A method and system for generating navigation information for assisting a neurovascular intervention procedure

Abstract

This invention discloses a method and system for generating navigation information to assist neurovascular interventional surgery. The method first acquires preoperative three-dimensional vascular images and extracts the vascular tree structure, then acquires an intraoperative two-dimensional perspective roadmap. A non-rigid registration algorithm based on thin-plate splines is used to spatially align the three-dimensional vessels with the two-dimensional roadmap, effectively handling the non-rigid deformation of the vessels. High-dimensional feature mapping of the three-dimensional vascular depth information is performed using sinusoidal position coding to generate a depth-coded feature map. A multi-scale attention mechanism is used to fuse the depth-coded features with the two-dimensional image features. Finally, an intelligent agent network based on deep reinforcement learning uses the fused features to make navigation decisions. This invention solves the problems of insufficient registration accuracy and inadequate utilization of depth information in existing technologies, providing rich three-dimensional spatial information for the intelligent agent and significantly improving the performance and safety of intelligent navigation in neurovascular interventional surgery.

Description

Technical Field

[0001] The embodiments of the present invention relate to the fields of medical image processing and artificial intelligence technology, specifically to a method and system for generating auxiliary navigation information for neurovascular interventional surgery. Background Technology

[0002] Neurovascular interventional surgery is an important minimally invasive surgical procedure for treating cerebrovascular diseases. During the procedure, doctors need to use digital subtraction angiography (DSA) to observe the position and movement of catheters, guidewires, and other instruments within the blood vessels in real time.

[0003] However, DSA imaging only provides two-dimensional projection images, requiring surgeons to reconstruct the three-dimensional vascular structure in their minds based on experience. This can easily lead to difficulties in instrument positioning in complex vascular bifurcation areas or tortuous vascular segments. To address this, existing technologies attempt to register and fuse preoperative three-dimensional vascular images (such as CTA, MRA, or 3D-DSA) with intraoperative two-dimensional fluoroscopic images to assist surgeons in surgical navigation.

[0004] In terms of registration and fusion, existing studies have employed non-rigid registration techniques to address tissue deformation issues. For example, CN121196735A discloses a 3D image navigation system for minimally invasive gynecological surgery, which uses a non-rigid registration algorithm of dual-stream convolution-deformation fusion to register the preoperative 3D model with the intraoperative endoscopic video; CN120656639A discloses an intelligent planning system for thoracic surgery, which uses a multi-level feature fusion algorithm to achieve non-rigid registration of multimodal images; and CN121601155A discloses a liver surgery simulation system, which uses a multi-scale dynamic registration module to achieve non-rigid registration of multimodal images and respiratory motion compensation.

[0005] However, directly applying the above techniques to neurovascular interventional surgery still presents the following technical challenges:

[0006] First, depth information is not fully utilized. Although existing non-rigid registration methods achieve geometric alignment between 3D models and 2D images, the registered 3D depth information is not effectively utilized by neural networks. Existing navigation systems are mainly trained based on 2D image features or only use 3D models as static overlay layers, lacking a mechanism to transform 3D vascular depth information into a learnable feature representation for neural networks. This limits the agent's ability to understand complex vascular spatial structures.

[0007] Second, there is insufficient preservation of vascular topology. Existing non-rigid registration methods (such as the two-stream network in CN121196735A) mainly rely on image features for deformation field estimation. For the specific scenario of neurovascular intervention, the vascular structure is delicate, with complex branches and dense bifurcation points. The lack of topological constraints can easily lead to topological errors such as bifurcation point misalignment and vascular crossing in the registration results, thus affecting navigation accuracy.

[0008] Third, navigation decisions lack adaptability. Existing surgical navigation systems mainly employ static path planning algorithms (such as the NSGA-II algorithm in CN120656639A and the improved A* algorithm in CN121196735A). The path schemes generated are predetermined and cannot be dynamically adjusted frame-by-frame according to real-time changes in the intraoperative environment (such as instrument position and vascular deformation). Although the aforementioned systems have the ability to update the path intraoperatively, the updates are conditionally triggered (such as registration error exceeding limits or manual intervention by the surgeon) rather than continuous real-time responses, making it difficult to meet the real-time adaptive navigation requirements of neurovascular interventional surgery. Summary of the Invention

[0009] Therefore, embodiments of the present invention provide a method and system for generating auxiliary navigation information for neurovascular interventional surgery to solve the technical problems of insufficient registration accuracy and inadequate utilization of depth information in the prior art.

[0010] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0011] According to a first aspect of the present invention, a method for generating navigation information for neurovascular interventional surgery is provided, the method comprising:

[0012] Acquire preoperative three-dimensional vascular imaging data and extract the three-dimensional vascular tree structure, including the vascular centerline, vascular radius, and vascular topology.

[0013] Acquire intraoperative two-dimensional fluoroscopic path image sequence;

[0014] For each frame in the two-dimensional perspective road map image sequence, the three-dimensional vascular tree and the two-dimensional road map are spatially aligned in real time frame by frame based on a non-rigid registration algorithm to establish the correspondence between three-dimensional spatial points and two-dimensional projection points;

[0015] Sine position encoding is performed on the depth information of each point in the registered 3D vascular tree along the projection direction to generate a depth-encoded feature map of the same size as the 2D road map.

[0016] Multi-scale attention fusion is performed between deep encoded feature maps and 2D road map features to generate fused feature maps;

[0017] The fused feature map and catheter status information are input into a deep reinforcement learning model trained offline, and navigation prompts are output in real time to assist doctors in decision-making. The navigation prompts include specific operation instructions for the catheter / guidewire, target path or operation warning.

[0018] Furthermore, non-rigid registration algorithms include:

[0019] The three-dimensional blood vessel centerline is sampled to obtain a set of control points;

[0020] Perform vessel segmentation on the current frame's two-dimensional road map, extract the two-dimensional vessel centerline, and obtain the target point set;

[0021] Establish a non-rigid deformation model based on thin plate splines;

[0022] An optimization strategy combining iterative nearest point and thin plate spline is adopted, and corresponding point matching and deformation parameter optimization are performed frame by frame using GPU acceleration, with a single frame processing time of less than 50ms;

[0023] By introducing hard constraints on vascular topology, the correspondence between vascular bifurcation points is forcibly maintained, thus avoiding navigation deviations caused by topological errors.

[0024] Further, the feature is that the vascular topological hard constraint includes:

[0025] Forced mapping: A mapping must be established between the three-dimensional blood vessel bifurcation point and the corresponding bifurcation point in the two-dimensional roadmap;

[0026] No crossing: The registered 3D vessel centerline projections must not cross each other to avoid topological ambiguity;

[0027] Connectivity maintenance: Maintaining consistent blood vessel connectivity before and after registration ensures the anatomical rationality of the navigation path.

[0028] Furthermore, depth information encoding methods include:

[0029] The depth value of each sampling point in the registered 3D vascular tree along the X-ray projection direction is obtained, and the depth value is the distance from the sampling point to the X-ray source or the distance to the imaging plane.

[0030] The depth values ​​are linearly normalized to the [0,1] interval;

[0031] A sinusoidal positional encoding method is used to perform D-dimensional high-dimensional feature mapping on the normalized depth values. The sinusoidal positional encoding method uses the following formula:

[0032]

[0033]

[0034] in, The normalized depth value is k=0,1,...,D / 2-1, where D is the encoding dimension and D≥64;

[0035] The encoded D-dimensional depth features are filled according to the corresponding two-dimensional projection positions, and the blood vessel width is covered by Gaussian diffusion to generate an H×W×D depth encoded feature map of the same size as the two-dimensional road map.

[0036] Furthermore, the feature fusion employs a multi-scale feature fusion strategy, including:

[0037] Multi-scale features of a two-dimensional road map are extracted using a convolutional neural network, wherein the multi-scale features include 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original map resolution;

[0038] The depth-encoded feature map is downsampled step by step through a convolutional layer with a stride of 2 to adjust it to the same spatial resolution as the two-dimensional features at each scale;

[0039] A cross-modal attention mechanism is used for fusion at various scales: two-dimensional features are used as queries, and deep features are used as keys and values. The attention weights are calculated and then added to the original two-dimensional features.

[0040] The fused multi-scale features are aggregated through a Feature Pyramid Network (FPN) to generate a fused feature map.

[0041] Furthermore, methods for generating navigation prompts include:

[0042] A deep reinforcement learning model trained offline is used to perform real-time inference processing on the fused feature map and output navigation prompt information. The inference time of the model is less than 10ms.

[0043] The deep reinforcement learning model employs a policy network trained using the PPO or SAC algorithm, and its state space includes: fused feature map, two-dimensional position of the catheter tip, catheter attitude angle, and the history of the previous 5 actions.

[0044] The action space is a discrete set of actions that directly outputs specific operational instructions for the catheter / guidewire rather than geometric paths.

[0045] Furthermore, the deep reinforcement learning model is trained offline in a simulated environment that includes the following elements:

[0046] The virtual vascular environment is constructed based on the patient's preoperative three-dimensional vascular tree, including physical parameters such as vascular elasticity, friction coefficient, and blood flow resistance.

[0047] A catheter dynamics model was used to simulate the bending and springback behavior of a flexible catheter under pushing and rotating operations.

[0048] A heart-respiratory motion model that simulates the periodic deformation of blood vessels during the physiological cycle;

[0049] Through millions of virtual operation iterations, the system learns a mapping strategy from state to action, enabling the trained policy network to output optimal operation instructions in real time during actual surgery.

[0050] According to a second aspect of the present invention, a neurovascular interventional surgery-aided navigation information generation system is provided, the system comprising:

[0051] The 3D vascular reconstruction module is used to acquire preoperative 3D vascular image data and extract the 3D vascular tree structure, including the vascular centerline, vascular radius, and vascular topology.

[0052] The two-dimensional image acquisition module is used to acquire intraoperative two-dimensional fluoroscopic path image sequences;

[0053] The non-rigid registration module is used to perform real-time spatial alignment of the three-dimensional vascular tree and the two-dimensional road map frame by frame in the two-dimensional perspective road map image sequence based on the non-rigid registration algorithm, and establish the correspondence between three-dimensional spatial points and two-dimensional projection points.

[0054] The depth encoding module is used to perform sinusoidal position encoding on the depth information of each point in the registered 3D vascular tree along the projection direction, generating a depth-encoded feature map of the same size as the 2D road map.

[0055] The feature fusion module is used to perform multi-scale attention fusion of deep encoded feature maps and two-dimensional road map features to generate a fused feature map;

[0056] The intelligent navigation module is used to input the fused feature map and catheter status information into the offline trained deep reinforcement learning model, and output navigation prompts in real time to assist doctors in decision-making. The navigation prompts include specific operation instructions for the catheter / guidewire, target path or operation warning.

[0057] Furthermore, it also includes a human-computer interaction module, configured as a medical display and touch interface, used to display the fused navigation image, specific operation instructions and confidence level of the catheter / guidewire in real time, and to receive confirmation or modification instructions from the doctor;

[0058] The specific operation instructions are presented in the form of discrete actions, including suggestions for forward distance and rotation angle. Doctors can confirm and execute the operation with one click through the touch interface or manually adjust the parameters before execution.

[0059] Furthermore, the intelligent navigation module employs a policy network trained offline using PPO or SAC reinforcement learning algorithms to generate navigation prompts.

[0060] The embodiments of the present invention have the following advantages:

[0061] (1) Realized real-time adaptive navigation during surgery: The deep reinforcement learning framework is adopted, and the fused feature map is used as the state input. The navigation decision is output frame by frame in real time through the offline trained policy network. The state space includes the catheter position, attitude and historical action information, which enables the navigation system to have temporal reasoning ability and can dynamically adjust the strategy according to the real-time changes in the intraoperative environment. This overcomes the defect that the static path planning scheme in the existing technology cannot continuously respond to environmental changes.

[0062] (2) Significantly improves three-dimensional spatial perception capability: The three-dimensional blood vessel depth information is mapped into a high-dimensional feature vector through sinusoidal position encoding, generating a depth encoding feature map of the same size as the two-dimensional road map, and the fusion of depth features and two-dimensional visual features is achieved by adopting a cross-modal attention mechanism, enabling the agent to explicitly perceive the three-dimensional spatial structure of blood vessels, thus solving the problem of insufficient utilization of depth information in the existing technology;

[0063] (3) Ensured the anatomical rationality of the registration results: In the non-rigid registration process, the topological constraints of blood vessels were introduced, and the correspondence of the bifurcation points of blood vessels was incorporated into the optimization objective as a hard constraint, which avoided topological errors such as bifurcation point misalignment and blood vessel crossing in the registration results and improved the registration accuracy;

[0064] (4) Optimized end-to-end decision quality: By designing a reward function that integrates distance, safety, efficiency and objective, and training the policy network with PPO or SAC algorithm, end-to-end optimization with dynamic trade-off of multiple factors is achieved. Compared with the existing multi-objective genetic algorithm or heuristic search, it is more adaptable to the complex decision-making needs of neurovascular interventional surgery. Attached Figure Description

[0065] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0066] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0067] Figure 1A schematic diagram of the logical structure of a neurovascular interventional surgery auxiliary navigation information generation system provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the overall process of a method for generating auxiliary navigation information for neurovascular interventional surgery provided in an embodiment of the present invention. Detailed Implementation

[0069] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] In order to solve the above-mentioned technical problems of insufficient registration accuracy and inadequate utilization of depth information.

[0071] refer to Figure 1 This invention discloses a neurovascular interventional surgery auxiliary navigation information generation system, including: a three-dimensional vascular reconstruction module, configured as a workstation-level computer, equipped with vascular segmentation and reconstruction software, receiving preoperative three-dimensional image data (DICOM format), and outputting three-dimensional vascular tree structure data;

[0072] The two-dimensional image acquisition module connects to the DSA device via DICOM or a dedicated interface to acquire perspective image sequences in real time at a frame rate of no less than 15fps.

[0073] The non-rigid registration module is implemented using a GPU-accelerated TPS registration algorithm, with a single-frame registration time of less than 50ms, meeting real-time requirements.

[0074] The depth coding module uses a tensor operation library to implement sinusoidal position coding and outputs a 64-channel depth-coded feature map.

[0075] The feature fusion module is implemented using a pre-trained ResNet-50 and a cross-modal attention network, and is deployed on a high-performance GPU (such as NVIDIA RTX 4090).

[0076] The intelligent navigation module includes an offline-trained PPO policy network with an inference time of less than 10ms;

[0077] The human-computer interaction module is configured as a medical display and touch interface, which displays the fused navigation image, recommended operation instructions and confidence level in real time, and receives confirmation or modification instructions from doctors.

[0078] Furthermore, the non-rigid registration module employs a GPU-accelerated thin-plate spline registration algorithm, with a single-frame registration time of less than 50ms.

[0079] Furthermore, the intelligent navigation module employs a policy network trained offline using PPO or SAC reinforcement learning algorithms to generate navigation prompts.

[0080] Corresponding to the aforementioned system for generating navigation information for neurovascular interventional surgery, this invention also discloses a method for generating navigation information for neurovascular interventional surgery. The following details a method for generating navigation information for neurovascular interventional surgery disclosed in this invention, in conjunction with the aforementioned system for generating navigation information for neurovascular interventional surgery.

[0081] refer to Figure 2 This invention discloses a method for generating navigation information to assist in neurovascular interventional surgery, the method comprising:

[0082] Acquire preoperative three-dimensional vascular image data and extract a three-dimensional vascular tree structure, which includes the vascular centerline and vascular topology;

[0083] Acquire intraoperative two-dimensional fluoroscopic path image sequence;

[0084] A non-rigid registration algorithm with vascular topology constraints is used to spatially align the 3D vascular tree with the 2D road map, establish the correspondence between each sampling point on the 3D vascular centerline and the 2D projection point, and maintain the topological correspondence consistency of vascular bifurcation points.

[0085] Based on the correspondence, the depth values ​​of each sampling point in the three-dimensional vascular tree along the projection direction are obtained. The depth values ​​are mapped to high-dimensional depth feature vectors using sinusoidal position encoding. The vectors are then filled according to the two-dimensional projection positions corresponding to each sampling point to generate a multi-channel depth-encoded feature map of the same size as the two-dimensional road map.

[0086] A cross-modal attention mechanism is used to fuse the deep encoded feature map with the multi-scale features of the convolutional neural network of the two-dimensional road map to generate a fused feature map;

[0087] The fused feature map is input into a deep reinforcement learning model trained offline, and navigation prompts are output to assist doctors in decision-making. The navigation prompts include at least one of the following: suggested direction of catheter / guidewire, target path, or operation warning.

[0088] Furthermore, non-rigid registration algorithms include:

[0089] The three-dimensional blood vessel centerline is sampled to obtain a set of control points;

[0090] Perform vessel segmentation on the 2D road map, extract the 2D vessel centerline, and obtain the target point set;

[0091] Establish a non-rigid deformation model based on thin plate splines;

[0092] An optimization strategy combining iterative nearest point and thin plate spline is adopted to perform corresponding point matching and deformation parameter optimization;

[0093] Introduce vascular topological constraints to maintain the correspondence between vascular bifurcation points.

[0094] Furthermore, depth information encoding methods include:

[0095] Obtain the depth value along the projection direction for each sampling point in the registered 3D vascular tree;

[0096] Normalize the depth values;

[0097] Sine positional coding is used to perform high-dimensional feature mapping on normalized depth values;

[0098] The encoded depth features are used to generate a depth-encoded feature map of the same size as the two-dimensional road map, based on the corresponding two-dimensional projection positions.

[0099] Furthermore, the sinusoidal position encoding adopts the following formula:

[0100]

[0101]

[0102] in, The normalized depth value is k=0,1,...,D / 2-1, where D is the encoding dimension.

[0103] Furthermore, the feature fusion employs a multi-scale feature fusion strategy, including:

[0104] Using convolutional neural networks to extract multi-scale features from two-dimensional road maps;

[0105] Adjust the deep encoded feature map to the same spatial resolution as the features at each scale;

[0106] An attention mechanism is used for cross-modal feature fusion.

[0107] Furthermore, methods for generating navigation prompts include:

[0108] An offline-trained deep reinforcement learning model is used to process the fused feature map and output navigation prompts.

[0109] The state space of the deep reinforcement learning model includes fused feature maps and duct state information, the action space includes suggestions for forward, backward and rotation operations, and the model's reward function is designed based on path planning efficiency and virtual collision detection in a simulated environment.

[0110] like Figure 1 As shown, a method for generating navigation information for neurovascular interventional surgery, applied to cerebral aneurysm embolization surgery, specifically includes the following steps:

[0111] Step 1: Preoperative three-dimensional vascular image processing

[0112] Preoperative 3D-DSA or CTA image data of the patient were acquired, and vascular region segmentation was performed using a deep learning-based vascular segmentation network (such as 3DU-Net). The segmentation results were then refined in three dimensions to extract the vascular centerlines. A distance-transform-based method was used to calculate the vascular radius at each centerline point. By analyzing the connectivity of the centerlines, a vascular topology tree structure was constructed.

[0113] Step 2: Intraoperative 2D Road Map Acquisition

[0114] Intraoperative fluoroscopic image sequences were acquired in real time using a DSA device. The raw images underwent preprocessing, including noise removal and contrast enhancement. Two-dimensional vascular structures were extracted using a vascular segmentation algorithm.

[0115] Step 3: Non-rigid registration

[0116] like Figure 2 As shown, the non-rigid registration process includes:

[0117] (1) Initialization: Based on the geometric parameters of the DSA device (source-to-detector distance, pixel spacing, etc.), an initial projection matrix is ​​established. The three-dimensional blood vessel centerline is projected onto a two-dimensional plane to obtain the initial correspondence.

[0118] (2) Corresponding point search: For the projection of each three-dimensional sampling point, the nearest neighbor point is searched on the two-dimensional blood vessel centerline as the corresponding point. Constraints based on blood vessel orientation are used to eliminate obviously erroneous correspondences.

[0119] (3) Thin plate spline deformation: Based on the current correspondence, solve for the TPS deformation parameters. Apply the deformation field to the three-dimensional blood vessel centerline.

[0120] (4) Iterative optimization: Repeat steps (2) and (3) until the registration error converges or the maximum number of iterations is reached. During the optimization process, hard constraints on the bifurcation points of blood vessels are introduced to ensure that the correspondence between the bifurcation points is correct.

[0121] Step 4: Depth Information Encoding

[0122] The deep encoding process includes:

[0123] (1) Depth value extraction: For each sampling point on the registered three-dimensional blood vessel centerline, calculate its depth value along the X-ray projection direction (i.e., the distance to the X-ray source or the distance to the imaging plane).

[0124] (2) Depth normalization: Statistically normalize the depth value range of all sampling points to the [0,1] interval.

[0125] (3) Sine position encoding: A coding dimension of D=64 is adopted, and the normalized depth values ​​are mapped to a high dimension using sine and cosine functions. After encoding, each sampling point corresponds to a 64-dimensional depth feature vector.

[0126] (4) Feature map generation: Initialize an H×W×D zero tensor as a depth encoding feature map. Based on the registration results, fill the depth feature vector of each 3D sampling point into the corresponding 2D projection position. Perform Gaussian diffusion on the blood vessel region to ensure that the depth encoding covers the entire width of the blood vessel.

[0127] Step 5: Feature Fusion

[0128] The following feature fusion strategy is adopted:

[0129] ResNet-50 was used as the backbone network for 2D feature extraction, extracting feature maps at four scales (1 / 4, 1 / 8, 1 / 16, and 1 / 32 resolution of the original image, respectively). The depth-encoded feature maps were downsampled stepwise through convolutional layers with a stride of 2 to obtain depth features at the same resolution as the 2D features. At each scale, a cross-modal attention mechanism was used for fusion: the 2D features were used as the query, and the depth features as the key and value; attention weights were calculated and then added to the original 2D features. The fused multi-scale features were aggregated using an FPN (Feature Pyramid Network) structure.

[0130] Step Six: Intelligent Navigation Decision

[0131] The agent network is trained using the PPO algorithm, with the following specific settings:

[0132] State space: fused feature map (256×256×256 dimensional) + catheter tip position (2 dimensional) + catheter attitude angle (1 dimensional) + history of the previous 5 steps (5 dimensional).

[0133] Action space: Discrete action set {move forward 1mm, move forward 2mm, move backward 1mm, rotate clockwise 10°, rotate counterclockwise 10°, remain stationary}.

[0134] Reward function design:

[0135] R = Rdist + Rsafe + Rtime + Rreach

[0136] Where: Rdist = -α·Δd, where Δd is the change in distance between the duct tip and the target; Rsafe = -β·I(collision), which is penalized when a collision occurs; Rtime = -γ, which is a time step penalty to encourage efficient operation; and Rreach = +R0, which is a large reward when the target is successfully reached.

[0137] The embodiments of the present invention have the following advantages:

[0138] (1) The non-rigid registration algorithm based on thin plate splines can effectively handle the deformation of blood vessels under the influence of physiological motion, and significantly improve the registration accuracy between three-dimensional blood vessels and two-dimensional road maps;

[0139] (2) Introduce a depth information encoding mechanism to transform the depth information of three-dimensional blood vessels into a feature representation that can be learned by the neural network, so that the intelligent agent can perceive the three-dimensional spatial structure of blood vessels.

[0140] (3) Using sinusoidal position coding to perform high-dimensional mapping of depth information preserves the relative relationship between depth values ​​and enhances the expressive power of features;

[0141] (4) Through the multi-scale feature fusion strategy, two-dimensional visual features and three-dimensional depth information are effectively integrated, providing the agent with a richer environmental representation;

[0142] (5) Registration optimization by introducing vascular topology constraints ensures the anatomical rationality of the registration results and avoids navigation deviation caused by topology errors.

[0143] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method of generating navigation information for assisting a neurovascular intervention procedure, characterized by, The method includes: Acquire preoperative three-dimensional vascular imaging data and extract the three-dimensional vascular tree structure, including the vascular centerline, vascular radius, and vascular topology. Acquire intraoperative two-dimensional fluoroscopic path diagram image sequence; For each frame in the two-dimensional perspective road map image sequence, the three-dimensional vascular tree and the two-dimensional road map are spatially aligned in real time frame by frame based on a non-rigid registration algorithm to establish the correspondence between three-dimensional spatial points and two-dimensional projection points; Sine position encoding is performed on the depth information of each point in the registered 3D vascular tree along the projection direction to generate a depth-encoded feature map of the same size as the 2D road map. Multi-scale attention fusion is performed between deep encoded feature maps and 2D road map features to generate fused feature maps; The fused feature map and catheter status information are input into a deep reinforcement learning model trained offline, and navigation prompts are output in real time to assist doctors in decision-making. The navigation prompts include specific operation instructions for the catheter / guidewire, target path or operation warning.

2. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 1, characterized in that, The non-rigid registration algorithm includes: The three-dimensional blood vessel centerline is sampled to obtain a set of control points; Perform vessel segmentation on the current frame's two-dimensional road map, extract the two-dimensional vessel centerline, and obtain the target point set; Establish a non-rigid deformation model based on thin plate splines; An optimization strategy combining iterative nearest point and thin plate spline is adopted, and corresponding point matching and deformation parameter optimization are performed frame by frame using GPU acceleration, with a single frame processing time of less than 50ms; By introducing hard constraints on vascular topology, the correspondence between vascular bifurcation points is forcibly maintained, thus avoiding navigation deviations caused by topological errors.

3. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 2, characterized in that, The hard constraint on vascular topology includes: Forced mapping: A mapping must be established between the three-dimensional blood vessel bifurcation point and the corresponding bifurcation point in the two-dimensional roadmap; No crossing: The registered 3D vessel centerline projections must not cross each other to avoid topological ambiguity; Connectivity maintenance: Maintaining consistent blood vessel connectivity before and after registration ensures the anatomical rationality of the navigation path.

4. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 1, characterized in that, Deep information encoding methods include: The depth value of each sampling point in the registered 3D vascular tree along the X-ray projection direction is obtained, and the depth value is the distance from the sampling point to the X-ray source or the distance to the imaging plane. The depth values ​​are linearly normalized to the [0,1] interval; A sinusoidal positional encoding method is used to perform D-dimensional high-dimensional feature mapping on the normalized depth values. The sinusoidal positional encoding method uses the following formula: in, The normalized depth value is k=0,1,...,D / 2-1, where D is the encoding dimension and D≥64; The encoded D-dimensional depth features are filled according to the corresponding two-dimensional projection positions, and the blood vessel width is covered by Gaussian diffusion to generate an H×W×D depth encoded feature map of the same size as the two-dimensional road map.

5. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 1, characterized in that, Feature fusion employs a multi-scale feature fusion strategy, including: Multi-scale features of a two-dimensional road map are extracted using a convolutional neural network, wherein the multi-scale features include 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original map resolution; The depth-encoded feature map is downsampled step by step through a convolutional layer with a stride of 2 to adjust it to the same spatial resolution as the two-dimensional features at each scale; A cross-modal attention mechanism is used for fusion at various scales: two-dimensional features are used as queries, and deep features are used as keys and values. The attention weights are calculated and then added to the original two-dimensional features. The fused multi-scale features are aggregated through a Feature Pyramid Network (FPN) to generate a fused feature map.

6. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 1, characterized in that, Methods for generating navigation tips include: A deep reinforcement learning model trained offline is used to perform real-time inference processing on the fused feature map and output navigation prompt information. The inference time of the model is less than 10ms. The deep reinforcement learning model employs a policy network trained using the PPO or SAC algorithm, and its state space includes: fused feature map, two-dimensional position of the catheter tip, catheter attitude angle, and the history of the previous 5 actions. The action space is a discrete set of actions that directly outputs specific operational instructions for the catheter / guidewire rather than geometric paths.

7. The method for generating auxiliary navigation information for neurovascular interventional surgery as described in claim 6, characterized in that, The deep reinforcement learning model is trained offline in a simulated environment that includes the following elements: The virtual vascular environment is constructed based on the patient's preoperative three-dimensional vascular tree, including physical parameters such as vascular elasticity, friction coefficient, and blood flow resistance. A catheter dynamics model was used to simulate the bending and springback behavior of a flexible catheter under pushing and rotating operations. A heart-respiratory motion model that simulates the periodic deformation of blood vessels during the physiological cycle; Through millions of virtual operation iterations, the system learns a mapping strategy from state to action, enabling the trained policy network to output optimal operation instructions in real time during actual surgery.

8. A system for generating navigation information to assist in neurovascular interventional surgery, characterized in that, The system includes: The 3D vascular reconstruction module is used to acquire preoperative 3D vascular image data and extract the 3D vascular tree structure, including the vascular centerline, vascular radius, and vascular topology. The two-dimensional image acquisition module is used to acquire intraoperative two-dimensional fluoroscopic path image sequences; The non-rigid registration module is used to spatially align the three-dimensional vascular tree with the two-dimensional road map frame by frame in real time based on the non-rigid registration algorithm for each frame in the two-dimensional perspective road map image sequence, and establish the correspondence between three-dimensional spatial points and two-dimensional projection points. The depth encoding module is used to perform sinusoidal position encoding on the depth information of each point in the registered 3D vascular tree along the projection direction, generating a depth-encoded feature map of the same size as the 2D road map. The feature fusion module is used to perform multi-scale attention fusion of deep encoded feature maps and two-dimensional road map features to generate a fused feature map; The intelligent navigation module is used to input the fused feature map and catheter status information into the offline trained deep reinforcement learning model, and output navigation prompts in real time to assist doctors in decision-making. The navigation prompts include specific operation instructions for the catheter / guidewire, target path or operation warning.

9. The neurovascular interventional surgery auxiliary navigation information generation system as described in claim 8, characterized in that, It also includes a human-computer interaction module, configured as a medical display and touch interface, used to display the fused navigation image, specific operation instructions and confidence level of the catheter / guidewire in real time, and to receive confirmation or modification instructions from doctors; The specific operation instructions are presented in the form of discrete actions, including suggestions for forward distance and rotation angle. Doctors can confirm and execute the operation with one click through the touch interface or manually adjust the parameters before execution.

10. The neurovascular interventional surgery auxiliary navigation information generation system as described in claim 8, characterized in that, The intelligent navigation module uses a policy network trained offline using PPO or SAC reinforcement learning algorithms to generate navigation prompts.

Images