Surgical navigation system based on multimode image

Abstract

The invention relates to a surgical navigation system based on a multimode image. The surgical navigation system comprises a multimode data acquisition module, an image preprocessing module, an image registration module, a path planning module, an intraoperative tracking module and an error correction module. The system constructs a structure bias registration model, fuses a reversible neural network and a dynamic depth divisible convolution mechanism, and improves the structure retention capability and registration robustness of a multi-modal image. A mutual information loss function and a CIDER bias regularization term of structure gradient regulation are introduced, and the suppression capability on pseudo factor interference is enhanced; in the aspect of path planning, a rare path perception SAC model is proposed, and track clustering and a three-dimensional risk map are combined to realize rare path avoidance and strategy optimization; the system has good registration precision, path robustness and tracking consistency, and is suitable for a complex navigation scene assisted by a multi-modal image.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a surgical navigation system based on multi-modal images. Background Art

[0002] With the development of medical imaging technology, the auxiliary role of multi-modal medical images such as CT, MRI, and ultrasound in the field of surgical navigation has been increasing day by day; multi-modal image fusion can make up for the incompleteness of single-modal information and provide accurate anatomical structure support for preoperative path planning and intraoperative real-time navigation; however, there are differences in resolution, gray value, imaging mechanism, etc. between different modal images, resulting in problems such as structural distortion, interference of pseudo factors, and insufficient fusion consistency in the registration process, making it difficult to meet the requirements of high-precision surgical navigation; currently, traditional image registration methods mainly adopt optimization algorithms based on rigid or non-rigid transformation, but such methods often lack the ability to deeply model structural features and are difficult to handle complex non-linear differences between modalities; in addition, although existing neural network registration models introduce a deep feature learning mechanism, their ability to suppress structural preservation and modal bias is still limited, especially in the presence of strong artifacts or structurally blurred areas, registration error accumulation may still occur, affecting the robustness and stability of the navigation system.

[0003] On the other hand, preoperative path planning is a key link in the surgical navigation system; although traditional path planning algorithms such as A* and RRT have a certain degree of real-time performance, in complex surgical anatomical scenarios, they lack the ability to deeply model historical path patterns and abnormal behavior patterns and are difficult to effectively identify and avoid rare risk paths; at the same time, existing path planning methods based on reinforcement learning mostly adopt a general strategy training mode, without considering rare behavior patterns and environmental risk factors in the intraoperative real navigation trajectory, resulting in poor interpretability and weak robustness of the output path strategy. Summary of the Invention

[0004] The present invention aims to overcome the problems in the prior art such as insufficient structural preservation ability in multimodal image registration, significant interference from pseudo factors, and poor adaptability of the path planning model. A surgical navigation system based on multimodal images is proposed to achieve the system integration optimization of structural bias registration, rare path perception path planning, and intraoperative tracking error correction. Its innovation lies in: constructing a structural bias registration model, combining a reversible neural network and a dynamic depthwise separable convolution mechanism to enhance the local structural preservation ability of cross-modal images; introducing a barrier mutual information loss function regulated by structural gradient and a kernel-estimated CIDER bias regular term to improve the robustness of the registration process against pseudo factor interference; in terms of path planning, a rare path perception navigation SAC model is proposed. By clustering and identifying historical navigation trajectories, extracting rare behavior patterns, constructing a three-dimensional risk map and embedding it into the strategy training process, risk avoidance and robust optimization of path planning are achieved. The present invention realizes the comprehensive performance improvement of the navigation system in terms of image processing accuracy, path robustness, and tracking stability by constructing a structural preservation and bias suppression mechanism in image registration, introducing rare behavior modeling and risk regulation strategies in path planning, and combining intraoperative real-time feedback, and has good technical integration and engineering adaptability.

[0005] The present invention provides a surgical navigation system based on multimodal images, which includes a multimodal data acquisition module, an image preprocessing module, an image registration module, a preoperative path planning module, an intraoperative tracking module, and a navigation feedback and error correction module;

[0006] The multimodal data acquisition module acquires CT images, MRI images, and ultrasound images, unifies their coordinate systems, and performs time synchronization using electrocardiogram gating to generate original multimodal image data;

[0007] The image preprocessing module normalizes the pixel intensity, suppresses noise, and enhances the ultrasound image of the original multimodal image data to obtain standardized image voxel data; non-local means and 3D Gaussian filtering methods are used for noise suppression, and Speckle denoising + contrast enhancement techniques are used for ultrasound image enhancement;

[0008] The image registration module constructs a structural bias registration model and uses the structural bias registration model to register the standardized image voxel data to generate a multimodal fusion registration image; the structural bias registration model includes a preprocessing unit, a modality translation unit, a bias optimization unit, a registration unit, a training unit, and a reversible neural network;

[0009] The preoperative path planning module establishes an SAC model, optimizes the strategy training of the SAC model, constructs a rare path perception navigation SAC model, and processes the multimodal fusion registration image through the rare path perception navigation SAC model to generate a preoperative navigation path. The rare path perception navigation SAC model includes the SAC model;

[0010] An intraoperative tracking module uses visual SLAM technology according to the preoperative navigation path to perform intraoperative registration during the operation, real-time match the preoperative path with the intraoperative image, provide an augmented reality view, perform real-time navigation display, and obtain real-time position coordinates and intraoperative images.

[0011] A navigation feedback and error correction module corrects the navigation trajectory through Kalman filtering based on the real-time position coordinates and intraoperative images.

[0012] Furthermore, the process of registering the standardized image voxel data using a structure-biased registration model to generate a multi-modal fusion registration image specifically includes the following steps:

[0013] Step S1: Perform intensity normalization, unify the gray scale range, unify the resolution, and remove the background from the standardized image voxel data to obtain a pair of standardized images.

[0014] Step S2: Respectively input the pair of standardized images as the source image and the real target image into a reversible neural network. Perform modal translation on the source image through an affine coupling structure to generate a pseudo-target image. Use the reversible structure to perform inverse mapping on the pseudo-target image to restore the structural information of the source image, ensuring that the structural information is not lost during the modal translation process. Embed a dynamic depthwise separable convolution mechanism in the affine coupling structure as a local attention path, perform sparse connection and dynamic weighting through the driven kernel combination to enhance the local structural detail extraction ability, and generate an enhanced pseudo-image. Introduce a barrier mutual information loss function regulated by structural gradients to measure the mutual information between the enhanced pseudo-image and the real target image, limit the degree of structural distortion between images, and obtain a structure-preserving image.

[0015] Step S3: Define a set of pseudo-factors. Input the structure-preserving image into the backbone network to generate a prediction result. Perform conditional distance correlation measurement on the prediction result and the set of pseudo-factors, and introduce a kernel-estimated CIDER regularization term to suppress pseudo-factor interference, realizing bias constraint optimization under the guidance of the structure, and obtaining a bias-constrained prediction image.

[0016] Step S4: Input the bias-constrained prediction image and the real target image into a registration network to learn a non-rigid deformation field, and use a spatial transformer to perform spatial torsion on the bias-constrained prediction image to generate an initial registration image.

[0017] Step S5: Combine steps S2 - S4 to construct a joint optimization objective function, train the structure-biased registration model, and further optimize the geometric consistency and modal fusion consistency of the initial registration image to generate a multi-modal fusion registration image.

[0018] Furthermore, the process of generating a preoperative navigation path through a rare path perception navigation SAC model specifically includes the following steps:

[0019] Step B1: Extract surgical anatomical structure information from the multimodal fusion registration image to construct a three-dimensional sparse voxel map. The three-dimensional sparse voxel map includes a point set and an edge set, where the point set represents path sampling points and the edge set represents feasible navigation connectivity, forming a structured three-dimensional navigation map;

[0020] Step B2: Based on the structured three-dimensional navigation map, collect the state-action pair sequence of its historical preoperative navigation trajectories. Semantically encode the state-action pair sequence through the Encoder network to obtain a feature embedding set, and then use the K-means algorithm to cluster the feature embedding set to generate a discretized label sequence, completing the structured abstraction of the trajectory behavior;

[0021] Step B3: Remove duplicates from the discretized label sequence to obtain a structural decision-making pattern, count the occurrence frequency of each pattern, and sort by frequency to obtain a rare pattern set; Based on the rare pattern set, construct the coverage ratio of rare paths;

[0022] Step B4: Based on the rare pattern set and the coverage ratio of rare paths, construct a three-dimensional space risk map, use the three-dimensional space risk map as external constraint information and embed it into the training environment of the SAC model to avoid the coverage ratio of high-rare paths and generate trajectory experience pairs; The three-dimensional space risk map includes risk scores, the coverage ratio of rare paths, and heat map mapping;

[0023] Step B5: Input the trajectory experience pairs into the SAC model for policy training, output the action distribution in the current state through the policy network, estimate the Q function through the double Q-network structure, and monitor the upper bound of the Q function valuation error through the Q-value estimation error formula, output the path planning policy, and obtain the preoperative navigation path.

[0024] Adopting the above solution, the beneficial effects of the present invention are as follows:

[0025] The present invention realizes the structure preservation and modality alignment in the registration process of multimodal images by constructing a structure-biased registration model, effectively improving the geometric consistency and local structure expression ability of the registration image; Combining the reversible neural network and the dynamic depthwise separable convolution mechanism enables the registration model to have the dynamic modeling ability for local edge features and nonlinear modality differences; Further introducing the barrier mutual information loss function and the CIDER bias regular term regulated by structural gradients effectively suppresses the interference of pseudo factors on the registration accuracy, solves the problem of unstable registration in the structure-blurred area of existing methods, and enhances the registration robustness and generalization ability of the system in the multimodal image fusion scenario.

[0026] In terms of path planning, the present invention proposes a rare path perception navigation SAC model, which for the first time combines the clustering modeling mechanism of historical trajectory behaviors with the three-dimensional space risk map construction method to identify and constrain the rare behavior paths existing in navigation, achieving effective avoidance and robust optimization of abnormal trajectories in the path strategy. By introducing the rare path coverage ratio and behavior pattern distribution information as training constraints, this model improves the stability and rationality of the policy output, solves the problems of unreasonable path selection and poor abnormal trajectory avoidance ability in traditional methods, and significantly improves the path generation quality and policy credibility of the system in complex three-dimensional structure diagrams.

[0027] In addition, the present invention integrates visual SLAM tracking and Kalman filter error correction mechanisms to achieve dynamic matching and position update between the intraoperative path and real-time images, effectively enhancing the spatio-temporal consistency and system stability during navigation. By introducing state estimation and error correction capabilities in continuous intraoperative feedback, the system can achieve real-time adjustment of problems such as path drift and positioning error, solve the common problem of decreasing navigation accuracy over time, improve the anti-interference ability and execution accuracy of the entire system during operation, and provide a solid guarantee for the continuity and reliability of the surgical navigation path. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the modules of a surgical navigation system based on multimodal images provided by the present invention;

[0029] Figure 2 It is a navigation feedback and error correction diagram provided in Embodiment 6.

[0030] Figure 2 In it, the blue line: the preoperative navigation path, as a reference; the red dashed line: the actual intraoperative trajectory, measured; the green dotted line: the Kalman filter corrected trajectory. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] Embodiment 1, according to Figure 1 , the present invention provides a surgical navigation system based on multimodal images, which includes a multimodal data acquisition module, an image preprocessing module, an image registration module, a preoperative path planning module, an intraoperative tracking module, and a navigation feedback and error correction module;

[0033] The multimodal data acquisition module acquires CT images, MRI images, and ultrasound images, unifies their coordinate systems, and performs time synchronization using electrocardiogram gating to generate original multimodal image data;

[0034] The image preprocessing module normalizes the pixel intensity, suppresses noise, and enhances the ultrasound image of the original multimodal image data to obtain standardized image voxel data; Non-local means and 3D Gaussian filtering methods are used for noise suppression, and Speckle denoising + contrast enhancement techniques are used for ultrasound image enhancement;

[0035] The image registration module constructs a structure bias registration model and uses the structure bias registration model to register the standardized image voxel data to generate a multimodal fusion registration image; The structure bias registration model includes a preprocessing unit, a modality translation unit, a bias optimization unit, a registration unit, a training unit, and a reversible neural network;

[0036] The preoperative path planning module establishes an SAC model, optimizes the policy training of the SAC model, constructs a rare path perception navigation SAC model, processes the multimodal fusion registration image through the rare path perception navigation SAC model to generate a preoperative navigation path, and the rare path perception navigation SAC model includes the SAC model;

[0037] The intraoperative tracking module uses visual SLAM technology to perform intraoperative registration based on the preoperative navigation path to match the preoperative path and the intraoperative image in real time, provides an augmented reality view, and performs real-time navigation display to obtain real-time position coordinates and intraoperative images;

[0038] The navigation feedback and error correction module corrects the navigation trajectory through Kalman filtering based on the real-time position coordinates and intraoperative images.

[0039] Embodiment 2. This embodiment is based on Embodiment 1. In this embodiment, the process of using the structure bias registration model to register the standardized image voxel data to generate a multimodal fusion registration image specifically includes the following steps:

[0040] Step S1: Perform intensity normalization, unify the gray scale range, unify the resolution, and remove the background on the standardized image voxel data to obtain a pair of standardized images;

[0041] Step S2: Use the standardized image pairs as the source image and the real target image and input them into the reversible neural network. Perform modal translation on the source image through the affine coupling structure to generate a pseudo-target image. Use the reversible structure to perform inverse mapping on the pseudo-target image to restore the structural information of the source image, ensuring that the structural information is not lost during the modal translation process. Embed the dynamic depthwise separable convolution mechanism as a local attention path in the affine coupling structure. Perform sparse connection and dynamic weighting through the driven kernel combination to enhance the ability to extract local structural details and generate an enhanced pseudo-image. Introduce a barrier mutual information loss function regulated by the structural gradient to measure the mutual information between the enhanced pseudo-image and the real target image, limit the degree of structural distortion between the images, and obtain a structure-preserving image;

[0042] Dynamic depthwise separable convolution mechanism: To enhance the ability to extract structural details during the image modal translation process, embed the dynamic depthwise separable convolution mechanism in the affine coupling structure of the reversible neural network. This mechanism uses the input-driven kernel generation module to generate candidate convolution kernels, and realizes the enhancement of local structural details through sparse connection and dynamic weighting, thereby enhancing the ability to express local structural information while maintaining the lightweight nature of the network and ensuring the geometric consistency of the images before and after translation;

[0043] Formula of the barrier mutual information loss function regulated by the structural gradient:

[0044] ;

[0045] where, represents the barrier mutual information loss function regulated by the structural gradient, represents the enhanced pseudo-image, represents the real target image, represents the barrier threshold, represents the rectified linear unit function, represents the normalized mutual information between the pseudo-target image and the real target image; represents the weight coefficient of the gradient term, represents the gradient map of represents the gradient map of represents the L1 norm;

[0046] Step S3: Define the pseudo-factor set. Input the structure-preserving image into the backbone network to generate a prediction result. Perform conditional distance correlation measurement between the prediction result and the pseudo-factor set. Introduce the kernel-estimated CIDER regularization term to suppress the interference of pseudo-factors and realize the bias constraint optimization guided by the structure to obtain a bias-constrained prediction image;

[0047] CIDER is the abbreviation of the Conditional Independent Distance Correlation Regularization Method;

[0048] The loss function in the kernel estimation CIDER regularization term is defined as follows:

[0049] ;

[0050] in, represents the CIDER bias regularization loss function, Represents the prediction result, represents the pseudo factor set, represents a conditional variable, Indicates that given a condition variable Under the premise of and pseudo factors The conditional distance correlation between represents the regularization weight coefficient, represents the high-order dependency penalty term based on kernel estimation;

[0051] Conditional distance correlation measurement is a method used to measure the degree of statistical dependence between two variables given a third conditional variable. It is a dependency measurement method that introduces conditional variables based on distance correlation.

[0052] The kernel estimation CIDER regularization term is a new bias constraint mechanism. It introduces a kernel estimation-based method for the first time to enhance the modeling of complex pseudo-factor dependencies. It builds a structure-aware bias control strategy by integrating a high-order kernel mapping mechanism, which significantly improves the robustness and interpretability of pseudo-factor interference in the multimodal registration process.

[0053] Step S4: input the bias-constrained predicted image and the real target image into the registration network, learn the non-rigid deformation field, use the spatial transformer to spatially twist the bias-constrained predicted image, and generate an initial registration image;

[0054] Step S5: Combine steps S2 to S4 to construct a joint optimization objective function, train the structural bias registration model, further optimize the geometric consistency and modality fusion consistency of the initial registration image, and generate a multimodal fusion registration image. The formula used is as follows:

[0055] ;

[0056] in, represents the total loss function of the joint optimization objective function, represents the smooth regularization term of the non-rigid deformation field, , and Represents the weight coefficient.

[0057] Embodiment 3. This embodiment is based on Embodiment 1. In this embodiment, the process of registering the standardized image voxel data to generate a multimodal fusion registration image specifically includes the following steps:

[0058] Step R1: Perform intensity normalization, unify the gray scale range, unify the resolution, and remove the background on the standardized image voxel data to obtain a pair of standardized images;

[0059] Step R2: Take the pair of standardized images as the source image and the target image and input them into the registration model respectively. Perform modal conversion processing on the source image to generate a pseudo-target image, and perform a structure restoration operation to obtain a structure-enhanced image;

[0060] Step R3: Input the structure-enhanced image into the prediction network, output the prediction image result, and perform constraint processing in combination with a predefined set of pseudo-factors to generate a constrained prediction image;

[0061] Step R4: Input the constrained prediction image and the target image into the registration network together, learn the non-rigid deformation relationship between the images, and perform a registration operation through the spatial transformation module to generate an initial registration image;

[0062] Step R5: Jointly train the various network modules involved in the above steps to optimize the registration effect between the images and output a multimodal fusion registration image.

[0063] Embodiment 4. This embodiment is based on Embodiment 2. In this embodiment, the process of generating a preoperative navigation path through the rare path perception navigation SAC model specifically includes the following steps:

[0064] Step B1: Extract surgical anatomical structure information from the multimodal fusion registration image to construct a three-dimensional sparse voxel map. The three-dimensional sparse voxel map includes a point set and an edge set, where the point set represents path sampling points and the edge set represents feasible navigation connectivity, forming a structured three-dimensional navigation map;

[0065] Step B2: Based on the structured three-dimensional navigation map, collect the state-action pair sequence of its historical preoperative navigation trajectories. Perform semantic encoding on the state-action pair sequence through the Encoder network to obtain a set of feature embeddings, and then use the K-means algorithm to cluster the set of feature embeddings to generate a discretized label sequence, completing the structured abstraction of the trajectory behavior;

[0066] Step B3: Remove duplicates from the discretized label sequence to obtain a structural decision-making pattern, count the occurrence frequency of each pattern, and sort by frequency to obtain a set of rare patterns; Based on the set of rare patterns, construct the coverage ratio of rare paths;

[0067] The structural decision-making patterns include path topology feature patterns, local behavior decision-making patterns, state-action nested sequence patterns, navigation target tendency patterns, and operation avoidance strategy patterns;

[0068] These patterns are obtained through: state-action pair sequence → semantic embedding → feature vector clustering → label sequence → deduplication and statistical frequency, resulting in structured trajectory abstract labels, and those with low frequencies are determined as rare path behavior patterns;

[0069] Step B4: Based on the rare pattern set and the coverage ratio of rare paths, construct a three-dimensional space risk map, use the three-dimensional space risk map as external constraint information and embed it into the training environment of the SAC model to avoid the coverage ratio of high rare paths and generate trajectory experience pairs; the three-dimensional space risk map includes risk scores, the coverage ratio of rare paths, and heat map mapping;

[0070] Step B5: Input the trajectory experience pairs into the SAC model for policy training, output the action distribution in the current state through the policy network, estimate the Q function through the double Q network structure, and monitor the upper bound of the Q function valuation error through the Q value estimation error formula, and output the path planning policy to obtain the preoperative navigation path. The formula used is as follows:

[0071] Q value estimation error formula:

[0072] ;

[0073] Where, represents the state, represents the action, represents the expectation of all pairs of the state-action distribution under the policy in, represents the optimal Q value, represents the Q value under the policy represents the valuation error, represents the discount factor, represents the maximum reward, represents the coverage ratio of rare paths,

[0074] represents the approximation error.

[0074] Example 5. This example is based on Example 2. In this example, the process of generating the preoperative navigation path through the rare path perception navigation SAC model specifically includes the following steps:

[0075] Step E1: Extract surgical anatomical structure information from the multimodal fusion registration image to construct a three-dimensional sparse voxel map, which includes a point set and an edge set. The point set represents path sampling points, and the edge set represents feasible navigation connectivity, forming a structured three-dimensional navigation map;

[0076] Step E2: Based on the structured three-dimensional navigation map, collect the state-action pair sequence of its historical preoperative navigation trajectories. Semantically encode the state-action pair sequence through the Encoder network to obtain a feature embedding set, and then use the K-means algorithm to cluster the feature embedding set to generate a discretized label sequence, completing the structured abstraction of the trajectory behavior;

[0077] Step E3: Remove duplicates from the discretized label sequence to obtain a structural decision-making pattern, count the occurrence frequency of each pattern, and sort by frequency to obtain a rare pattern set; Based on the rare pattern set, construct the coverage ratio of rare paths;

[0078] Step E4: Based on the rare pattern set and the coverage ratio of rare paths, construct a three-dimensional space risk map, embed the three-dimensional space risk map as external constraint information into the training environment of the SAC model, avoid high coverage ratios of rare paths, and generate trajectory experience pairs;

[0079] Step E5: Input the trajectory experience pairs into the SAC model for policy training, output the action distribution in the current state through the policy network, estimate the Q function through the double Q network structure, output the path planning policy, and obtain the preoperative navigation path.

[0080] Example 6, according to Figure 2 , this example is based on Example 5. In this example, a preoperative path planning module is established to build a SAC model, optimize the policy training of the SAC model, construct a rare path-aware navigation SAC model, and process the multimodal fusion registration image through the rare path-aware navigation SAC model to generate a preoperative navigation path;

[0081] Preoperative navigation path:

[0082] Path length: 85.3 mm; Navigation obstacle avoidance radius: 4.5 mm; Average spacing: 1.02 mm;

[0083] The intraoperative tracking module uses visual SLAM technology to perform intraoperative registration based on the preoperative navigation path to match the preoperative path and the intraoperative image in real time, provide an augmented reality view, and perform real-time navigation display to obtain real-time position coordinates and intraoperative images;

[0084] Camera frame rate: 30 fps;

[0085] The delay is controlled within <80ms;

[0086] Real-time position coordinates:

[0087] t = 0 min: Real-time coordinates (134.2, 88.3, 45.1);

[0088] t = 7 min: Real-time coordinates (139.8, 93.2, 50.6);

[0089] t = 15 min: Real-time coordinates (188.5, 123.6, 66.3);

[0090] The navigation feedback and error correction module corrects the navigation trajectory through Kalman filtering based on the real-time position coordinates and intraoperative images;

[0091] Generate a navigation feedback and error correction map.

[0092] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto; Generally speaking, if those of ordinary skill in the art are inspired by it and without departing from the purpose of the present invention, without creative design, structures and embodiments similar to the technical solution should fall within the protection scope of the present invention.

Claims

1. A surgical navigation system based on multimodal images, including an image preprocessing module, which generates standardized image voxel data; characterized in that: The system further includes an image registration module and a preoperative path planning module; The image registration module constructs a structure bias registration model, uses the structure bias registration model to register the standardized image voxel data, and generates a multi-modal fusion registration image; The structure bias registration model includes a preprocessing unit, a modality translation unit, a bias optimization unit, a registration unit, a training unit, and a reversible neural network; The preoperative path planning module processes the multi-modal fusion registration image through a rare path perception navigation SAC model to generate a preoperative navigation path. The rare path perception navigation SAC model includes a SAC model.

2. The surgical navigation system based on multi-modal images according to claim 1, wherein: The preprocessing unit preprocesses the standardized image voxel data to obtain a pair of standardized images.

3. The surgical navigation system based on multimodal images according to claim 2, characterized in that: The modality translation unit divides the pair of standardized images into a source image and a true target image and inputs them into the reversible neural network. The modality of the source image is translated using an affine coupling structure; a dynamic depthwise separable convolution mechanism is embedded in the affine coupling structure as a local attention path, and sparse connection and dynamic weighting are performed through a driven kernel combination to generate an enhanced pseudo-image; a barrier mutual information loss function regulated by a structural gradient is introduced to measure the mutual information between the enhanced pseudo-image and the true target image to obtain a structure-preserving image.

4. The surgical navigation system based on multimodal images according to claim 3, wherein: The bias optimization unit defines a set of pseudo-factors, combines the structure-preserving image, generates a prediction result, performs a conditional distance correlation measurement on the prediction result and the set of pseudo-factors, and introduces a kernel-estimated CIDER regularization term to suppress the interference of pseudo-factors to obtain a bias-constrained prediction image.

5. The surgical navigation system based on multimodal images according to claim 4, wherein: The registration unit generates an initial registration image based on the bias-constrained prediction image and the true target image.

6. The surgical navigation system based on multimodal images according to claim 5, characterized in that: The training unit constructs a joint optimization objective function, trains the structure bias registration model, optimizes the initial registration image, and generates a multi-modal fusion registration image.

7. The surgical navigation system based on multimodal images according to claim 1, wherein: The process of generating a preoperative navigation path through the rare path perception navigation SAC model specifically includes the following steps: Step B1: Form a structured three-dimensional navigation map based on the multi-modal fusion registration image; Step B2: Generate a discretized label sequence based on the structured three-dimensional navigation map; Step B3: Remove duplicates from the discretized label sequence, obtain a structural decision-making pattern, count the occurrence frequency of each pattern in the structural decision-making pattern, and sort by frequency to obtain a set of rare patterns; based on the set of rare patterns, construct a coverage ratio of rare paths; Step B4: Based on the set of rare patterns and the coverage ratio of rare paths, construct a three-dimensional space risk map as external constraint information and embed it into the training environment of the SAC model to avoid high coverage ratios of rare paths and generate trajectory experience pairs; Step B5: Input the trajectory experience pairs into the SAC model for policy training, estimate the Q function through a double Q network structure, and monitor the upper bound of the Q function valuation error through a Q value estimation error formula, output a path planning policy, and obtain a preoperative navigation path.

8. The surgical navigation system based on multimodal images according to claim 7, wherein: The three-dimensional space risk map includes a risk score, a coverage ratio of rare paths, and a heat map mapping.

Images