Three-dimensional reconstruction system for preoperative image analysis of spine surgery

Abstract

The invention relates to the technical field of medical image processing, in particular to a three-dimensional reconstruction system for preoperative image analysis of spine surgery. The system comprises a data import and preprocessing module, a three-dimensional reconstruction module, a visual interaction module, a virtual operation planning module, a risk assessment report generation module and a system management module. According to the scheme, a generative adversarial network model is introduced to construct a spine three-dimensional reconstruction network model, and dynamic weight distribution is performed by calculating a domain distance and reconstructing quality similarity, so that a sample which is closer to real data and easier to perform high-quality reconstruction occupies a more important position in training, and therefore, the training process is stabilized, and the effectiveness of alignment is improved; self-similarity is adopted to supervise priori knowledge in loss mining data, cross-region connection is established, the similarity is calculated, characteristics of the most similar position are found for a reconstruction region, it is ensured that only the really meaningful similarity is used for guiding reconstruction, and introduction of error information is avoided.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a three-dimensional reconstruction system for preoperative image analysis in spinal surgery. Background Technology

[0002] In spinal surgery, accurate preoperative assessment of the patient's spinal anatomy, lesion location, and adjacent structures is crucial. Traditional preoperative analysis relies primarily on the physician's observation and experience with two-dimensional CT or MRI slices, which has several limitations: two-dimensional images cannot intuitively and comprehensively present the complex three-dimensional anatomy of the spine, and existing three-dimensional reconstruction software is mostly a general-purpose medical image processing tool, lacking optimization for spinal anatomy. When the input images have quality differences such as imaging noise, blurring, or artifacts, the reconstructed three-dimensional model is prone to distortion, resulting in a discrepancy between the spinal structure and the patient's actual anatomy. Furthermore, for areas lacking high-quality reference data, existing methods lack effective supervision mechanisms, leading to uneven reconstruction quality and affecting the overall reliability of the model. Therefore, developing a three-dimensional reconstruction system specifically designed for spinal surgery, which is easy to operate, provides accurate reconstruction, and supports in-depth preoperative analysis has significant clinical value. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a three-dimensional reconstruction system for preoperative image analysis in spinal surgery. Addressing the problem that existing three-dimensional reconstruction software is mostly general-purpose medical image processing tools lacking optimization for spinal anatomical features, and prone to distortion in reconstructed three-dimensional models when input images exhibit quality differences such as imaging noise, blurring, or artifacts, this solution introduces a generative adversarial network (GAN) model to construct a spinal three-dimensional reconstruction network model. Through adversarial learning, the reconstruction network matches the feature distribution of low-quality images with that of high-quality reference images in the feature space, thereby improving reconstruction quality. By calculating domain distance and reconstruction quality similarity, weights are dynamically allocated, giving greater importance to samples that are closer to real data and easier to reconstruct with high quality during training, thus stabilizing the training process and improving alignment effectiveness. For regions lacking high-quality reference data, existing methods lack effective supervision mechanisms, leading to uneven reconstruction quality. This solution employs self-similarity supervised loss to mine prior knowledge within the data, establishes cross-regional connections, calculates similarity to find the most similar features in the reconstructed region, and filters out irrelevant matches, ensuring that only truly meaningful similarities are used to guide reconstruction and avoiding the introduction of erroneous information.

[0004] The present invention provides a three-dimensional reconstruction system for preoperative image analysis in spinal surgery. The system includes a data import and preprocessing module, a three-dimensional reconstruction module, a visualization and interaction module, a virtual surgery planning module, a risk assessment report generation module, and a system management module.

[0005] The data import and preprocessing module is used to receive the patient's DICOM format CT and MRI raw image data, and to perform data format conversion, image registration, noise filtering, and image enhancement operations on the raw image data to obtain preprocessed images.

[0006] The 3D reconstruction module uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine, generates a 3D model of the spine using preprocessed images, and supports visualization by assigning different colors and transparency to different structures;

[0007] The visualization and interaction module provides a three-dimensional view window, which supports users to perform virtual interaction and in-depth observation of the generated three-dimensional model of the spine.

[0008] The virtual surgical planning module allows doctors to select the target vertebra and screw placement location on the generated 3D model of the spine. It automatically recommends safe channels using the D* algorithm, allowing doctors to adjust the screw diameter, length, insertion angle, and depth. It also displays the relative position of the screw to surrounding nerves and blood vessels in real time, performs virtual screw placement, and assesses the safety of screw placement.

[0009] After the virtual pin placement and implant placement, the risk assessment report generation module automatically performs collision detection using a collision detection algorithm, highlights risk areas that may damage nerve roots, spinal cord, or blood vessels, and automatically summarizes the patient's basic information, imaging diagnosis, three-dimensional reconstruction results, virtual surgical planning scheme, and risk assessment results to generate a standardized and visualized preoperative analysis report.

[0010] The system management module includes user access control, patient data management, and data backup and recovery functions. It provides standard interfaces for data interaction with hospital PACS (Picture Archiving and Communication System) and HIS (Hospital Information System) to achieve workflow integration.

[0011] Furthermore, the 3D reconstruction module uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine, specifically including the following steps:

[0012] Step S1: Data input, collect a set of high-quality, high-resolution spinal CT and MRI body data as reference CT-MR data;

[0013] Step S2: Multiple randomization. A multiple randomization model is used to simulate the image quality degradation process commonly seen in clinical practice for each reference CT-MR data, generating corresponding synthetic low-quality volume data, and obtaining a synthetic low-quality spinal image set.

[0014] Step S3: Train the basic 3D reconstruction network. The synthesized low-quality spinal image set and the reference CT-MR data are divided into data pairs and input into the 3D residual dense network model for training. After training, a pre-trained basic 3D reconstruction model is obtained.

[0015] Step S4: Dynamic domain adaptive optimization. An adaptive reconstruction framework is constructed by combining the pre-trained basic 3D reconstruction model and the adversarial loss is optimized using dynamic weights to obtain the spinal 3D reconstruction network model.

[0016] Step S5: 3D reconstruction and preoperative analysis. The patient's complete large-scale low-resolution spinal images are input into the 3D reconstruction network model of the spine to obtain a high-resolution, detailed 3D model of the spine.

[0017] Furthermore, in step S4, the dynamic domain adaptive optimization specifically includes the following steps:

[0018] Step S41: Data segmentation, acquire low-resolution spinal imaging data of the patient's whole body and large area as low-resolution volume data, and acquire high-resolution imaging data of key areas of the same patient's spine as high-resolution volume data.

[0019] Step S42: Data matching. Use a rigid registration algorithm to align the high-resolution volume data to the physical location corresponding to the low-resolution volume data. After registration, use the high-resolution volume data as the reference data for the sub-region of the low-resolution volume data. Use the registered sub-region as the overlapping region and the unregistered sub-region as the non-overlapping region.

[0020] Step S43: Initialize the domain adaptive reconstruction framework. Based on the pre-trained basic 3D reconstruction model, construct a generative adversarial network architecture containing two 3D reconstruction networks and a 3D discriminator. The 3D reconstruction networks are respectively... and 3D reconstruction network 3D reconstruction network The weights are copied to the 3D reconstruction network. Use 3D reconstruction network as initial weights and Generate reconstructed volume data and use a 3D discriminator to distinguish the reconstructed volume data from the reference data;

[0021] Step S44: Generate dynamic weights, calculate the domain distance between the reconstructed volume data and the overlapping region using the maximum mean difference, calculate the similarity between the reconstructed volume data and the high-resolution volume data using SSIM, obtain the reconstruction quality score, and generate dynamic weights by combining the domain distance and the reconstruction quality score.

[0022] Step S45: Optimize adversarial loss and update the 3D reconstruction network using dynamic weights. Optimize the parameters of the 3D discriminator and the adversarial loss of the 3D reconstruction network and the 3D discriminator;

[0023] Step S46: Optimize the self-similar anatomical structure loss, use the optimized overlapping regions to guide the reconstruction of non-overlapping regions, and refine the 3D reconstruction network. The parameters are updated to obtain a stable 3D reconstruction network model of the spine, which includes the following steps:

[0024] Step S461: Establish cross-regional feature associations, and input overlapping and non-overlapping regions into the 3D reconstruction network. In the middle, extract the feature map before the upsampling layer. and For feature maps For each pixel in the image, use cosine similarity in... Find the most similar pixel in the array;

[0025] Step S462: Filtering and enhancing connections. Set a similarity threshold and use it to filter out irrelevant matches with low similarity, thus improving the 3D reconstruction network. When reconstructing non-overlapping regions, the most similar pixels known in overlapping regions are referenced to reconstruct a high-resolution 3D model of the spine.

[0026] Step S463: Calculate the total loss. The total loss is calculated using dynamic adversarial loss and self-similarity supervised loss, and the 3D reconstruction network is updated using the exponential moving average algorithm. The parameters are used to obtain a stable three-dimensional reconstruction network model of the spine, and the formula used is as follows: ; ;

[0027] In the formula, , and Representing feature maps respectively and Scales in the first, second, and third spatial dimensions, For sparsification functions, The preset similarity threshold, and For pixel index, Representation of feature map Medium pixel eigenvectors, Representation of feature map Most similar pixels eigenvectors, For similarity measurement, Indicates the current similarity. To dynamically mitigate losses, For self-similarity supervision loss, and These are the weight coefficients for the dynamic adversarial loss and the self-similarity supervision loss, respectively. This represents the total loss.

[0028] The beneficial effects achieved by the present invention using the above solution are as follows:

[0029] (1) As existing 3D reconstruction software is mostly a general medical image processing tool, it lacks optimization for spinal anatomical features. When the input image has quality differences such as imaging noise, blur or artifacts, the reconstructed 3D model is easy to be distorted, resulting in the spinal structure not matching the actual anatomy of the patient. This solution introduces a generative adversarial network model to construct a 3D reconstruction network model for the spine. By constructing two reconstruction networks for adversarial learning, the reconstruction network generates results for low-quality images, which match the feature distribution of reference high-quality images in the feature space, thereby improving the reconstruction quality. By calculating the domain distance and reconstruction quality similarity, and shrinking by a preset threshold, the weights are dynamically allocated, so that samples that are closer to the real data and easier to reconstruct with high quality occupy a more important position in the training, while the influence of those outliers and difficult to reconstruct is reduced, thereby stabilizing the training process and improving the effectiveness of alignment.

[0030] (2) For regions lacking high-quality reference data, existing methods lack effective supervision mechanisms, resulting in uneven reconstruction quality. Self-similar supervised loss is used to mine prior knowledge within the data, establish cross-regional connections, calculate similarity, find the most similar features for the reconstruction region, and filter out irrelevant matches to ensure that only truly meaningful similarities are used to guide reconstruction, avoid introducing erroneous information, and maintain good anatomical or structural continuity in the entire image or the entire spine, thus avoiding the problem that the reconstruction quality of local regions is significantly lower than that of other regions. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the process of a three-dimensional reconstruction system for preoperative image analysis in spinal surgery proposed in this invention. Figure 2 A schematic diagram of the data processing flow for constructing a three-dimensional reconstruction network model of the spine.

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example 1, see Figure 1 The present invention provides a three-dimensional reconstruction system for preoperative image analysis in spinal surgery. The system includes a data import and preprocessing module, a three-dimensional reconstruction module, a visualization and interaction module, a virtual surgery planning module, a risk assessment report generation module, and a system management module.

[0035] The data import and preprocessing module is used to receive the patient's DICOM format CT and MRI raw image data, and to perform data format conversion, image registration, noise filtering, and image enhancement operations on the raw image data to obtain preprocessed images.

[0036] The 3D reconstruction module uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine, generates a 3D model of the spine using preprocessed images, and supports visualization by assigning different colors and transparency to different structures;

[0037] The visualization and interaction module provides a three-dimensional view window, which supports users to perform virtual interaction and in-depth observation of the generated three-dimensional model of the spine.

[0038] The virtual surgical planning module allows doctors to select the target vertebra and screw placement location on the generated 3D model of the spine. It automatically recommends safe channels using the D* algorithm, allowing doctors to adjust the screw diameter, length, insertion angle, and depth. It also displays the relative position of the screw to surrounding nerves and blood vessels in real time, performs virtual screw placement, and assesses the safety of screw placement.

[0039] After the virtual pin placement and implant placement, the risk assessment report generation module automatically performs collision detection using a collision detection algorithm, highlights risk areas that may damage nerve roots, spinal cord, or blood vessels, and automatically summarizes the patient's basic information, imaging diagnosis, three-dimensional reconstruction results, virtual surgical planning scheme, and risk assessment results to generate a standardized and visualized preoperative analysis report.

[0040] The system management module includes user permission management, patient data management, data backup and recovery functions, manages the functions and permissions of other modules, provides standard interfaces for data interaction with hospital PACS and HIS, and realizes workflow integration.

[0041] Example 2, based on the above examples, describes a 3D reconstruction module that uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine. The specific steps include:

[0042] Step S1: Data input, collect a set of high-quality, high-resolution spinal CT and MRI body data as reference CT-MR data;

[0043] Step S2: Multiple randomization. A multiple randomization model is used to simulate the image quality degradation process commonly seen in clinical practice for each reference CT-MR data, generating corresponding synthetic low-quality volume data, and obtaining a synthetic low-quality spinal image set.

[0044] Step S3: Train the basic 3D reconstruction network. The synthesized low-quality spinal image set and the reference CT-MR data are divided into data pairs and input into the 3D residual dense network model for training. After training, a pre-trained basic 3D reconstruction model is obtained.

[0045] Step S4: Dynamic domain adaptive optimization. An adaptive reconstruction framework is constructed by combining the pre-trained basic 3D reconstruction model and the adversarial loss is optimized using dynamic weights to obtain the spinal 3D reconstruction network model.

[0046] Step S5: 3D reconstruction and preoperative analysis. The patient's complete large-scale low-resolution spinal images are input into the 3D reconstruction network model of the spine to obtain a high-resolution, detailed 3D model of the spine.

[0047] Example 3, based on the above examples, in step S2, multiple randomization is used to generate corresponding synthetic low-quality image data by simulating a complex real degradation process, including the following:

[0048] Image blurring: Apply isotropic Gaussian kernel and anisotropic Gaussian kernel to the image used to reconstruct 3D volume data in sequence;

[0049] Image downsampling: Downsampling is performed using nearest neighbor, bilinear, or bicubic interpolation. Before nearest neighbor downsampling, two-dimensional linear grid interpolation is applied to reduce pixel offset. Downsampling-upsampling strategy is used to simulate some volume effects and interpolation artifacts, which reduces image quality.

[0050] Add noise: simulate electronic noise or low signal-to-noise ratio during the scanning process, add Gaussian noise, sample the standard deviation according to the empirical value of different scanning protocols, introduce speckle noise to simulate ultrasound images, and Rician noise to simulate MRI.

[0051] Example 4, based on the above examples, includes the following steps in step S4: Dynamic domain adaptive optimization.

[0052] Step S41: Data segmentation, acquire low-resolution spinal imaging data of the patient's whole body and large area as low-resolution volume data, and acquire high-resolution imaging data of key areas of the same patient's spine as high-resolution volume data.

[0053] Step S42: Data matching. Use a rigid registration algorithm to align the high-resolution volume data to the physical location corresponding to the low-resolution volume data. After registration, use the high-resolution volume data as the reference data for the sub-region of the low-resolution volume data. Use the registered sub-region as the overlapping region and the unregistered sub-region as the non-overlapping region.

[0054] Step S43: Initialize the domain adaptive reconstruction framework. Based on the pre-trained basic 3D reconstruction model, construct a generative adversarial network architecture containing two 3D reconstruction networks and a 3D discriminator. The 3D reconstruction networks are respectively... and 3D reconstruction network 3D reconstruction network The weights are copied to the 3D reconstruction network. Use 3D reconstruction network as initial weights and Generate reconstructed volume data and use a 3D discriminator to distinguish the reconstructed volume data from the reference data;

[0055] Step S44: Generate dynamic weights. Calculate the domain distance between the reconstructed volume data and the overlapping region using the maximum mean difference. Calculate the similarity between the reconstructed volume data and the high-resolution volume data using SSIM to obtain the reconstruction quality score. Combine the domain distance and the reconstruction quality score to generate dynamic weights. The formula used is as follows: ; ;

[0056] In the formula, and They represent the first The data of the first reconstructed body and the first High-resolution volume data, For similarity functions, The preset threshold, Represents the maximum value function. It is a constant, and its value is , For similarity weight, For the first Domain distance between reconstructed volume data and overlapping regions This indicates the number of reconstructed body data in the same batch. For the generated dynamic weights;

[0057] Step S45: Optimize adversarial loss and update the 3D reconstruction network using dynamic weights. Optimize the parameters of the 3D discriminator and the adversarial loss of the 3D reconstruction network and the 3D discriminator;

[0058] Step S46: Optimize the self-similar anatomical structure loss, use the optimized overlapping regions to guide the reconstruction of non-overlapping regions, and refine the 3D reconstruction network. The parameters are updated to obtain a stable 3D reconstruction network model of the spine.

[0059] By performing the aforementioned operations, this solution addresses the problem that existing 3D reconstruction software, mostly general-purpose medical image processing tools, lacks optimization for spinal anatomical features. When the input image has quality differences such as imaging noise, blurring, or artifacts, the reconstructed 3D model is prone to distortion, leading to a discrepancy between the spinal structure and the patient's actual anatomy. This solution introduces a generative adversarial network (GAN) model to construct a 3D reconstruction network model for the spine. By constructing two reconstruction networks and conducting adversarial learning, the reconstruction network's results generated from low-quality images are matched with the feature distribution of a reference high-quality image in the feature space, thereby improving reconstruction quality. By calculating domain distance and reconstruction quality similarity, and applying a preset threshold for shrinkage, weights are dynamically allocated. This ensures that samples that are closer to the real data and easier to reconstruct with high quality occupy a more important position in training, while the influence of outliers and difficult-to-reconstruct samples is reduced, thus stabilizing the training process and improving the effectiveness of alignment.

[0060] Example 5, based on the above examples, involves constructing a 3D reconstruction network in step S43, comprising two 3D reconstruction networks. As the base model, it utilizes the supervised data synthesized in step S3 for fully supervised training, learning the mapping relationship from simulated, low-quality inputs to high-quality outputs; 3D reconstruction network It is a 3D reconstruction network The learned knowledge is transferred to real, unlabeled clinical data and optimized through domain adaptation and self-supervised learning. Adversarial learning is used to align the reconstruction results of real low-quality data with high-quality reference data in terms of feature distribution, thereby bridging the gap between the synthetic domain and the real domain.

[0061] Example 6, based on the above examples, optimizes the self-similar anatomical structure loss in step S46, specifically including the following steps:

[0062] Step S461: Establish cross-regional feature associations, and input overlapping and non-overlapping regions into the 3D reconstruction network. In the middle, extract the feature map before the upsampling layer. and For feature maps For each pixel in the image, use cosine similarity in... Find the most similar pixel in the array;

[0063] Step S462: Filtering and enhancing connections. Set a similarity threshold and use it to filter out irrelevant matches with low similarity, thus improving the 3D reconstruction network. When reconstructing non-overlapping regions, the most similar pixels known in overlapping regions are referenced to reconstruct a high-resolution 3D model of the spine.

[0064] Step S463: Calculate the total loss. The total loss is calculated using dynamic adversarial loss and self-similarity supervised loss, and the 3D reconstruction network is updated using the exponential moving average algorithm. The parameters are used to obtain a stable three-dimensional reconstruction network model of the spine, and the formula used is as follows: ; ;

[0065] In the formula, , and Representing feature maps respectively and Scales in the first, second, and third spatial dimensions, For sparsification functions, The preset similarity threshold, and For pixel index, Representation of feature map Medium pixel eigenvectors, Representation of feature map Most similar pixels eigenvectors, For similarity measurement, Indicates the current similarity. To dynamically mitigate losses, For self-similarity supervision loss, and These are the weight coefficients for the dynamic adversarial loss and the self-similarity supervision loss, respectively. This represents the total loss.

[0066] By performing the aforementioned operations, existing methods lack effective supervision mechanisms for regions lacking high-quality reference data, resulting in uneven reconstruction quality. This approach employs self-similarity supervised loss to mine prior knowledge within the data, establishes cross-regional connections, calculates similarity to find the most similar features for the reconstructed region, and filters out irrelevant matches. This ensures that only truly meaningful similarities are used to guide reconstruction, avoiding the introduction of erroneous information. This allows the reconstruction results to maintain good anatomical or structural continuity across the entire image or the entire spine, preventing the problem of significantly lower reconstruction quality in local areas compared to other areas.

[0067] Example 7, based on the above examples, describes a three-dimensional reconstruction module that uses a regional refined reconstruction strategy when generating a three-dimensional model of the spine. The anatomical structure of the spine is divided into multiple independent parts according to functional and morphological characteristics and modeled separately. Specifically, these include major segments such as the cervical, thoracic, lumbar, sacral, and coccygeal vertebrae, which can be further subdivided into key anatomical units such as vertebral bodies, pedicles, laminae, spinous processes, transverse processes, intervertebral joints, and intervertebral discs. Each part of the model retains accurate spatial coordinates and topological relationships, supporting seamless combination and assembly under a unified coordinate system, thereby constructing a complete, continuous, and anatomically clear three-dimensional model of the entire spine.

[0068] Example 8, based on the above examples, provides a 3D view window with free rotation, scaling, and translation operations; multi-plane reconstructed views and 3D views are displayed in conjunction; layered display / hiding and sectioning functions of the model are supported; internal structures can be observed; the 3D model can be overlaid or compared with the original image data; and the reconstructed vertebral bodies, pedicles, lamina, spinous processes, transverse processes, intervertebral discs, spinous cords, nerve roots, and major vascular key anatomical structures are labeled.

[0069] The virtual surgical planning module supports simulating different types of osteotomy surgery on a 3D model, rehearsing the osteotomy range and angle, and evaluating the correction effect; it has a built-in 3D model library of commonly used spinal internal fixation devices (such as pedicle screws, interbody fusion devices, and artificial vertebrae), which can be directly dragged and dropped to the planned screw placement position for matching and adjustment.

[0070] The basic patient information described in the risk assessment report generation module includes patient identity information, medical record information, examination-related information, and clinical information. Patient information includes name, gender, date of birth, age, and unique patient identification number. Medical record information includes medical record number, consultation number, and hospitalization number. Examination-related information includes the date of CT or MRI scan, examination number, and examination type. Clinical information includes clinical diagnosis results, department, and attending physician.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0073] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A three-dimensional reconstruction system for preoperative image analysis in spinal surgery, characterized in that: The system includes a data import and preprocessing module, a 3D reconstruction module, a visualization and interaction module, a virtual surgery planning module, a risk assessment report generation module, and a system management module; The data import and preprocessing module is used to receive the patient's DICOM format CT and MRI raw image data, and to perform data format conversion, image registration, noise filtering, and image enhancement operations on the raw image data to obtain preprocessed images. The 3D reconstruction module uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine, generates a 3D model of the spine using preprocessed images, and supports visualization by assigning different colors and transparency to different structures; The visualization and interaction module provides a three-dimensional view window, which supports users to perform virtual interaction and in-depth observation of the generated three-dimensional model of the spine. The virtual surgery planning module selects the target vertebra and screw placement position on the generated three-dimensional model of the spine, automatically recommends a safe passage using the D* algorithm, allows doctors to adjust the diameter, length, insertion angle and depth of the screw, displays the relative position of the screw to the surrounding nerves and blood vessels in real time, performs virtual screw placement, and assesses the safety of screw placement. After the virtual nail placement and implant placement, the risk assessment report generation module automatically performs collision detection using a collision detection algorithm, automatically summarizes the patient's basic information, imaging diagnosis, three-dimensional reconstruction results, virtual surgical planning scheme and risk assessment results, and generates a standardized and visualized preoperative analysis report. The system management module includes user permission management, patient data management, data backup and recovery functions, and provides standard interfaces for data interaction with hospital PACS and HIS to achieve workflow integration.

2. The three-dimensional reconstruction system for preoperative image analysis in spinal surgery according to claim 1, characterized in that: The 3D reconstruction module uses generative adversarial networks and self-similar anatomical structure supervised loss to construct a 3D reconstruction network model of the spine, specifically including the following steps: Step S1: Data input, collect a set of high-quality, high-resolution spinal CT and MRI body data as reference CT-MR data; Step S2: Multiple randomization. A multiple randomization model is used to simulate the image quality degradation process commonly seen in clinical practice for each reference CT-MR data, generating corresponding synthetic low-quality volume data, and obtaining a synthetic low-quality spinal image set. Step S3: Train the basic 3D reconstruction network. The synthesized low-quality spinal image set and the reference CT-MR data are divided into data pairs and input into the 3D residual dense network model for training. After training, a pre-trained basic 3D reconstruction model is obtained. Step S4: Dynamic domain adaptive optimization. An adaptive reconstruction framework is constructed by combining the pre-trained basic 3D reconstruction model and the adversarial loss is optimized using dynamic weights to obtain the spinal 3D reconstruction network model. Step S5: 3D reconstruction and preoperative analysis. The patient's complete large-scale low-resolution spinal images are input into the 3D reconstruction network model of the spine to obtain a high-resolution, detailed 3D model of the spine.

3. A three-dimensional reconstruction system for preoperative image analysis in spinal surgery according to claim 2, characterized in that: In step S4, the dynamic domain adaptive optimization specifically includes the following steps: Step S41: Data segmentation, acquire low-resolution spinal imaging data of the patient's whole body and large area as low-resolution volume data, and acquire high-resolution imaging data of key areas of the same patient's spine as high-resolution volume data. Step S42: Data matching. Use a rigid registration algorithm to align the high-resolution volume data to the physical location corresponding to the low-resolution volume data. After registration, use the high-resolution volume data as the reference data for the sub-region of the low-resolution volume data. Use the registered sub-region as the overlapping region and the unregistered sub-region as the non-overlapping region. Step S43: Initialize the domain adaptive reconstruction framework. Based on the pre-trained basic 3D reconstruction model, construct a generative adversarial network architecture containing two 3D reconstruction networks and a 3D discriminator. The 3D reconstruction networks are respectively... and 3D reconstruction network 3D reconstruction network The weights are copied to the 3D reconstruction network. Use 3D reconstruction network as initial weights and Generate reconstructed volume data and use a 3D discriminator to distinguish the reconstructed volume data from the reference data; Step S44: Generate dynamic weights, calculate the domain distance between the reconstructed volume data and the overlapping region using the maximum mean difference, calculate the similarity between the reconstructed volume data and the high-resolution volume data using SSIM, obtain the reconstruction quality score, and generate dynamic weights by combining the domain distance and the reconstruction quality score. Step S45: Optimize adversarial loss and update the 3D reconstruction network using dynamic weights. Optimize the parameters of the 3D discriminator and the adversarial loss of the 3D reconstruction network and the 3D discriminator; Step S46: Optimize the self-similar anatomical structure loss, use the optimized overlapping regions to guide the reconstruction of non-overlapping regions, and refine the 3D reconstruction network. The parameters are updated to obtain a stable 3D reconstruction network model of the spine.

Images