Method, device and medium for spinal medical image registration under implant occlusion

Abstract

The present invention discloses a spinal medical image registration method, device and medium under implant occlusion, which is characterized by comprising the following steps: S100: inputting an intraoperative X-ray image with implant occlusion, and obtaining a FeatureMap through CNN; S200: extracting regional occlusion information from a single implant image through spatial pyramid pooling, and adding the regional occlusion information to the FeatureMap obtained in S100; S300: performing hierarchical decoupling processing on the implant area and the spinal area in the feature map through GCN to obtain hidden features; S400: using a posture encoder to extract spinal information, reconstructing the corresponding posture latent space, and then using MLP to regress the latent space to obtain the required posture parameters, projecting the preoperative CT with the posture parameters to obtain the final registration image. This method can accurately and quickly regress the corresponding 3D posture parameters in the preoperative CT image, improving the problem that traditional optimization methods cannot quickly perform registration.

Description

Technical Field

[0001] The present invention relates to the field of image registration under pedicle screw implantation during spinal orthopedic surgery, and in particular to a spinal medical image registration method, device and medium under the occlusion of an implant. Background Art

[0002] Pedicle screw placement is a common spinal treatment in minimally invasive spinal surgery. Accurately inserting the screw while avoiding the risk of nerve damage through the spinal canal or intervertebral foramen places extremely high demands on the surgeon's precision. Intraoperative X-ray images can be used to assess screw placement, but these are typically low-dose, resulting in limited clarity and inadequate visualization of the positional relationship between the screw and anatomical structures. Furthermore, capturing multiple X-rays can easily cause radiation damage to the patient. Utilizing 2D / 3D registration technology, blurred intraoperative X-ray images are used to generate clear, registered preoperative 3D CT images in the corresponding position. Finally, by fusing the X-ray image with the registered image, the surgeon can refer to the clear image to assess screw placement, improving surgical efficiency and reducing surgical risk.

[0003] Conventional optimization-based 2D / 3D registration methods, when performed without screw implantation, can achieve registration through iterative digital reconstruction (DRR), demonstrating high robustness and accuracy. Similarity measurement and DRR generation are two key components of optimized registration, but these two steps require extensive computation during the iterative process, taking up a significant amount of time during surgery and failing to meet the surgeon's need for real-time registration.

[0004] In recent years, deep learning-based methods have been gradually used in medical image registration tasks. These methods treat the registration task as a regression problem and train a model to estimate the pose parameters in the input image. However, the accuracy that can be achieved by this method is much lower than that achieved by using optimization-based methods.

[0005] When pedicle screws are implanted in the spine, the high pixel values of the screw images in intraoperative X-ray images obstruct the spinal area, affecting the similarity calculation during iterative optimization. This makes the registration process prone to falling into a local optimum and preventing further convergence. Learning-based methods use CNNs to automatically extract features, effectively improving registration efficiency. However, registration accuracy still needs to be improved, and under the influence of screw occlusion, registration accuracy is further compromised. Summary of the Invention

[0006] The technical problem solved by the present invention is to provide a spinal medical image registration method, device and medium under implant occlusion that can avoid the influence of screw occlusion on the extracted posture while quickly registering and ensure the registration accuracy.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a spinal medical image registration method under implant occlusion, characterized by comprising the following steps:

[0008] S100: Input an intraoperative X-ray image with implant occlusion and obtain the Feature Map through CNN.

[0009] S200: A single implant image is subjected to spatial pyramid pooling to extract regional occlusion information, and the regional occlusion information is added to the Feature Map obtained in S100 to obtain a feature map.

[0010] S300: performing hierarchical decoupling processing on the implant region and the spine region in the feature map in step S200 through GCN to obtain hidden features;

[0011] S400: Use the posture encoder to extract spinal information and reconstruct the corresponding posture latent space. Then use MLP to regress the latent space to obtain the required posture parameters. The preoperative CT is projected with the posture parameters to obtain the final registration image.

[0012] Furthermore, in step S200, in the SPP network, the screw image is first subjected to a convolution operation to obtain the corresponding feature map, and then three global average pooling layers are input, and the pooling pyramid parameters are defined as (1, 4, 8), with a step size of 1. Then, the overall features of the occluded image are concatenated with the feature vectors calculated by the three global average pooling layers and filled, and finally a convolution calculation is performed to form the final output.

[0013] Furthermore, step S300 is to use GCN as an occlusion processing module, where each graph node in the feature map represents a single pixel on the feature map, and the adjacency matrix between each pixel serves as the edge relationship between the nodes. Therefore, the graph convolution operation is expressed as:

[0014] Z=σ(W z AX)+X;

[0015] A=soft max(f(X i ,X j ));

[0016] f(X i ,X j )=θ(X i ) T φ(X j );

[0017] In the above formula, X is the input image feature, +X represents a residual connection, A represents the adjacency matrix that defines the adjacent relationship of the graph nodes through the dot product, and W Z is the learnable parameter matrix of the output transformation, and finally the output feature Z is obtained through a softmax function and a nonlinear function;

[0018] Among them, θ and φ are two trainable transformation functions implemented by 1×1 convolution to construct the adjacency matrix and graph nodes.

[0019] Furthermore, in step S400, a posture encoder is used to extract spinal information and construct an autoencoder architecture. The autoencoder architecture consists of four convolutions, each layer uses a 3×3 convolution kernel with a step size of 2, and each convolution is followed by a LeakyReLU layer with a parameter of 0.2. 2D convolution is applied in the encoding process, and spinal X-ray image features under different postures are input to obtain posture information, which is then connected to a 128->64->32->16->3 fully connected layer to regress the posture parameters.

[0020] Furthermore, when performing registration, specifically:

[0021]

[0022] In the above formula, I CT is the CT image, I X is the X-ray image, I metal is the occlusion image, is the posture regression process model of posture encoder + MLP, proj is DRR projection, I reg To register the image.

[0023] Furthermore, the registration network framework is implemented by a multi-loss function L total The end-to-end training is defined, and the overall objective function is:

[0024] L total =λ1L L1 +λ2L cosine +λ3L NCC

[0025] Among them, λ1, λ2 and λ3 are hyperparameters to balance the weights of different loss terms;

[0026] L L1 is the predicted posture parameter and the real posture r (i) The L1 norm loss is calculated as follows:

[0027]

[0028] Lcosine is the latent variable Z extracted from the X-ray image i The latent variable Z extracted from the DRR registration map i ''s cosine similarity loss. During the training process, the DRR image is obtained by unobstructed CT projection. The spinal posture features extracted from this image are free of occlusion interference, which is in contrast to the latent variables of the X-ray image with occlusion. The specific calculation is:

[0029]

[0030] L NCC is the similarity loss between the X-ray image and the registered image. The loss function can be expressed as:

[0031]

[0032] In the above formula, I Xl with I DRR Corresponding to the intensity values of X-ray image and DRR image in region (i, j), I Xl with I DRR is the mean of the images in the overlapping region (i, j)∈T.

[0033] The present invention also discloses a computer device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the above-mentioned spinal medical image registration method under implant occlusion.

[0034] The present invention also discloses a computer storage medium, wherein the computer storage medium stores at least one executable instruction, and the executable instruction enables a processor to execute the above-mentioned spinal medical image registration method under implant occlusion.

[0035] The beneficial effects of the present invention are:

[0036] (1) The present invention uses a posture encoder to predict posture parameters, avoiding the traditional iterative process and significantly shortening the registration time.

[0037] (2) The present invention introduces a graph convolutional network to process the occluded area of the implant, avoiding the influence of occlusion on posture extraction and ensuring the registration accuracy.

[0038] (3) The present invention uses a synthetic data set to train the model, and only CT data is needed to complete the entire model training.

[0039] (4) The present invention adopts a posture prediction method for registration, which can be applied to other orthopedic images except the spine. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a framework diagram of the 2D / 3D alignment method under implant occlusion in an embodiment of the present application.

[0041] Figure 2 This is a flowchart of the pyramid pooling (SPP) network of an embodiment of the present application.

[0042] Figure 3 This is a flowchart of the GCN non-local operation of an embodiment of the present application.

[0043] Figure 4 This is a flowchart of the implant synthesis dataset according to an embodiment of the present application.

[0044] Figure 5 These are comparative test diagrams of different metal implantations of the embodiments of the present application and other methods.

[0045] Figure 6 This is a diagram showing the module ablation experiment results of an embodiment of the present application.

[0046] Figure 7 This is a diagram of the model registration results under different occlusion area ranges in an embodiment of the present application.

[0047] Figure 8 This is a table showing the comparison results between the method of the present invention and other methods.

[0048] Figure 9 Data table for performance experiment analysis of each module.

[0049] Figure 10 mTRE (mm) results of grid search on (0, 0.5) for λ2 and λ3. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, the embodiment of the present application discloses a spinal medical image registration method under implant occlusion. During the training process, X-ray images in different postures are input into the network to obtain the 3D posture parameters of the corresponding rigid body. The posture parameters are used to perform DRR projection on the preoperative CT to obtain a 2D registration image. The method includes a registration network framework, which includes the following steps:

[0052] S100: Input an intraoperative X-ray image with implant occlusion and obtain the Feature Map through the Convolutional Neural Network (CNN).

[0053] S200: extracting regional occlusion information from a single implant image through spatial pyramid pooling, and adding the regional occlusion information to the Feature Map obtained in S100 to obtain a feature map;

[0054] Specifically, before the image features are fed into the GCN, the contours and positional features of the occluded object are extracted and added to the image's feature map, helping the GCN to clearly process the occluded area. Considering that the size of the metal implant varies from patient to patient during surgery, or that the same metal may appear at different scales in X-ray images taken in different postures, resulting in varying sizes of rectangular regions segmented by thresholding, a spatial pyramid pooling module is used.

[0055] The spatial pyramid pooling module can be used to solve the problem of repeated extraction of image features by CNN. It converts the image into multiple different feature blocks and aggregates the local features in each feature block. It can also convert two-dimensional input of any size into a fixed-size representation without being affected by the image size / scale, effectively avoiding the problem of distorting or cropping the input in general practices. This method extracts local metal feature information and then integrates it with the global image features to provide rich occlusion clues, such as the shape and position of the occluded area, thereby improving the detection accuracy of the occluded area.

[0056] Specifically, such as Figure 2 As shown, in step S200, in the SPP network, the screw image is first subjected to a convolution operation to obtain the corresponding feature map, and then three global average pooling layers are input, and the pooling pyramid parameters are defined as (1, 4, 8), with a step size of 1. Then, the overall features of the occluded image are concatenated with the feature vectors calculated by the three global average pooling layers and filled, and finally a convolution calculation is performed to form the final output.

[0057] S300: performing hierarchical decoupling processing on the implant region and the spine region in the feature map in step S200 through GCN to obtain hidden features;

[0058] Given highly overlapping objects, due to the non-local properties of GCN, non-local relationships between pixels can be considered, allowing information to be propagated across pixels in the presence of occluded areas, thereby locating the boundaries of the object and helping the Encoder avoid occluded areas during feature extraction.

[0059] Here, GCN is used as the occlusion processing module, where each graph node in the feature map represents a single pixel on the feature map, and the adjacency matrix between each pixel is used as the edge relationship between the nodes. Therefore, the graph convolution operation is expressed as:

[0060] Z=σ(W z AX)+X;

[0061] A=soft max(f(X i ,X j ));

[0062] f(X i ,X j )=θ(X i ) T φ(X j );

[0063] In the above formula, X is the input image feature, +X represents a residual connection, A represents the adjacency matrix that defines the adjacent relationship of the graph nodes through the dot product, and W Z is the learnable parameter matrix of the output transformation, and finally the output feature Z is obtained through a softmax function and a nonlinear function;

[0064] Among them, θ and φ are two trainable transformation functions implemented by 1×1 convolution to construct the adjacency matrix and graph nodes.

[0065] Specifically, in order to better classify or generate data, it is necessary to represent the features of the data. However, the data has many features, which are interrelated and highly coupled, making it difficult for the model to figure out the relationship between them, resulting in low learning efficiency. Therefore, it is necessary to find the deep relationships hidden under these surface features, decouple these relationships, and obtain hidden features. The latent space is a representation of compressed data. Its function is to learn data features and simplify data representation in order to find patterns. In the latent space feature representation, the difference features between similar sample images will be eliminated as non-core information, and only the core feature information will be retained and learned. Therefore, after the data feature points are mapped to the latent space, the distance between similar feature points will be closer. Therefore, the next step is:

[0066] S400: Use the posture encoder to extract spinal information and reconstruct the corresponding posture latent space. Then use MLP to regress the latent space to obtain the required posture parameters. The preoperative CT is projected with the posture parameters to obtain the final registration image.

[0067] Specifically, an autoencoder architecture is constructed. The autoencoder architecture consists of four convolutions. Each layer uses a 3×3 convolution kernel with a stride of 2. Each convolution is followed by a LeakyReLU layer with a parameter of 0.2. 2D convolution is applied during the encoding process. The spinal X-ray image features under different postures (e.g., 128×128×1) are input to obtain posture information. Then, the architecture is connected to a 128->64->32->16->3 fully connected layer to regress the posture parameters.

[0068] The attitude parameters here refer to the three rotation attitudes rx, ry, rz

[0069] In this embodiment, when performing registration, specifically:

[0070]

[0071] In the above formula, I CT is the CT image, I X is the X-ray image, I metal is the occlusion image, is the posture regression process model of posture encoder + MLP, proj is DRR projection, I reg To register the image.

[0072] In this embodiment, when the registration network framework is trained, the registration network framework is trained by the multi-loss function L total The end-to-end training is defined, and the overall objective function is:

[0073] L total =λ1L L1 +λ2L cosine +λ3L NCC

[0074] Among them, λ1, λ2 and λ3 are hyperparameters to balance the weights of different loss terms;

[0075] L L1 is the predicted posture parameter and the real posture r (i) The L1 norm loss is calculated as follows:

[0076]

[0077] L cosine is the latent variable Z extracted from the X-ray image i The latent variable Z extracted from the DRR registration map i ''s cosine similarity loss. During the training process, the DRR image is obtained by unobstructed CT projection. The spinal posture features extracted from this image are free of occlusion interference, which is in contrast to the latent variables of the X-ray image with occlusion. The specific calculation is:

[0078]

[0079] L NCC is the similarity loss between the X-ray image and the registered image. The loss function can be expressed as:

[0080]

[0081] In the above formula, I Xl with I DRR Corresponding to the intensity values of X-ray image and DRR image in region (i, j), I Xl with I DRR is the mean of the images in the overlapping region (i, j)∈T.

[0082] In order to verify the performance and advantages of this method, the following experiments were conducted:

[0083] First, three different similarity measurement methods were combined with stochastic gradient descent (SGD) to form three optimization-based methods: Opt-GC (gradient correlation), Opt-NCC (normalized cross-correlation), and Opt-MI (mutual information). Opt-C2F, an optimization method combining "coarse and fine" registration, was also compared. When the initial pose differs significantly from the groundTruth (true value), MSE (mean square error) is more effective. As the DRR image and X-ray image gradually align, this loss captures a relatively large range, lacks fine accuracy, and becomes less sensitive to subtle differences. In this case, NCC refined registration is applied. Furthermore, comparisons were made with occlusion-based repair methods such as Repair. This method uses artificial intelligence algorithms to repair occluded areas on 2D images, making the image information more complete and thus extracting features from the entire image.

[0084] The experimental data used the public CTSpine1K spine dataset released by Beijing Jishuitan Hospital in China, and 80 patient CT scans were selected for the experiment. Figure 4 The process shown here synthesizes the required X-ray image dataset and the corresponding metal region dataset. First, a DRR image is projected within the (-20°, 20°) rotation range of the CT spine coronal plane. Occlusions are then integrated into the image (given that the pedicle screws implanted during spinal surgery are made of metal, metal implants are extracted from other orthopedic datasets containing metal). To fully account for the characteristics of intraoperative X-ray images, the fused image is Gaussian blurred to ensure that the synthesized image is as consistent as possible with the X-ray image, allowing the trained registration network to generalize well to real X-ray images. Finally, a threshold-based segmentation method is used to extract the metal region image from the synthesized image.

[0085] In the same experimental environment, 200 registrations were performed on the CT scans of each patient, and a quantitative comparison was made in terms of registration accuracy, registration success rate, and registration time. The results are shown in Table 1.

[0086] (1) The mTRE and MAE values of the optimization-based methods (Opt-GC, Opt-NCC, Opt-MI) all have certain errors. Their registration process is prone to falling into local optimality. In addition, the interference caused by metal implantation increases the probability of occurrence, so the registration success rate is low. The Opt-C2F method uses different losses according to the convergence of different loss stages to avoid local optimal values. Therefore, its registration accuracy and registration success rate are better than other optimization methods. However, the large amount of DRR rendering and similarity measurement calculations in the optimization process still cause the registration time to be too long. Compared with the optimization method, the Repair method is dozens of times faster, but the overall process still takes 11.5 seconds because it involves the image completion process. The repaired area cannot perfectly match the original image, which easily affects the calculation of image similarity during the registration training process. The registration accuracy and success rate results cannot meet the registration requirements.

[0087] By t-test, such as Figure 8 As shown in the figure, in terms of mTRE and MAE indicators, the method proposed in this application has a significant improvement over the comparison method (p<0.05). In terms of the registration success rate, it is significantly higher than other comparison methods, reflecting good robustness. In terms of the registration time indicator, the proposed method is superior to the method based on optimization iteration and can meet the clinical demand of completing registration within 3 seconds. Therefore, this model can effectively reduce the occlusion effect of the implant area, thereby ensuring high registration accuracy. At the same time, it greatly speeds up the registration speed to meet the requirements of rapid intraoperative registration.

[0088] Figure 5 The results of visual comparison with other methods after integrating three different metals into three different parts of the spine are randomly displayed. The yellow line is the posture outline of the spine in the X-ray image. It can be seen that Figure 5 The second column shows that the spinal margins of the DRR-registered image using this method are essentially consistent with the reference image. The registration results of the Opt-GC, Opt-NCC, Opt-MI, and Opt-C2F methods (columns 3 through 6) all exhibit significant deviations from the reference image due to occlusion by the implant. The Repair registration result, shown in column 7, exhibits even greater error, essentially indicating a registration failure.

[0089] (2) Model ablation experiments are conducted to analyze the effectiveness of each module in detail, and the results are also tested on mTRE, MAE, and Rot_Error (rotation angle error).

[0090] 1. Encoder only: To avoid the occlusion effect of different metals implanted at different positions, the GCN non-local operation module is introduced. To prove its performance effectiveness, only the encoder is used for experiments.

[0091] 2. Only Encoder (Metal-free): This encoder reconstructs the spinal posture latent space through a posture encoder, thereby regressing accurate posture parameters. It uses X-ray images of only the spine (without metal implants) as input to test its ability to accurately extract posture features from normal spinal images.

[0092] according to Figure 9 The data in Figure 6 As shown in the figure, without GCN processing, the encoder extracts features from the entire X-ray image, including metal occlusions. These high-pixel-valued occlusions significantly interfere with the regression results of the pose parameters, resulting in large registration errors. Furthermore, the encoder's feature extraction capabilities are demonstrated in spinal images without metal implants, keeping the rotation angle error within 1°.

[0093] (3) Considering the different sizes and shapes of pedicle screws used in different hospitals and different surgical situations, the occlusion range of the spine also varies. This summary experiment incorporates implants of different sizes into X-ray images, occluding the spine within a ratio range of approximately (10%, 50%). Similarly, the three indicators of mTRE, MAE, and Rot_Error are used to test the maximum ability of this chapter's model to resist occlusion interference.

[0094] The results are Figure 7 As shown in the figure, the combined results of the three indicators show that when the metal occludes the spine within approximately 35% during the model registration process, the registration result is within an acceptable range, and the rotation angle is still controlled within 1°. When the occlusion range exceeds 35%, the registration accuracy begins to decline. When the occlusion range reaches half, the registration result indicator value can basically be considered a registration failure.

[0095] Figure 10 The table shows the mTRE results of model training under different values of the three regularization parameters λ1, λ2 and λ3. L1 Based on training experience, considering the small distribution range of posture parameter values, λ1 is set to 1. λ2 and λ3 are both set to (0, 0.5) using grid search for parameter adjustment experiments. Figure 10 It can be seen that when λ2 is 0.3 and λ3 is 0.2, L total The training effect is good.

[0096] The present invention also discloses a computer device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the spinal medical image registration method under implant occlusion as described above.

[0097] The present invention also discloses a computer storage medium storing at least one executable instruction, wherein the executable instruction causes a processor to execute the spinal medical image registration method under implant occlusion as described above. The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A spinal medical image registration method under implant occlusion, characterized in that: It includes a registration network framework, which includes the following steps: S100: Input an intraoperative X-ray image with implant occlusion and obtain the Feature Map through CNN. S200: extracting regional occlusion information from a single implant image through spatial pyramid pooling, and adding the regional occlusion information to the FeatureMap obtained in S100 to obtain a feature map; S300: performing hierarchical decoupling processing on the implant region and the spine region in the feature map in step S200 through GCN to obtain hidden features; S400: Use the posture encoder to extract spinal information and reconstruct the corresponding posture latent space. Then use MLP to regress the latent space to obtain the required posture parameters. The preoperative CT is projected with the posture parameters to obtain the final registration image.

2. The spinal medical image registration method under implant occlusion according to claim 1, characterized in that: In step S200, in the SPP network, the screw image is first subjected to a convolution operation to obtain the corresponding Featuremap, and then three global average pooling layers are input, and the pooling pyramid parameters are defined as (1, 4, 8), with a step size of 1. Then, the overall features of the occluded image are concatenated and filled with the feature vectors calculated by the three global average pooling layers, and finally a convolution calculation is performed to form the final output.

3. The spinal medical image registration method under implant occlusion according to claim 1, characterized in that: Step S300 is to use GCN as an occlusion processing module, where each graph node in the feature map represents a single pixel on the feature map, and the adjacency matrix between each pixel serves as the edge relationship between the nodes. Therefore, the graph convolution operation is expressed as: Z=σ(W z AX)+X; A=softmax(f(X i ,X j )); f(X i ,X j )=θ(X i ) T φ(X j ); In the above formula, X is the input image feature, +X represents a residual connection, A represents the adjacency matrix that defines the adjacent relationship of the graph nodes through the dot product, and W Z is the learnable parameter matrix of the output transformation, and finally the output feature Z is obtained through a softmax function and a nonlinear function; Among them, θ and φ are two trainable transformation functions implemented by 1×1 convolution to construct the adjacency matrix and graph nodes.

4. The spinal medical image registration method under implant occlusion according to claim 1, characterized in that: In step S400, a posture encoder is used to extract spinal information and construct an autoencoder architecture. The autoencoder architecture consists of four convolutions. Each layer uses a 3×3 convolution kernel with a step size of 2. Each convolution is followed by a LeakyReLU layer with a parameter of 0.

2. 2D convolution is applied during the encoding process, and spinal X-ray image features under different postures are input to obtain posture information. Then, a 128->64->32->16->3 fully connected layer is connected to regress the posture parameters.

5. The spinal medical image registration method under implant occlusion according to claim 4, characterized in that: When performing registration, specifically: In the above formula, I CT is the CT image, I X is the X-ray image, I metal is the occlusion image, is the posture regression process model of posture encoder + MLP, proj is DRR projection, I reg To register the image.

6. The spinal medical image registration method under implant occlusion according to claim 4, characterized in that: The registration network framework is implemented through multiple loss functions L total The end-to-end training is defined, and the overall objective function is: L total =λ l L L1 +λ2L cosine +λ3L NCC Among them, λ1, λ2 and λ3 are hyperparameters to balance the weights of different loss terms; L L1 is the predicted posture parameter and the real posture r (i) The L1 norm loss is calculated as follows: L cosine is the latent variable Z extracted from the X-ray image i The latent variable Z extracted from the DRR registration map i ''s cosine similarity loss. During the training process, the DRR image is obtained by unobstructed CT projection. The spinal posture features extracted from this image are free of occlusion interference, which is in contrast to the latent variables of the X-ray image with occlusion. The specific calculation is: L NCC is the similarity loss function between the X-ray image and the registered image, and the similarity loss function is expressed as: In the above formula, I X1 with I DRR Corresponding to the intensity values of X-ray image and DRR image in region (i, j), and is the mean of the images in the overlapping region (i, j)∈T.

7. A computer device, characterized in that: include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus, and the memory is used to store at least one executable instruction, wherein the executable instruction enables the processor to perform an operation corresponding to the spinal medical image registration method under implant occlusion as described in any one of claims 1 to 6.

8. A computer storage medium, characterized in that: The computer storage medium stores at least one executable instruction, and the executable instruction enables the processor to execute the spinal medical image registration method under implant occlusion according to any one of claims 1 to 6.

Images