentation is performed by collecting preoperative and intraoperative spine CT images, using a three-dimensional V-shaped depth segmentation network model to eliminate isolated points, convert them into three-dimensional point clouds, and obtain the most by iterating from nearest neighbor points. Optimal rotation matrix and translation vector to achieve rapid registration of the lumbar spine.

Benefits of technology

Reduced the number of image shots, achieved rapid registration of lumbar spine images, reduced the risk of radiation exposure, and simplified the registration process, improving the safety and efficiency of the surgery.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115049709B_ABST
    Figure CN115049709B_ABST
Patent Text Reader

Abstract

The present invention discloses a deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation, comprising: collecting preoperative and intraoperative spinal CT images; inputting the spinal CT images into a trained three-dimensional V-shaped deep segmentation network model for segmentation to obtain preoperative and intraoperative lumbar CT images; eliminating isolated points in the lumbar CT images by a morphological method; converting the preoperative and intraoperative lumbar CT images after eliminating isolated points into corresponding lumbar three-dimensional point cloud sets by a point cloud conversion method; performing nearest neighbor point iteration on the preoperative and intraoperative lumbar three-dimensional point cloud sets to complete lumbar registration in the preoperative and intraoperative spinal CT images; the present invention can reduce the number of image capture times and simultaneously achieve rapid registration of lumbar images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, belonging to the technical field of medical image processing. Background Art

[0002] Image-guided surgery (IGS) is a system that combines medical image processing, spatial positioning, and computer vision technologies. It obtains the relative position and posture relationship between surgical instruments and the anatomical structures of the patient's surgical area through three-dimensional reconstruction, target recognition, and image registration of the anatomical structures in the surgical area. It also uses augmented reality display technology to display the navigation results in real time to guide the implementation of minimally invasive spinal surgery. Image-guided surgery systems can solve the problem of invisible surgery, especially in the process of percutaneous puncture, and have important practical significance for clinical surgical operations such as surgical instrument implantation and vertebral degenerative lesions. Hwang et al. used multiple two-dimensional images obtained during surgery to reconstruct a three-dimensional model and achieve registration with the preoperative three-dimensional image. However, the process of taking multiple X-ray images during surgery increases the radiation exposure of doctors and patients, especially for surgeons who perform surgery for a long time. The accumulation of X-ray radiation can lead to potential pathogenic risks. Moura et al. proposed a dual-plane reconstruction technology, which uses a deformable hinge model as statistical prior knowledge. By extracting the lumbar midline and reconstructing a three-dimensional model of the lumbar spine using two X-ray images in the anteroposterior and lateral positions, the 2D-3D lumbar image registration is achieved. However, the registration algorithms in different dimensions are relatively complex. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation, which can reduce the number of image capture times and achieve rapid registration of lumbar images.

[0004] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0005] In a first aspect, the present invention provides a deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation, comprising:

[0006] Collect preoperative and intraoperative spinal CT images;

[0007] The spinal CT images are input into the trained three-dimensional V-shaped deep segmentation network model for segmentation to obtain preoperative and intraoperative lumbar CT images;

[0008] The isolated points in lumbar spine CT images were eliminated by morphological methods;

[0009] The preoperative and intraoperative lumbar spine CT images after eliminating isolated points are converted into corresponding lumbar spine three-dimensional point cloud sets by point cloud conversion method;

[0010] Perform nearest neighbor iteration on the lumbar 3D point cloud set before and during surgery to obtain the optimal rotation matrix and translation vector;

[0011] The three-dimensional point cloud set is transformed through the optimal rotation matrix and translation vector to complete the lumbar spine registration in the preoperative and intraoperative spinal CT images.

[0012] Optionally, the training of the three-dimensional V-shaped deep segmentation network model includes:

[0013] Obtain spinal CT images and add lumbar contour annotations to generate sample images, and build a training set based on the sample images;

[0014] Based on the training set, the Adam algorithm is used to train the three-dimensional V-shaped deep segmentation network model, and the model parameters are adjusted until the loss function is minimized.

[0015] Optionally, the adding of lumbar vertebrae contour annotation includes annotating voxels belonging to the lumbar vertebrae in the spinal CT image as 1, and annotating voxels not belonging to the lumbar vertebrae as 0;

[0016] The loss function L dice for:

[0017]

[0018] In the formula, G i is the annotation information of voxel i in the spine CT image, P i is the probability that voxel i in the spinal CT image is labeled as 1; N is the number of voxels in the spinal CT image.

[0019] Optionally, the three-dimensional V-shaped deep segmentation network model includes a symmetrical encoder and decoder;

[0020] The encoder includes an input layer and four downsampling layers connected in sequence, the input layer performs unit addition with the input image through a residual operation, and then downsamples and outputs a feature map layer by layer through the four downsampling layers; the image input dimension of the input layer is 64×64×64, and the number of channels is increased to 16 using a 5×5×5 convolution kernel; after the first downsampling layer performs two convolutions, it performs unit addition with the output of the input layer through a residual operation, and then downsamples using a 2×2×2 convolution kernel with a step size of 2; after the last three downsampling layers perform three convolutions, they perform unit addition with the output of the previous downsampling layer through a residual operation, and then downsample using a 2×2×2 convolution kernel with a step size of 2;

[0021] The decoder includes four upsampling layers and an output layer connected in sequence. The four upsampling layers upsample the feature map output by the last downsampling layer layer by layer to output the feature map, and then perform unit addition with the feature map output by the last upsampling layer through a residual operation through the output layer to output the segmentation result; after the first three upsampling layers perform three convolutions, they perform unit addition with the output of the previous upsampling layer through a residual operation, and then use a 2×2×2 convolution kernel with a step size of 2 for upsampling; after the last upsampling layer performs two convolutions, it performs unit addition with the output of the output layer through a residual operation, and then uses a 2×2×2 convolution kernel with a step size of 2 for upsampling; the image output dimension of the output layer is 64×64×64, and a 5×5×5 convolution kernel is used to reduce the number of channels to 1.

[0022] Optionally, eliminating isolated points in the lumbar CT image by a morphological method includes: performing morphological opening and closing operations on the lumbar CT image, and the template sizes of the opening and closing operations are adjusted according to the sizes of the isolated points to be segmented.

[0023] Optionally, the conversion into a corresponding lumbar vertebrae three-dimensional point cloud set includes: performing surface drawing on the preoperative and intraoperative lumbar vertebrae CT images after eliminating isolated points through a Marching Cubes algorithm, traversing the vertex coordinates of the surface drawing results and converting them into point cloud coordinates, and obtaining the preoperative and intraoperative lumbar vertebrae three-dimensional point cloud set.

[0024] Optionally, performing nearest neighbor iteration on the lumbar 3D point cloud set before and during surgery to obtain the optimal rotation matrix and translation vector includes:

[0025] Initialize the number of iterations i = 1;

[0026] The lumbar 3D point cloud collection C before surgery source Pick point set p i ∈C source ;

[0027] Lumbar 3D point cloud collection C during surgery target Pick a point set q i ∈C target , so that ||p i -q i || is the minimum, ||·|| is the Euclidean distance;

[0028] Construct the error function:

[0029]

[0030] The error function is minimized by singular value decomposition to obtain the rotation matrix R i and the translation vector T i ;

[0031] According to the rotation matrix R k and the translation vector T k Point set p i Perform rotation and translation transformation to obtain the point set p′ i ;

[0032] Calculate the point set p′ i and the corresponding point set q i The average distance d;

[0033] Determine whether the average distance d is less than the preset threshold. If so, stop the iteration and output the rotation matrix R i and the translation vector T i ;

[0034] If not, determine whether the number of iterations i is greater than the preset number of iterations. If so, stop the iteration and output the rotation matrix R i and the translation vector T i ;

[0035] If not, set the number of iterations i=i+1 and repeat the above iteration steps.

[0036] In a second aspect, the present invention provides a deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, the device comprising:

[0037] Image acquisition module, used to acquire spinal CT images before and during surgery;

[0038] An image segmentation module is used to input the spinal CT image into the trained three-dimensional V-shaped deep segmentation network model for segmentation, and obtain the preoperative and intraoperative lumbar CT images;

[0039] An image processing module, used to eliminate isolated points in lumbar CT images by using a morphological method;

[0040] A point cloud conversion module, used for converting the preoperative and intraoperative lumbar CT images after eliminating isolated points into corresponding lumbar 3D point cloud sets through a point cloud conversion method;

[0041] The image registration module is used to perform nearest neighbor iteration on the preoperative and intraoperative lumbar 3D point cloud set to complete the lumbar registration in the preoperative and intraoperative spinal CT images.

[0042] In a third aspect, the present invention provides a deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, including a processor and a storage medium;

[0043] The storage medium is used to store instructions;

[0044] The processor is used to operate according to the instructions to execute the steps according to the above method.

[0045] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention provides a deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation. The spinal CT image is input into a trained three-dimensional V-shaped deep segmentation network model for segmentation. There is no need to collect images multiple times, and the lumbar CT image can be quickly acquired. The lumbar CT image is optimized by a morphological method, and an optimized point cloud set is obtained according to point cloud conversion. Then, fast 3D-3D lumbar image registration is achieved through nearest neighbor iteration, which provides a guarantee for solving the invisible problem of clinical minimally invasive lumbar surgery. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of a deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation provided in Example 1 of the present invention;

[0049] Figure 2 It is a schematic diagram of the structure of a three-dimensional V-shaped deep segmentation network model provided in Example 1 of the present invention;

[0050] Figure 3 is a schematic diagram of a lumbar 3D point cloud set before and during surgery provided by the first embodiment of the present invention;

[0051] Figure 4 It is a schematic diagram of the point cloud lumbar vertebra registration result provided by the first embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0053] Embodiment 1:

[0054] like Figure 1 As shown, an embodiment of the present invention provides a deep learning point cloud lumbar registration method for spinal minimally invasive surgery navigation, comprising the following steps:

[0055] 1. Collect spinal CT images before and during surgery.

[0056] 2. Input the spinal CT images into the trained three-dimensional V-shaped deep segmentation network model for segmentation to obtain preoperative and intraoperative lumbar CT images;

[0057] Among them, the training of the three-dimensional V-shaped deep segmentation network model includes:

[0058] Obtain spinal CT images and add lumbar contour annotations to generate sample images, and build a training set based on the sample images;

[0059] Based on the training set, the Adam algorithm is used to train the three-dimensional V-shaped deep segmentation network model, and the model parameters are adjusted until the loss function is minimized.

[0060] Among them, adding lumbar vertebrae contour annotation includes annotating voxels belonging to the lumbar vertebrae in the spinal CT image as 1, and annotating voxels not belonging to the lumbar vertebrae as 0;

[0061] Loss function L dice for:

[0062]

[0063] In the formula, G i is the annotation information of voxel i in the spine CT image, P i is the probability that voxel i in the spinal CT image is labeled as 1; N is the number of voxels in the spinal CT image.

[0064] 3. Eliminate isolated points in lumbar CT images by morphological methods;

[0065] Eliminating isolated points in the lumbar CT image by using a morphological method includes: performing morphological opening and closing operations on the lumbar CT image, and adjusting the template sizes of the opening and closing operations according to the sizes of the isolated points to be segmented.

[0066] 4. The preoperative and intraoperative lumbar CT images after eliminating isolated points are converted into corresponding lumbar 3D point cloud sets by using the point cloud conversion method;

[0067] The conversion to the corresponding lumbar 3D point cloud set includes: using the Marching Cubes algorithm to perform surface drawing on the preoperative and intraoperative lumbar CT images after eliminating isolated points, traversing the vertex coordinates of the surface drawing results and converting them into point cloud coordinates, and obtaining the preoperative and intraoperative lumbar 3D point cloud set. The schematic diagram of the lumbar 3D point cloud set is shown in Figure 3 As shown, (a) and (b) are the 3D point cloud sets of the lumbar spine before and during surgery, respectively.

[0068] 5. Perform nearest neighbor iteration on the lumbar 3D point cloud set before and during surgery to obtain the optimal rotation matrix and translation vector; including:

[0069] Initialize the number of iterations i = 1;

[0070] The lumbar 3D point cloud collection C before surgery source Pick point set p i ∈C source ;

[0071] Lumbar 3D point cloud collection C during surgery target Pick a point set q i ∈C target , so that ||p i -q i || is the minimum, ||·|| is the Euclidean distance;

[0072] Construct the error function:

[0073]

[0074] The error function is minimized by singular value decomposition to obtain the rotation matrix R i and the translation vector T i ;

[0075] According to the rotation matrix R k and the translation vector T k Point set p i Perform rotation and translation transformation to obtain the point set p′ i ;

[0076] Calculate the point set p′ i and the corresponding point set q i The average distance d;

[0077] Determine whether the average distance d is less than the preset threshold. If so, stop the iteration and output the rotation matrix R i and the translation vector T i ;

[0078] If not, determine whether the number of iterations i is greater than the preset number of iterations. If so, stop the iteration and output the rotation matrix R i and the translation vector T i ;

[0079] If not, set the number of iterations i=i+1 and repeat the above iteration steps.

[0080] 6. The three-dimensional point cloud set is transformed by the optimal rotation matrix and translation vector to complete the lumbar registration in the spinal CT images before and during surgery. The schematic diagram of lumbar registration is shown in the figure below. Figure 4 shown.

[0081] like Figure 2 As shown, the three-dimensional V-shaped deep segmentation network model includes a symmetrical encoder and decoder;

[0082] The encoder includes an input layer and four downsampling layers connected in sequence. The input layer performs unit addition with the input image through the residual operation, and then downsamples the output feature map layer by layer through the four downsampling layers. The image input dimension of the input layer is 64×64×64, and the number of channels is increased to 16 using a 5×5×5 convolution kernel. After the first downsampling layer performs two convolutions, it performs unit addition with the output of the input layer through the residual operation, and then downsamples using a 2×2×2 convolution kernel with a stride of 2. After the next three downsampling layers perform three convolutions, they perform unit addition with the output of the previous downsampling layer through the residual operation, and then downsample using a 2×2×2 convolution kernel with a stride of 2.

[0083] The decoder includes four upsampling layers and an output layer connected in sequence. The four upsampling layers upsample the feature map output by the last downsampling layer layer by layer to output the feature map, and then the output layer performs unit addition with the feature map output by the last upsampling layer through the residual operation to output the segmentation result; after the first three upsampling layers perform three convolutions, they perform unit addition with the output of the previous upsampling layer through the residual operation, and then use a 2×2×2 convolution kernel with a step size of 2 for upsampling; after the last upsampling layer performs two convolutions, it performs unit addition with the output of the output layer through the residual operation, and then uses a 2×2×2 convolution kernel with a step size of 2 for upsampling; the image output dimension of the output layer is 64×64×64, and a 5×5×5 convolution kernel is used to reduce the number of channels to 1.

[0084] Embodiment 2:

[0085] The embodiment of the present invention provides a deep learning point cloud lumbar registration device for spinal minimally invasive surgery navigation, the device comprising:

[0086] Image acquisition module, used to acquire spinal CT images before and during surgery;

[0087] An image segmentation module is used to input the spinal CT image into the trained three-dimensional V-shaped deep segmentation network model for segmentation, and obtain the preoperative and intraoperative lumbar CT images;

[0088] An image processing module, used to eliminate isolated points in lumbar CT images by using a morphological method;

[0089] A point cloud conversion module, used for converting the preoperative and intraoperative lumbar CT images after eliminating isolated points into corresponding lumbar 3D point cloud sets through a point cloud conversion method;

[0090] The image registration module is used to perform nearest neighbor iteration on the preoperative and intraoperative lumbar 3D point cloud set to complete the lumbar registration in the preoperative and intraoperative spinal CT images.

[0091] Embodiment three:

[0092] Based on the first embodiment, the embodiment of the present invention provides a deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, including a processor and a storage medium;

[0093] The storage medium is used to store instructions;

[0094] The processor is used to operate according to the instructions to execute the steps according to the above method.

[0095] Embodiment 4:

[0096] Based on the first embodiment, the embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0097] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0098] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0099] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation, characterized in that: include: Collect preoperative and intraoperative spinal CT images; The spinal CT images are input into the trained three-dimensional V-shaped deep segmentation network model for segmentation to obtain preoperative and intraoperative lumbar CT images; The isolated points in lumbar spine CT images were eliminated by morphological methods; The preoperative and intraoperative lumbar spine CT images after eliminating isolated points are converted into corresponding lumbar spine three-dimensional point cloud sets by point cloud conversion method; Perform nearest neighbor iteration on the lumbar 3D point cloud set before and during surgery to obtain the optimal rotation matrix and translation vector; The three-dimensional point cloud set is transformed through the optimal rotation matrix and translation vector to complete the lumbar spine registration in the preoperative and intraoperative spinal CT images; Wherein, the three-dimensional V-shaped deep segmentation network model includes a symmetrical encoder and decoder; The encoder includes an input layer and four downsampling layers connected in sequence, the input layer performs unit addition with the input image through a residual operation, and then downsamples and outputs a feature map layer by layer through the four downsampling layers; the image input dimension of the input layer is 64×64×64, and the number of channels is increased to 16 using a 5×5×5 convolution kernel; after the first downsampling layer performs two convolutions, it performs unit addition with the output of the input layer through a residual operation, and then downsamples using a 2×2×2 convolution kernel with a step size of 2; after the last three downsampling layers perform three convolutions, they perform unit addition with the output of the previous downsampling layer through a residual operation, and then downsample using a 2×2×2 convolution kernel with a step size of 2; The decoder includes four upsampling layers and an output layer connected in sequence, the four upsampling layers upsample the feature map output by the last downsampling layer layer by layer to output the feature map, and then the output layer performs unit addition with the feature map output by the last upsampling layer through a residual operation to output the segmentation result; after the first three upsampling layers perform three convolutions, the residual operation is performed with the output of the previous upsampling layer, and then a 2×2×2 convolution kernel with a step size of 2 is used for upsampling; after the last upsampling layer performs two convolutions, the residual operation is performed with the output of the output layer, and then a 2×2×2 convolution kernel with a step size of 2 is used for upsampling; The image output dimension of the output layer is 64×64×64, and a 5×5×5 convolution kernel is used to reduce the number of channels to 1.

2. According to claim 1, a deep learning point cloud lumbar vertebra registration method for minimally invasive spinal surgery navigation is characterized in that: The training of the three-dimensional V-shaped deep segmentation network model includes: Obtain spinal CT images and add lumbar contour annotations to generate sample images, and build a training set based on the sample images; Based on the training set, the Adam algorithm is used to train the three-dimensional V-shaped deep segmentation network model, and the model parameters are adjusted until the loss function is minimized.

3. According to claim 2, a deep learning point cloud lumbar vertebra registration method for minimally invasive spinal surgery navigation is characterized in that: The adding of lumbar vertebrae contour annotation includes annotating voxels belonging to the lumbar vertebrae in the spinal CT image as 1, and annotating voxels not belonging to the lumbar vertebrae as 0; The loss function for: ; In the formula, is the voxel in the spine CT image The annotation information, is the voxel in the spine CT image The probability marked as 1; is the number of voxels in the spine CT image.

4. The deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation according to claim 1 is characterized in that: The method of eliminating isolated points in the lumbar CT image by using a morphological method includes: performing morphological opening and closing operations on the lumbar CT image, and adjusting the template size of the opening and closing operations according to the size of the isolated points to be segmented.

5. The deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation according to claim 1, characterized in that: The conversion into the corresponding lumbar 3D point cloud set includes: performing surface drawing on the preoperative and intraoperative lumbar CT images after eliminating isolated points using the Marching Cubes algorithm, traversing the vertex coordinates of the surface drawing results and converting them into point cloud coordinates, and obtaining the preoperative and intraoperative lumbar 3D point cloud set.

6. The deep learning point cloud lumbar registration method for minimally invasive spinal surgery navigation according to claim 1, characterized in that: The method of performing nearest neighbor iteration on the lumbar 3D point cloud set before and during surgery to obtain the optimal rotation matrix and translation vector comprises: Initialize the number of iterations ; 3D point cloud collection of the lumbar spine before surgery Pick point set ; Lumbar spine 3D point cloud collection during surgery Pick point set , so that Minimum, is the Euclidean distance; Construct the error function: ; The rotation matrix is ​​obtained by solving the minimum error function through the singular value decomposition method and translation vectors ; According to the rotation matrix and translation vectors Point Set Perform rotation and translation transformation to obtain point set ; Calculation point set And the corresponding point set The average distance ; Determine the average distance Is it less than the preset threshold? If so, stop the iteration and output the rotation matrix and translation vectors ; If not, determine the number of iterations Is it greater than the preset number of iterations? If so, stop the iteration and output the rotation matrix and translation vectors ; If not, let the number of iterations be , and repeat the above iterative steps.

7. A deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, characterized in that: The device comprises: Image acquisition module, used to acquire spinal CT images before and during surgery; An image segmentation module is used to input the spinal CT image into the trained three-dimensional V-shaped deep segmentation network model for segmentation, and obtain the preoperative and intraoperative lumbar CT images; An image processing module, used to eliminate isolated points in lumbar CT images by using a morphological method; A point cloud conversion module, used for converting the preoperative and intraoperative lumbar CT images after eliminating isolated points into corresponding lumbar 3D point cloud sets through a point cloud conversion method; The image registration module is used to perform nearest neighbor point iteration on the lumbar 3D point cloud set before and during surgery to complete the lumbar registration in the spinal CT images before and during surgery; Wherein, the three-dimensional V-shaped deep segmentation network model includes a symmetrical encoder and decoder; The encoder includes an input layer and four downsampling layers connected in sequence, the input layer performs unit addition with the input image through a residual operation, and then downsamples and outputs a feature map layer by layer through the four downsampling layers; the image input dimension of the input layer is 64×64×64, and the number of channels is increased to 16 using a 5×5×5 convolution kernel; after the first downsampling layer performs two convolutions, it performs unit addition with the output of the input layer through a residual operation, and then downsamples using a 2×2×2 convolution kernel with a step size of 2; after the last three downsampling layers perform three convolutions, they perform unit addition with the output of the previous downsampling layer through a residual operation, and then downsample using a 2×2×2 convolution kernel with a step size of 2; The decoder includes four upsampling layers and an output layer connected in sequence. The four upsampling layers upsample the feature map output by the last downsampling layer layer by layer to output the feature map, and then perform unit addition with the feature map output by the last upsampling layer through a residual operation through the output layer to output the segmentation result; after the first three upsampling layers perform three convolutions, they perform unit addition with the output of the previous upsampling layer through a residual operation, and then use a 2×2×2 convolution kernel with a step size of 2 for upsampling; after the last upsampling layer performs two convolutions, it performs unit addition with the output of the output layer through a residual operation, and then uses a 2×2×2 convolution kernel with a step size of 2 for upsampling; the image output dimension of the output layer is 64×64×64, and a 5×5×5 convolution kernel is used to reduce the number of channels to 1.

8. A deep learning point cloud lumbar registration device for minimally invasive spinal surgery navigation, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.