Path planning system and method for spine and intervertebral disc operation positioning

Abstract

The invention discloses a path planning system and method for spine and intervertebral disc operation positioning, and the method comprises the steps: constructing an orthogonal biplane feature through employing a plurality of orthogonal X-ray images, and carrying out the three-dimensional reconstruction of the orthogonal biplane feature, and obtaining a three-dimensional model; inputting the three-dimensional body model into a lesion extraction model constructed based on a deep learning algorithm to obtain a plurality of lesion positions of the three-dimensional body; and performing path planning on the plurality of lesion positions based on an A * algorithm to obtain a planned path. The invention relates to the technical field of medical image processing, and solves the technical problems of large positioning error and insufficient lesion detection precision of a traditional two-dimensional image.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing and relates to deep learning technology, specifically a path planning system and method for positioning spine and intervertebral disc surgery. Background Art

[0002] In spine and intervertebral disc surgery, accurate surgical positioning and path planning are the core links to ensure the safety and effectiveness of the surgery. Traditional technologies mainly rely on two-dimensional X-rays, CT or MRI images and the doctor's clinical experience for surgical positioning, which has significant technical bottlenecks:

[0003] First, the spatial information limitations of two-dimensional images lead to insufficient positioning accuracy. Existing methods use orthogonal biplane X-rays or CT images to manually interpret anatomical structures. This requires the physician to subjectively construct three-dimensional spatial relationships, which is susceptible to factors such as bone overlap and low tissue contrast. For example, in complex cases such as scoliosis and vertebral rotational deformity, two-dimensional images cannot accurately reflect the spatial proximity of the lesion to the surrounding nerves and blood vessels, leading to deviations in the positioning of surgical instruments and increasing the risk of spinal cord injury and vascular rupture.

[0004] Secondly, lesion detection and classification rely on manual interpretation, making accuracy difficult to guarantee. For minor lesions such as herniated discs, vertebral microfractures, and early-stage tumors, traditional methods are limited by image noise and resolution, resulting in a high rate of missed detection. Furthermore, they lack the ability to automatically distinguish lesion types, relying on physician experience to distinguish between calcifications, soft tissue protrusions, or bony abnormalities, which carries the risk of subjective misjudgment. Especially in the case of multi-segment lesions, manual analysis is time-consuming and labor-intensive, making it difficult to meet the efficient preoperative planning requirements of precision medicine.

[0005] Finally, surgical path planning lacks quantitative evaluation and intelligent optimization. Existing path planning often relies on the empirical principle of "shortest distance" and fails to assess high-risk areas such as nerves, blood vessels, and the spinal cord, which can easily lead to paths crossing high-risk areas. Furthermore, traditional surgical path planning relies on physician experience, which can lead to inconsistent and inefficient results due to differences in clinical experience, professional skills, and understanding of anatomical structures. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a path planning system and method for positioning of spine and intervertebral disc surgery, which is used to solve the technical problems of large errors in traditional two-dimensional image positioning and insufficient lesion detection accuracy.

[0007] To achieve the above objectives, a first aspect of the present invention provides a path planning method for positioning spine and intervertebral disc surgery, comprising:

[0008] Using multiple orthogonal X-ray images to construct orthogonal biplane features, and then reconstructing the orthogonal biplane features in three dimensions to obtain a three-dimensional body model;

[0009] The 3D model is input into a lesion extraction model built based on a deep learning algorithm to obtain several lesion locations in the 3D body;

[0010] Path planning is performed on several lesion locations based on the A* algorithm to obtain a planned path.

[0011] Furthermore, before constructing the orthogonal biplane features, the orthogonal X-ray images are pre-processed:

[0012] At least two orthogonal X-ray images are collected using X-ray anteroposterior and lateral imaging techniques to obtain multiple orthogonal X-ray images;

[0013] The Gaussian kernel function is used to perform weighted averaging on the orthogonal X-ray images to obtain the denoised image;

[0014] Adaptive histogram equalization is used to improve the contrast of the noise-reduced image to obtain a preprocessed orthogonal X-ray image.

[0015] Furthermore, the method of constructing orthogonal biplane features using multiple orthogonal X-ray images includes:

[0016] The scale-invariant feature transformation algorithm is used to extract the feature points in the pre-processed orthogonal X-ray image to obtain several feature points;

[0017] The feature points in different orthogonal X-ray images are matched by the nearest neighbor matching algorithm, and orthogonal biplane features are constructed according to the matched feature points to obtain orthogonal biplane features.

[0018] A scale-invariant feature transformation algorithm is used to extract feature points, and a nearest neighbor matching algorithm is used to match feature points across views to construct orthogonal biplane features. This can stably extract key points of the spinal structure, such as vertebral edges and articular processes; and ensure the precise correspondence of feature points under different viewing angles, establishing a geometric mapping relationship between two-dimensional images and three-dimensional space.

[0019] Furthermore, the three-dimensional reconstruction is based on a multi-layer perceptron and volume rendering technology, including:

[0020] Input the orthogonal biplane features into the multi-layer perceptron, extract the mapping relationship between the orthogonal biplane features and the three-dimensional space coordinates, and obtain the three-dimensional point cloud data;

[0021] The three-dimensional point cloud data is rendered into a continuous three-dimensional volume using volume rendering technology to obtain a converted three-dimensional volume model.

[0022] The multi-layer perceptron (MLP) is used to automatically learn the nonlinear mapping from two-dimensional features to three-dimensional coordinates to adapt to the geometric transformation of complex spinal structures; volume rendering technology can avoid the discretization errors of traditional surface modeling, making the generated three-dimensional model more continuous and smooth, and able to fully retain tissue density information.

[0023] Furthermore, the lesion extraction model constructed based on the deep learning algorithm includes:

[0024] The encoder includes a parallel 3D residual path and a spatial attention path for extracting multi-level feature maps of the 3D body model;

[0025] The three-dimensional residual path includes a multi-level downsampling module, each of which includes a plurality of residual blocks composed of deformable three-dimensional convolution, GroupNorm and GELU activation layers, to obtain an initial feature set A = {A1, A2, A3, A4, A5}; wherein A i (i=1,2,3,4,5) represent feature maps of different sizes respectively;

[0026] The spatial attention path is embedded in a 3D Transformer module after the A3 and A4 feature maps in the initial feature set, resulting in a feature map set {B3, B4} with spatial weights.

[0027] The feature map set {B3, B4} with spatial weights is combined with the feature maps A1, A2, and A5 in the initial feature set to form a multi-level feature map C = {C1, C2, C3, C4, C5}; where C1 = A1, C2 = A2, C3 = B3, C4 = B4, and C5 = A5;

[0028] The feature fusion module includes anisotropic three-dimensional convolution combination and dynamic channel weighting mechanism, which is used to perform cross-scale fusion on the multi-level feature maps to obtain a multi-scale fusion feature set F fusion ={S1,S2,S3,S4,S5};

[0029] The decoder includes a hierarchical upsampling structure and a spatially constrained attention gate, which is used to reconstruct the multi-scale fusion features to the original resolution of the 3D model to obtain a segmentation probability map P; wherein,

[0030] The hierarchical upsampling structure includes multiple upsampling modules, each of which includes a transposed convolution. The output of the transposed convolution is jump-connected to a feature map of the same size in the encoder's multi-scale fusion feature set, and then used as the input of the next upsampling module. The expression is: D=Up(Up(Up(S5)+S4)+S3)+S2)+S1; where D represents the output of the hierarchical upsampling structure, and the resolution of D is the same as the original resolution of the 3D model. Up(·) represents the upsampling module.

[0031] The spatially constrained attention gate is to align the resolution of the prior mask with the output feature map of the upsampling structure through three-dimensional bilinear interpolation, and then perform three-dimensional convolution and Sigmoid function calculation on the output feature map of the upsampling structure to obtain a segmentation probability map P, which is expressed as: P = Sigmoid(Conv3D([D,Mask])); where Mask represents a predefined pyramidal binary prior mask, and Conv3D(·) represents a three-dimensional convolution;

[0032] The output module includes a three-dimensional candidate region network 3D-RPN, which is used to extract the coordinates of the candidate box containing the lesion area and the lesion type from the segmentation probability map; wherein,

[0033] The process of obtaining the candidate frame coordinates includes: using the formula: Convert the voxel coordinates of the candidate frame output by 3D-RPN into physical coordinates (X, Y, Z); where (X origin ,Y origin ,Z origin ) represents the origin coordinates of the 3D model, (Δx, Δy, Δz) represents the voxel spacing of the 3D model, (X v ,Y v ,Z v ) represents the voxel coordinates of the candidate box;

[0034] The process of acquiring the lesion type includes: using a three-dimensional convolution classifier based on ROI Align features to output the lesion type of the cubic area corresponding to the candidate box coordinates.

[0035] It should be noted that the lesion extraction model is applied in the present invention after training and verification.

[0036] In the lesion extraction model, the encoder extracts local detail features and global context information through parallel three-dimensional residual paths and spatial attention paths, respectively. The three-dimensional residual path uses deformable convolution to adapt to complex anatomical structures, while the spatial attention path assigns dynamic weights to feature maps through the Transformer module, enhancing the model's sensitivity to subtle lesions. The dynamic channel weighting mechanism generates channel weights through global pooling and multi-layer perceptrons, optimizing the fusion effect of multi-scale features, enabling the model to accurately identify the location and type of lesions, reducing missed detections or false detections, and effectively improving detection accuracy.

[0037] Furthermore, the feature fusion module in the lesion extraction model includes anisotropic three-dimensional convolution combination and dynamic channel weighting mechanism, wherein,

[0038] The input of the anisotropic 3D convolution is a multi-level feature map C = {C1, C2, C3, C4, C5}, which includes the following operations:

[0039] Each level feature map C is processed by 1×1×1 convolution i Perform feature alignment to obtain standardized feature C i ':C′ i =Conv3D 1×1×1 (C i );

[0040] For the standardized feature C i Apply 5×3×3, 3×5×3, and 3×3×5 three-dimensional convolution kernels in parallel to extract sagittal features, coronal features, and cross-sectional features, respectively, and perform channel splicing to obtain the splicing features of each level.

[0041] C i+1 'Upsample to C by trilinear interpolation i ' resolution, and perform 1×1×1 three-dimensional convolution to obtain the predicted offset Δi; where i=1,2,3,4, when i=5, Δi=C i ';

[0042] Perform deformable 3D convolution on the spliced features to obtain the interactive features of each level Among them, DeformConv3D(·) represents deformable three-dimensional convolution;

[0043] The input of the dynamic channel weighting mechanism is the interaction features of each level The following operations are included:

[0044] Interaction features Perform global average pooling and maximum pooling respectively, and input them into the two-layer multi-layer perceptron MLP to obtain the dynamic weight of each level Among them, F avg 、F max They represent the global average pooling result and the maximum pooling result respectively, and σ(·) represents the Sigmoid function;

[0045] Interaction Features Channel compression is performed through 1×1×1 convolution to obtain compressed features After that, with dynamic weight Perform channel-by-channel point multiplication to obtain the weighted fusion result S i :

[0046] Output the weighted fusion results of each level to obtain the multi-scale fusion feature set F fusion ={S1,S2,S3,S4,S5}.

[0047] In the feature fusion module, anisotropic convolution designs different convolution kernels according to the anatomical direction of the spine, which can efficiently extract multi-planar structural features; dynamic channel weighting adaptively adjusts feature weights according to the context, strengthens lesion-related features, and can improve the lesion extraction model's ability to detect subtle lesions, avoiding the risk of path deviation in surgical path planning due to lesion positioning deviation.

[0048] Furthermore, the path planning for several lesion positions based on the A* algorithm includes:

[0049] Construct a three-dimensional risk field function φ(x, y, z) based on predefined risk locations: Among them, p = (x, y, z), represents the coordinates of the three-dimensional space point, c k represents the center coordinate of the kth risk location, σ k represents the risk attenuation radius, ω k represents the weight of the k-th risk position;

[0050] The state transition cost function g(n) of the A* algorithm is defined as: Among them, α, β, and δ represent the weight coefficients of each item respectively; among them, φ(q i ) represents the path point q i The three-dimensional risk field value at is calculated by the three-dimensional risk field function. is the indicator function, representing the path point q i Lesion at the jth lesion location j When inside, the value is 1;

[0051] The heuristic function h(n) of the A* algorithm is defined as: in represents the centroid coordinates obtained from the j-th candidate box coordinates output by the lesion extraction model, L2 represents the Euclidean norm, which is used to measure the distance between two points in space;

[0052] The A* algorithm is used to search the path from the predefined surgical entrance position to several lesion positions in the three-dimensional space grid to obtain a sequence of path points. Wherein, the three-dimensional space grid is obtained by discretizing the three-dimensional body model;

[0053] The objective function is defined as: F(Q) = γ1L(Q) + γ2C(Q); where, represents the path length cost term, represents the curvature penalty term, γ1 and γ2 represent the weight coefficients of each term, which are obtained based on experimental calibration;

[0054] Use the objective function to optimize the path point sequence Q and obtain the optimized path;

[0055] The optimized path is smoothed by B-spline to obtain the smoothed surgical path.

[0056] Path planning based on the A* algorithm, combined with a predefined three-dimensional risk field function, quantifies the distance relationship between the surgical path and the risk area. Path length, risk field value, and curvature penalty terms are introduced into the state transition cost function and objective function to ensure that the planned path avoids high-risk areas while minimizing length and curvature. Finally, B-spline smoothing is used to ensure that the resulting path conforms to the operating characteristics of the surgical instrument while minimizing damage to normal tissue.

[0057] Furthermore, the objective function includes a path length cost term and a curvature penalty term, wherein,

[0058] The expression of the path length cost term is:

[0059] The expression of the curvature penalty term is:

[0060] A first aspect of the present invention provides a path planning system for positioning spine and intervertebral disc surgery, comprising:

[0061] Model construction module: used to construct orthogonal biplane features using multiple orthogonal X-ray images, and perform three-dimensional reconstruction of the orthogonal biplane features to obtain a three-dimensional model;

[0062] Lesion localization module: used to input the 3D model into the lesion extraction model built based on the deep learning algorithm to obtain several lesion locations and lesion types in the 3D body;

[0063] Path planning module: used to perform path planning for several lesion locations based on the A* algorithm to obtain the planned path.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] Traditional technologies rely on two-dimensional images and physician experience, resulting in missing spatial information and subjective errors. This invention uses three-dimensional reconstruction technology from multiple orthogonal X-ray images, combined with a multi-layer perceptron and volume rendering method, to construct a high-precision three-dimensional model that fully restores the spinal anatomical structure and the spatial relationship between the lesion and surrounding tissue. Furthermore, the deep learning-based lesion extraction model uses an encoder and a dynamic channel weighting mechanism to effectively integrate multi-scale features, significantly improving the detection accuracy and type discrimination capabilities of small lesions and providing a reliable anatomical basis for path planning.

[0066] Existing surgical path planning relies on empirical judgment and lacks quantitative assessment, which can easily lead to lengthy paths and the risk of accidental injury. This paper proposes a three-dimensional path planning framework based on the A* algorithm. By defining a three-dimensional risk field function to quantify the avoidance weights of key areas such as blood vessels and nerves, and combining a dual optimization mechanism of a state transition cost function and an objective function, it generates the shortest, safest, and smoothest surgical path. B-spline curve post-processing further ensures that the path conforms to the operating characteristics of surgical instruments. This makes it more suitable for complex spinal deformities or minimally invasive surgery scenarios, and has important clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 A schematic diagram of a framework of a path planning system for positioning spine and intervertebral disc surgery provided by the present invention;

[0069] Figure 2 A schematic diagram of a process flow of a path planning system for positioning spine and intervertebral disc surgery provided by the present invention;

[0070] Figure 3 A schematic diagram of the process of constructing a three-dimensional model using multiple orthogonal X-ray images provided by the present invention. DETAILED DESCRIPTION

[0071] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0072] The path planning method for positioning spine and intervertebral disc surgery provided in the embodiment of the present application can be applied to Figure 1 The path planning system shown in Figure 1 is as follows. Figure 1 As shown, the system includes: a model building module, a lesion positioning module and a path planning module.

[0073] Among them, the model construction module is used to use multiple orthogonal X-ray images to construct orthogonal biplane features, and to perform three-dimensional reconstruction on the orthogonal biplane features to obtain a three-dimensional body model.

[0074] The lesion localization module is used to input the three-dimensional body model into the lesion extraction model built based on the deep learning algorithm to obtain several lesion locations and lesion types of the three-dimensional body.

[0075] The path planning module is used to perform path planning for several lesion locations based on the A* algorithm to obtain a planned path.

[0076] To address existing technical issues such as large positioning errors and blind surgical path planning caused by the reliance on two-dimensional imaging and surgeon experience in traditional spinal and intervertebral disc surgery, this application provides a path planning method. This method achieves precise surgical path planning through three-dimensional reconstruction, accurate lesion extraction, and intelligent path planning, effectively improving surgical safety and accuracy.

[0077] like Figure 2 As shown, the path planning method provided in the embodiment of the present application includes:

[0078] S1. Use multiple orthogonal X-ray images to construct orthogonal biplane features, and perform three-dimensional reconstruction on the orthogonal biplane features to obtain a three-dimensional body model.

[0079] Among them, constructing orthogonal dual-plane features is the basis of three-dimensional reconstruction, which aims to establish a geometric mapping relationship between two-dimensional images and three-dimensional space by matching multi-view image features.

[0080] In some implementations, the orthogonal X-ray images need to be preprocessed before constructing the orthogonal biplane features:

[0081] First, at least two orthogonal X-ray images are collected using X-ray anteroposterior and lateral imaging techniques to obtain multiple orthogonal X-ray images. Then, the orthogonal X-ray images are subjected to weighted averaging noise reduction using a Gaussian kernel function to obtain denoised images. Finally, adaptive histogram equalization is used to improve the contrast of the denoised images to obtain preprocessed orthogonal X-ray images.

[0082] It should be pointed out that the purpose of preprocessing is to improve image quality, reduce noise and enhance contrast to ensure the accuracy of subsequent feature extraction.

[0083] In a possible implementation of the embodiment of the present application, combined with Figure 2 ,like Figure 3 As shown, step S1 can be specifically implemented by the following steps:

[0084] S101. Extract feature points from the preprocessed orthogonal X-ray image using a scale-invariant feature transform algorithm to obtain a number of feature points containing key point position, scale, and orientation information. The scale-invariant feature transform algorithm (SIFT) can stably extract key points of spinal structures, such as vertebral edges and articular processes, and is scale and rotation invariant, making it suitable for feature extraction from images at different viewing angles.

[0085] The accuracy of feature points directly affects the subsequent feature matching and 3D reconstruction accuracy, so it is necessary to ensure the positioning accuracy and robustness of feature points;

[0086] In some implementations, a stable feature point set can be obtained by performing steps such as Gaussian pyramid construction, extreme point detection, and feature point descriptor generation on an image using the SIFT algorithm.

[0087] S102. Match feature points in different orthogonal X-ray images using a nearest neighbor matching algorithm, and construct orthogonal biplane features based on the matched feature points to obtain orthogonal biplane features. The nearest neighbor matching algorithm calculates the Euclidean distance between feature point descriptors and considers the closest feature point pairs as matching points, thereby ensuring accurate correspondence between feature points under different viewing angles.

[0088] In some implementations, a bidirectional matching strategy may be used, that is, performing nearest neighbor matching from image A to image B and from image B to image A, respectively, and taking the intersection as the final matching point pair to improve matching accuracy;

[0089] The matched feature point pairs constitute orthogonal biplane features, providing the necessary corresponding point pairs for 3D reconstruction and solving the ambiguity problem of single-view positioning.

[0090] S103, inputting the orthogonal biplane features into a multilayer perceptron, extracting the mapping relationship between the orthogonal biplane features and the three-dimensional space coordinates, and obtaining three-dimensional point cloud data; and rendering the three-dimensional point cloud data into a continuous three-dimensional volume using volume rendering technology, and obtaining a converted three-dimensional volume model;

[0091] Among them, the multi-layer perceptron (MLP) adopts a multi-layer fully connected network, with the input being the two-dimensional coordinates of orthogonal dual-plane features and the output being the corresponding three-dimensional space coordinates. An accurate mapping model is obtained through training with a large amount of labeled data; volume rendering uses ray casting to generate a continuous three-dimensional volume model from three-dimensional point cloud data.

[0092] Based on the above technical solution, through the combination of SIFT feature extraction, nearest neighbor matching, MLP and volume rendering technology, reconstruction from orthogonal X-ray images to high-precision three-dimensional volume models is achieved, providing reliable three-dimensional geometric information for subsequent lesion positioning and path planning.

[0093] S2. Input the 3D model into a lesion extraction model built based on a deep learning algorithm to obtain several lesion locations and lesion types in the 3D volume. The lesion extraction model is the core of achieving accurate lesion localization. By fusing multi-scale features and a spatial attention mechanism, it improves the detection capability of small lesions. Specifically, it includes an encoder, a feature fusion module, a decoder, and an output module.

[0094] In some implementations, the lesion extraction model is trained using 500 clinical X-ray image data, employing a joint optimization of Dice loss and cross-entropy loss. In the encoder, the 3D residual path adapts to complex anatomical structures through deformable convolution, while the spatial attention path utilizes the Transformer module to enhance global context awareness. The feature fusion module's anisotropic convolution designs different kernel functions for spinal anatomical orientations, while the dynamic channel weighting mechanism generates channel weights through global pooling and MLP to enhance lesion-related features. The decoder improves segmentation accuracy through hierarchical upsampling and spatially constrained attention gates, combined with vertebral prior masks. The output module generates candidate boxes using a 3D-RPN, converts voxel coordinates to physical coordinates, and classifies lesion types, outputting lesion location coordinates and lesion type to provide clear guidance for path planning.

[0095] In the embodiments of this application, the lesion extraction model is the core technology for achieving accurate lesion localization. Its architecture is designed to meet the needs of complex spinal anatomical structures and small lesion detection. It adopts an encoder-decoder framework, combined with a three-dimensional residual network, spatial attention mechanism, multi-scale feature fusion and hierarchical supervision strategy. The specific details are as follows:

[0096] (1) Encoder

[0097] The encoder consists of a 3D residual path and a spatial attention path to extract local detail features and global context information, specifically:

[0098] 1. 3D residual path, including a multi-level downsampling module: Input 3D volume model with a resolution of H×W×D×C (height×width×depth×number of channels), and gradually extract multi-scale features through 5 levels of downsampling:

[0099] Level 1: Directly input the original pre-processed 3D model, and output the feature map A1 through 1×1×1 convolution, with a resolution of H×W×D and 64 channels;

[0100] Level 2-5: Each level contains 2 residual blocks, and the structure of each residual block is:

[0101] Deformable 3D convolution: kernel size 3×3×3, dilation rate 1, the number of output channels is the same as the number of input channels; → GroupNorm: group number 8, used to normalize the spatial dimensions → GELU activation function → Skip connection: add the input features of the residual block to the output features;

[0102] Each level of downsampling is achieved through three-dimensional convolution with a stride of 2, and the initial feature set A = {A1, A2, A3, A4, A5} is finally generated. The resolution is 1 / 1, 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image, and the number of channels is 64, 128, 256, 512, and 1024 respectively.

[0103] It should be noted that the deformable 3D convolution introduces an offset prediction network based on the traditional convolution to address the irregular shape of the spinal structure. Its expression is: Among them, p k is the original convolution kernel sampling point, that is, the sampling position of the traditional convolution kernel when it is not deformed, Δp k To predict the offset, a 1×1×1 convolution is used to predict the sample point so that it can dynamically adapt to the anatomical structure boundary. x is the input feature of the deformable 3D convolution, p is the current pixel coordinate, w(k) is the weight of the convolution kernel, and y(p) represents the output feature of the 3D convolution. K = 27, that is, a 3×3×3 convolution kernel.

[0104] 2. Spatial attention path, embedded in the 3D Transformer module:

[0105] Insert a 3D Transformer module after A3 (1 / 4 resolution, 256 channels) and A4 (1 / 8 resolution, 512 channels) to strengthen long-range spatial dependencies:

[0106] (1) Multi-head self-attention: Expand the feature map into a sequence Generate query vector Q, key vector K, and value vector V through three sets of linear transformations, and calculate attention weights The number of self-attention heads can be set to 4, the dimension of each head is 32, and the number of output channels remains unchanged;

[0107] (2) Position encoding: Perform sine and cosine encoding on the three-dimensional spatial coordinates (x, y, z) to generate a position embedding vector, which is added to the feature sequence and then input into the Transformer;

[0108] Finally, the feature map {B3, B4} with spatial weights is output, which together with the feature maps A1, A2 and A5 form a multi-level feature map C = {C1, C2, C3, C4, C5}, and the resolution and number of channels correspond to the A sequence.

[0109] (2) Feature Fusion Module

[0110] The feature fusion module includes cross-scale interaction and channel weighting, which is used to perform cross-scale fusion on the multi-level feature maps to obtain a multi-scale fusion feature set F fusion , specifically including:

[0111] 1. Anisotropic 3D convolution combination: Based on the anatomical directionality of the spine (sagittal, coronal, and transverse planes), different convolution kernels are designed to extract multi-planar features:

[0112] (1) Feature alignment: For each level of feature map C i First, perform 1×1×1 convolution, unify the number of channels to 256, and obtain the standardized feature C i ', formula: C' i =Conv3D 1×1×1 (C i );

[0113] (2) Multi-plane feature extraction: Three convolution kernels are applied in parallel:

[0114] Sagittal plane (YZ plane): 5 × 3 × 3 kernel (expanded along the X axis to extract longitudinal structures);

[0115] Coronal plane (XZ plane): 3 × 5 × 3 kernel (expanded along the Y axis to extract anterior and posterior structures);

[0116] Cross-section (XY plane): 3 × 3 × 5 kernels (expanded along the Z axis to extract transverse structures);

[0117] The number of output feature channels is 256, and after splicing, we get

[0118] (3) Deformable convolution adjusts the receptive field: i+1 'Perform trilinear interpolation upsampling to C i'resolution, the offset Δi of the deformable convolution is predicted by 1×1×1 convolution, according to the formula: Output interaction features Among them, DeformConv3D(·) represents deformable three-dimensional convolution;

[0119] 2. Dynamic channel weighting mechanism

[0120] (1) Global feature statistics: interactive features Perform global average pooling and maximum pooling respectively, formula:

[0121] (2) Weight generation and feature fusion: The pooling results are processed by a two-layer MLP and the dynamic weights are generated by Sigmoid normalization.

[0122] Interaction Features Perform 1×1×1 convolution to compress the channel and get and Channel-by-channel point multiplication, output weighted fusion features Finally, a multi-scale fusion feature set F is formed fusion ={S1,S2,S3,S4,S5}.

[0123] (3) Decoder

[0124] The decoder includes a hierarchical upsampling structure and a spatially constrained attention gate to reconstruct the multi-scale fusion features to the original resolution of the 3D model to obtain a segmentation probability map P, which specifically includes:

[0125] 1. Hierarchical upsampling structure: Starting from the minimum resolution S5 (1 / 16 size, 128 channels), it is restored to the original resolution through a four-level upsampling module:

[0126] Use transposed convolution for upsampling: kernel size 4×4, stride 2, padding 1, the number of output channels is the same as the input; then add the result of transposed convolution to F fusion Same size feature map S i Perform channel splicing and use it as the input of the next level upsampling module;

[0127] Finally, the feature map D with the same resolution as the original 3D model is obtained, which is expressed as: D = Up(Up(Up(Up(S5)+S4)+S3)+S2)+S1); where Up(·) represents the upsampling module;

[0128] 2. Spatially constrained attention gate: First, a pyramidal binary mask is predefined and its resolution is adjusted to be consistent with D through 3D bilinear interpolation, highlighting the pyramidal region with a value of 1 and suppressing the background region with a value of 0.

[0129] D is then concatenated with Mask, and the segmentation probability map P is calculated through a 3×3×3 convolution and a Sigmoid function: P = Sigmoid(Conv3D([D,Mask])). By focusing on the vertebral body and its surrounding lesion area, the interference of paravertebral tissue can be reduced, and the segmentation accuracy of fine structures such as the intervertebral disc and vertebral endplate can be improved.

[0130] (4) Output module

[0131] The output module includes a 3D candidate region network (3D-RPN) for lesion detection and classification, which can include:

[0132] 1. 3D-RPN

[0133] (1) Anchor frame preset: Three anchor frames are preset for common sizes of spinal lesions:

[0134] Small size: 16×16×16, suitable for calcifications less than 5mm;

[0135] Medium size: 32×32×32, suitable for 5-15mm disc herniation;

[0136] Large size: 64×64×64, suitable for vertebral tumors larger than 15mm;

[0137] And using a cube with a length-width-height ratio of 1:1:1, symmetrical lesions in the sagittal, coronal, and transverse planes were covered;

[0138] (2) Coordinate transformation and physical positioning

[0139] The voxel coordinates (X v ,Y v ,Z v ), converted to physical coordinates (X, Y, Z) through DICOM header file parameters: Where (X origin ,Y origin ,Z origin ) is the origin coordinate of the 3D model, usually located in the upper left corner, and (Δx, Δy, Δz) is the voxel spacing;

[0140] 2. Lesion type classifier

[0141] (1) ROIAlign feature extraction: Perform 3D ROIAlign on the candidate box area and output a fixed-size 8×8×8 feature block to avoid quantization error;

[0142] (2) 3D convolution classification: Deep features are extracted through three layers of 3D convolution with a kernel size of 3×3×3 and channel numbers of 128, 64, and 32, respectively. After global average pooling, the softmax layer is input to output the probability of five types of lesions: 0: intervertebral disc herniation; 1: vertebral fracture; 2: spinal canal stenosis; 3: vertebral tumor; 4: normal tissue.

[0143] Through the above model design, the lesion extraction model achieves accurate output from the three-dimensional model to the lesion location and type. Its encoding structure can effectively integrate local details and global context. The dynamic feature fusion and spatial constraint mechanism significantly improve the detection performance in complex anatomical environments, providing a reliable anatomical basis for the subsequent A* algorithm path planning.

[0144] S3. Perform path planning for several lesion locations based on the A* algorithm to obtain a planned path.

[0145] Among them, the A* algorithm is combined with the three-dimensional risk field function to optimize the path length and curvature while avoiding high-risk areas, ensuring the safety and feasibility of the path.

[0146] In some implementations, path planning includes steps such as constructing a three-dimensional risk field function, defining a state transition cost function and a heuristic function, searching for a path point sequence, optimizing the path, and performing smoothing. Specifically, the steps include:

[0147] The three-dimensional risk field function φ(x, y, z) quantifies the risk value through predefined risk locations (such as nerve and blood vessel areas) and guides the algorithm to avoid dangerous areas. Specifically: Where p = (x, y, z) represents the coordinates of a point in three-dimensional space, c k represents the center coordinate of the kth risk location, σ k represents the risk attenuation radius, ω k represents the weight of the k-th risk position;

[0148] The state transition cost function g(n) of the A* algorithm is defined as: Among them, α, β, and δ represent the weight coefficients of each item, φ(q i ) represents the path point q i The three-dimensional risk field value at is calculated by the three-dimensional risk field function. is the indicator function, representing the path point q i Lesion at the jth lesion location j If it is inside, the value is 1, otherwise the value is 0;

[0149] The heuristic function h(n) of the A* algorithm is defined as: in represents the centroid coordinates obtained from the j-th candidate box coordinates output by the lesion extraction model, represented by L2;

[0150] The A* algorithm is used to search the path from the predefined surgical entrance position to several lesion positions in the three-dimensional space grid to obtain a sequence of path points.

[0151] The objective function is defined as: F(Q) = γ1L(Q) + γ2C(Q); where, represents the path length cost term, represents the curvature penalty term, γ1 and γ2 represent the weight coefficients of each term, which are obtained based on experimental calibration;

[0152] Use the objective function to optimize the path point sequence Q and obtain the optimized path;

[0153] The optimized path is smoothed using B-spline technology to make it conform to the operating characteristics of surgical instruments, reduce the risk of damage to normal tissues, and obtain a smoothed surgical path.

[0154] It should be noted that the three-dimensional space grid is obtained by discretizing the three-dimensional body model, which may include:

[0155] First, the basic parameters of the grid are determined based on the physical voxel spacing (Δx, Δy, Δz) and resolution of the 3D model, such as H×W×D, to ensure that the grid nodes are aligned with the voxel positions.

[0156] Then, the physical origin of the 3D model (X origin ,Y origin ,Z origin ) as the grid origin, and combining the voxel spacing and data size, calculate the grid boundary coordinates (X max , Y max , Z max ), generating a 3D spatial extent covering the entire anatomical region:

[0157] During the specific construction, a three-dimensional array of dimensions H×W×D is first initialized as a grid carrier. Each array element corresponds to a grid node, storing its physical coordinates (x, y, z) and risk field value φ(x, y, z). The node coordinates are calculated by the origin coordinates and the voxel index, that is, x = X origin +i·Δx,y=Y origin +j·Δy,z=Y origin +k·Δz, where i, j, and k are voxel indices. At the same time, the adjacency relationship between nodes is defined, with a preference for twenty-four neighborhood connections. This includes up-down, front-back, left-right, and diagonal directions to accommodate the multi-angle requirements of spinal surgical paths. Bone hard tissue or surgical restricted areas are marked as inaccessible nodes to ensure automatic algorithm avoidance.

[0158] Next, based on predefined key risk structures (such as spinal cord, nerve roots, aorta, etc.), the risk field value of each node is calculated by superposition of Gaussian kernel functions. Finally, the coordinates of the lesion centroid (physical coordinates) output by the lesion extraction model are back-calculated into the grid index i using the voxel spacing: Determine the target node of the path planning so that the end point corresponds to the core position of the lesion.

[0159] During path planning, the weight of the risk field function is set based on the anatomical risk level, with the spinal cord receiving the highest weight, followed by nerve roots and blood vessels. The attenuation radius can be set to 1.5 times the diameter of the risk structure to ensure a safe distance between the path and critical structures. B-spline smoothing employs periodic boundary conditions, resampling path points to generate a parameterized curve. Curvature analysis removes sudden changes to ensure path smoothness meets clinical operational requirements.

[0160] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0161] Working principle of the present invention:

[0162] First, multiple orthogonal images are acquired using X-ray anteroposterior and lateral imaging techniques. After preprocessing with noise reduction and contrast enhancement, feature points are extracted using a scale-invariant feature transformation algorithm and orthogonal biplane features are constructed through nearest neighbor matching. These features are then reconstructed into a high-precision 3D volume model containing the spinal anatomical structure using a multi-layer perceptron and volume rendering technology.

[0163] Next, the 3D volume model is fed into a deep learning-based lesion extraction model. This model uses a hybrid encoder to fuse local detail features extracted by the 3D residual path with global contextual information captured by the spatial attention path. Combined with a feature fusion module and a hierarchical decoder, the model accurately locates the lesion and identifies the lesion type.

[0164] Finally, a three-dimensional risk field is constructed based on the predefined key risk structure, and the A* algorithm is used to search for a path from the surgical entrance to the lesion location in the three-dimensional spatial grid. During the process, the path length, risk field value, and curvature penalty are comprehensively considered for optimization, and the path is smoothed using B-spline processing to generate a final planned path that is safe, short, and consistent with the operating characteristics of surgical instruments.

[0165] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A path planning method for positioning spine and intervertebral disc surgery, characterized in that: include: Using multiple orthogonal X-ray images to construct orthogonal biplane features, and then reconstructing the orthogonal biplane features in three dimensions to obtain a three-dimensional body model; The 3D model is input into a lesion extraction model built based on a deep learning algorithm to obtain several lesion locations; Path planning is performed on several lesion locations based on the A* algorithm to obtain a planned path.

2. A path planning method for positioning spine and intervertebral disc surgery according to claim 1, characterized in that: Before constructing the orthogonal biplane features, the orthogonal X-ray images are pre-processed: At least two orthogonal X-ray images are collected using X-ray anteroposterior and lateral imaging techniques to obtain multiple orthogonal X-ray images; The Gaussian kernel function is used to perform weighted averaging on the orthogonal X-ray images to obtain the denoised image; Adaptive histogram equalization is used to improve the contrast of the noise-reduced image to obtain a preprocessed orthogonal X-ray image.

3. A path planning method for positioning spine and intervertebral disc surgery according to claim 1, characterized in that: The method of constructing orthogonal biplane features using multiple orthogonal X-ray images includes: The scale-invariant feature transformation algorithm is used to extract feature points from multiple orthogonal X-ray images to obtain several feature points. The feature points in different orthogonal X-ray images are matched by the nearest neighbor matching algorithm, and orthogonal biplane features are constructed according to the matched feature points to obtain orthogonal biplane features.

4. A path planning method for positioning spine and intervertebral disc surgery according to claim 1, characterized in that: The 3D reconstruction is based on a multi-layer perceptron and volume rendering technology, including: Input the orthogonal biplane features into the multi-layer perceptron, extract the mapping relationship between the orthogonal biplane features and the three-dimensional space coordinates, and obtain the three-dimensional point cloud data; The three-dimensional point cloud data is rendered into a continuous three-dimensional volume using volume rendering technology to obtain a converted three-dimensional volume model.

5. A path planning method for positioning spine and intervertebral disc surgery according to claim 1, characterized in that: The lesion extraction model constructed based on the deep learning algorithm includes: The encoder consists of a 3D residual path and a spatial attention path to extract multi-level feature maps of the 3D volume model; The feature fusion module includes anisotropic three-dimensional convolution combination and dynamic channel weighting mechanism, which is used to perform cross-scale fusion on the multi-level feature maps to obtain a multi-scale fusion feature set F fusion ={S1,S2,S3,S4,S5}; The decoder includes a hierarchical upsampling structure and a spatially constrained attention gate, which is used to reconstruct the multi-scale fusion features to the original resolution of the 3D model to obtain a segmentation probability map P; wherein, The hierarchical upsampling structure includes multiple upsampling modules, each of which includes a transposed convolution. The output of the transposed convolution is jump-connected to a feature map of the same size in the encoder's multi-scale fusion feature set, and then used as the input of the next upsampling module. The expression is: D=Up(Up(Up(S5)+S4)+S3)+S2)+S1; where D represents the output of the hierarchical upsampling structure, and the resolution of D is the same as the original resolution of the 3D model. Up(·) represents the upsampling module. The spatially constrained attention gate is to align the resolution of the predefined pyramidal binary prior mask Mask with the output feature map of the upsampling structure through three-dimensional bilinear interpolation, and then perform three-dimensional convolution and Sigmoid function calculation on the output feature map of the upsampling structure to obtain a segmentation probability map P, which is expressed as: P = Sigmoid(Conv3D([D,Mask])); wherein Conv3D(·) represents a three-dimensional convolution; The output module includes a three-dimensional candidate region network 3D-RPN, which is used to extract the candidate frame coordinates and lesion type containing the lesion area from the segmentation probability map; wherein, The process of obtaining the candidate frame coordinates includes: using the formula: Convert the voxel coordinates of the candidate frame output by 3D-RPN into physical coordinates (X, Y, Z); where (X origin ,Y origin ,Z origin ) represents the origin coordinates of the 3D model, (Δx, Δy, Δz) represents the voxel spacing of the 3D model, (X v ,Y v ,Z v ) represents the voxel coordinates of the candidate box; The process of acquiring the lesion type includes: using a three-dimensional convolution classifier based on ROI Align features to output the lesion type of the cubic area corresponding to the candidate box coordinates.

6. A path planning method for positioning spine and intervertebral disc surgery according to claim 5, characterized in that: The encoder in the lesion extraction model contains a 3D residual path and a spatial attention path, where The three-dimensional residual path includes a multi-level downsampling module, each of which includes a plurality of residual blocks composed of deformable three-dimensional convolution, GroupNorm and GELU activation layers, to obtain an initial feature set A = {A1, A2, A3, A4, A5}; wherein A i (i=1,2,3,4,5) represent feature maps of different sizes respectively; The spatial attention path is embedded in a 3D Transformer module after the A3 and A4 feature maps in the initial feature set, resulting in a feature map set {B3, B4} with spatial weights. The feature map set {B3, B4} with spatial weights is combined with the feature maps A1, A2 and A5 in the initial feature set to form a multi-level feature map C = {C1, C2, C3, C4, C5}; among them, C1 = A1, C2 = A2, C3 = B3, C4 = B4, C5 = A5.

7. A path planning method for positioning spine and intervertebral disc surgery according to claim 6, characterized in that: The feature fusion module in the lesion extraction model includes anisotropic three-dimensional convolution combination and dynamic channel weighting mechanism, wherein, The input of the anisotropic 3D convolution is a multi-level feature map C = {C1, C2, C3, C4, C5}, which includes the following operations: Each level feature map C is processed by 1×1×1 convolution i Perform feature alignment to obtain standardized feature C i ':C″ i =Conv3D 1×1×1 (C i ); For the standardized feature C i Apply 5×3×3, 3×5×3, and 3×3×5 three-dimensional convolution kernels in parallel to extract sagittal features, coronal features, and cross-sectional features, respectively, and perform channel splicing to obtain the splicing features of each level. C i+1 'Upsample to C by trilinear interpolation i ' resolution, and perform 1×1×1 three-dimensional convolution to obtain the predicted offset Δi; where i=1,2,3,4, when i=5, Δi=C i '; Perform deformable 3D convolution on the spliced features to obtain the interactive features of each level : Among them, DeformConv3D(·) represents deformable three-dimensional convolution; The input of the dynamic channel weighting mechanism is the interaction features of each level The following operations are included: Interaction features Perform global average pooling and maximum pooling respectively, and input them into the two-layer multi-layer perceptron MLP to obtain the dynamic weight of each level : Among them, F avg 、F max They represent the global average pooling result and the maximum pooling result respectively, and σ(·) represents the Sigmoid function; Interaction Features Channel compression is performed through 1×1×1 convolution to obtain compressed features After that, with dynamic weight Perform channel-by-channel point multiplication to obtain the weighted fusion result S i : Output the weighted fusion results of each level to obtain the multi-scale fusion feature set F fusion ={S1,S2,S3,S4,S5}.

8. A path planning method for positioning spine and intervertebral disc surgery according to claim 1, characterized in that: The path planning for several lesion locations based on the A* algorithm includes: Construct a three-dimensional risk field function φ(x, y, z) based on predefined risk locations: Among them, p = (x, y, z), represents the coordinates of the three-dimensional space point, c k represents the center coordinate of the kth risk location, σ k represents the risk attenuation radius, ω k represents the weight of the k-th risk position; The state transition cost function g(n) of the A* algorithm is defined as: Among them, α, β, and δ represent the weight coefficients of each item, φ(q i ) represents the path point q i The three-dimensional risk field value at is calculated by the three-dimensional risk field function. is the indicator function, representing the path point q i Lesion at the jth lesion location j When inside, the value is 1; Define the heuristic function h of the A* algorithm j (n) is: in represents the centroid coordinates of the jth lesion location output by the lesion extraction model, and L2 represents the Euclidean norm; The A* algorithm is used to search the path from the predefined surgical entrance position to several lesion positions in the three-dimensional space grid to obtain a sequence of path points. Wherein, the three-dimensional space grid is obtained by discretizing the three-dimensional body model; The objective function is defined as: F(Q) = γ1L(Q) + γ2C(Q); where, represents the path length cost term, represents the curvature penalty term, γ1 and γ2 represent the weight coefficients of each term respectively; Use the objective function to optimize the path point sequence Q and obtain the optimized path; The optimized path is smoothed by B-spline to obtain the smoothed surgical path.

9. A path planning method for positioning spine and intervertebral disc surgery according to claim 8, characterized in that: The objective function includes a path length cost term and a curvature penalty term, where: The expression of the path length cost term is: The expression of the curvature penalty term is:

10. A path planning system for positioning spine and intervertebral disc surgery, applied to a path planning method for positioning spine and intervertebral disc surgery according to claims 1-9, characterized in that: include: Model construction module: used to construct orthogonal biplane features using multiple orthogonal X-ray images, and perform three-dimensional reconstruction of the orthogonal biplane features to obtain a three-dimensional model; Lesion localization module: used to input the 3D model into the lesion extraction model built based on the deep learning algorithm to obtain several lesion locations and lesion types in the 3D body; Path planning module: used to perform path planning for several lesion locations based on the A* algorithm to obtain the planned path.

Images