Cerebrovascular path planning method and system based on artificial intelligence

Abstract

The invention provides a cerebrovascular path planning method and system based on artificial intelligence, and the method comprises the steps: obtaining cerebrovascular three-dimensional image data, and generating an initial vascular network map; aggregating the node features of the blood vessel network atlas, calculating the weight in attention calculation according to the distance between the node and the key blood vessel region and morphological parameters during each iteration of the model, and outputting the path feature vector of the node; based on the path feature vector, performing path search by adopting an extended polyhedron algorithm EPA; in the iterative expansion process of the EPA algorithm, if a currently searched polyhedron cannot effectively approach a target point within a preset number of iterations or continuously violates any high-priority constraint in the multi-target optimization function for multiple times, adjusting the weight of a target function or temporarily relaxing a secondary constraint condition according to a pre-stored importance level of a blood vessel region; the expansion process is restarted from the adjacent candidate nodes; and calculating and outputting a cerebrovascular navigation path meeting clinical requirements.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a cerebrovascular path planning method and system based on artificial intelligence. Background Art

[0002] Cerebrovascular diseases, such as stroke, intracranial aneurysms, and arteriovenous malformations, are among the leading causes of death and severe disability. Endovascular intervention, with its minimally invasive and precise advantages, has gradually become an important approach for treating these complex cerebrovascular lesions. However, endovascular intervention requires precise navigation of minimally invasive instruments such as guidewires and catheters to the lesion site within the complex cerebral vascular network for treatment, placing extremely high demands on the accuracy, safety, and efficiency of preoperative path planning. The cerebral vascular system is characterized by its unique three-dimensional structure, highly tortuous course, and significant individual variability. This poses significant challenges to physicians when navigating instruments during surgery, and any minor operational error can lead to serious complications. Traditional interventional path planning relies primarily on manual or semi-manual decision-making by clinicians, combining their expertise and extensive experience with two-dimensional or three-dimensional medical imaging such as digital subtraction angiography, computed tomography angiography, or magnetic resonance angiography. This process is not only time-consuming and labor-intensive, but also highly dependent on the physician's experience. Moreover, due to the influence of human subjective factors, it is difficult to ensure the optimality, repeatability, and consistency of the planned path. Especially when faced with complex or variable vascular structures, its limitations are more prominent.

[0003] Computer-assisted vascular path planning algorithms, for example, extract the centerline skeleton of the vascular tree and then find the shortest path connecting the starting point and the target point. However, the effectiveness of such methods is largely limited by the accuracy and completeness of the vascular centerline extraction, and it is often difficult to extract the centerline of small blood vessels, diseased blood vessels, or complex bifurcation areas. At the same time, these algorithms usually only use path length as a single optimization goal, and fail to fully integrate and optimize other key clinical considerations, such as the local curvature of the blood vessel, the minimum lumen diameter, the branch angle, and the risk areas that need to be avoided, which may cause the planned path to be difficult to pass safely in clinical practice or not the best choice. Existing technologies still have a lot of room for improvement in improving the comprehensive clinical quality of planned paths, ensuring the robustness of the planning process, and the intelligent adaptability to complex anatomical structures. Summary of the Invention

[0004] To address technical issues with existing cerebrovascular path planning methods, such as insufficient depth in learning complex vascular network features, limited multi-objective optimization capabilities, and a lack of effective adjustments when path search encounters difficulties, this application proposes an artificial intelligence-based cerebrovascular path planning method, including:

[0005] Acquire 3D cerebral vascular imaging data and generate an initial vascular network map; perform regional division and hierarchical analysis on the initial vascular network map to identify trunks, branch vessels, and key vascular regions;

[0006] The Graph Transformer model is used to aggregate the node features of the vascular network graph. At each iteration, the model calculates the weight in the attention calculation based on the distance between the node and the key vascular area and the morphological parameters, and outputs the node's path feature vector;

[0007] Based on the path feature vector, an extended polyhedron algorithm (EPA) is used to perform path search. The EPA includes a multi-objective optimization function for path length, tortuosity, minimum vessel diameter, and safety distance. During the iterative expansion process of the EPA algorithm, if the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, the objective function weight is adjusted or secondary constraints are temporarily relaxed based on pre-stored vessel region importance levels, and the expansion process is restarted from a neighboring candidate node.

[0008] Calculate and output cerebral vascular navigation paths that meet clinical requirements.

[0009] Optionally, the model calculates the weight in the attention calculation based on the distance between the node and the key blood vessel area and the morphological parameters at each iteration, and outputs the path feature vector of the node, including:

[0010] Obtain node information of the vascular network map and identified key vascular area information;

[0011] For each node in the graph, in each iteration, the distance between the current node and each key vascular region is calculated, and the morphological parameters of the current node are extracted. Based on the distance and morphological parameters, the attention weight of the current node relative to each key vascular region is calculated using a preset attention mechanism function. The calculated attention weight is used to aggregate the features of neighboring nodes and update the features of the current node.

[0012] Through the iterative processing of multi-layer Graph Transformer until the preset maximum number of iterations is reached, the path feature vector from each node to the key vascular area is output.

[0013] Optionally, the calculation of the attention weight of the current node relative to each key blood vessel region by a preset attention mechanism function based on the distance and morphological parameters includes:

[0014] The distance and morphological parameters are digitized and normalized. Through the preset attention mechanism function, the inverse of the distance and the score of the morphological parameters are combined to calculate a raw attention score for each key vascular area for the current node.

[0015] The Softmax function is applied to the original attention scores of all key vascular regions to obtain the final attention weight distribution of the current node to each key vascular region.

[0016] Optionally, the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function for multiple consecutive times, including:

[0017] During the iterative expansion process of the EPA algorithm, the distance between the current search polyhedron and the target point is obtained, and the number of iterations is recorded;

[0018] If the geometric distance reduction of the polyhedron to the target point is less than a predetermined threshold within a preset number of consecutive iterations, it means that the target point has not been effectively approached.

[0019] For each path segment generated by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches a preset threshold, it is considered as a multiple consecutive violation of the high-priority constraint.

[0020] Optionally, adjusting the weight of the objective function or temporarily relaxing the secondary constraint condition according to the pre-stored importance level of the blood vessel region includes:

[0021] Obtaining pre-stored importance level information of each blood vessel region;

[0022] If the path deviates from the target area of ​​high importance due to the trend towards the low-importance area, the weight of the path characteristic item that guides the path to the high-importance target area in the multi-objective optimization function is increased; if the path stagnates near the high-importance area due to strict secondary constraints, the threshold of one or more secondary constraints is temporarily relaxed according to the importance level of the area;

[0023] Apply the adjusted weights or relaxed constraints to subsequent iterations of EPA.

[0024] Optionally, restarting the expansion process from the adjacent candidate node includes:

[0025] identifying a set of leading-edge nodes of the current search polyhedron and / or backtracking nodes on a path segment that causes problems;

[0026] Screening at least one candidate restart node based on the node's path feature vector and its connectivity to the target area;

[0027] Based on the adjusted objective function weights or relaxed constraints, the success probability of restarting the search from each candidate restart node is predicted, and the candidate restart node with the highest success probability is used as the starting extension point of the new EPA algorithm;

[0028] Continue the path search from the starting extension point.

[0029] This application also proposes an artificial intelligence-based cerebrovascular path planning system, including:

[0030] The recognition unit is used to obtain 3D imaging data of cerebral blood vessels and generate an initial vascular network map; perform regional division and hierarchical analysis on the initial vascular network map to identify trunk vessels, branch vessels, and key vascular regions;

[0031] A path feature extraction unit is used to aggregate the node features of the vascular network graph using the Graph Transformer model. The model calculates the weight in the attention calculation based on the distance between the node and the key vascular area and the morphological parameters at each iteration, and outputs the node's path feature vector;

[0032] a path search unit configured to perform a path search based on the path feature vector using an extended polyhedron algorithm (EPA), wherein the EPA includes a multi-objective optimization function for path length, tortuosity, minimum vessel diameter, and safety distance; and wherein during an iterative extension process of the EPA algorithm, if the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, then adjusting the weight of the objective function or temporarily relaxing secondary constraints based on pre-stored vessel region importance levels, and restarting the extension process from a neighboring candidate node.

[0033] The path output unit is used to calculate and output a cerebral vascular navigation path that meets clinical requirements.

[0034] Optionally, the model calculates the weight in the attention calculation based on the distance between the node and the key blood vessel area and the morphological parameters at each iteration, and outputs the path feature vector of the node, including:

[0035] Obtain node information of the vascular network map and identified key vascular area information;

[0036] For each node in the graph, in each iteration, the distance between the current node and each key vascular region is calculated, and the morphological parameters of the current node are extracted. Based on the distance and morphological parameters, the attention weight of the current node relative to each key vascular region is calculated using a preset attention mechanism function. The calculated attention weight is used to aggregate the features of neighboring nodes and update the features of the current node.

[0037] Through the iterative processing of multi-layer Graph Transformer until the preset maximum number of iterations is reached, the path feature vector from each node to the key vascular area is output.

[0038] Optionally, the calculation of the attention weight of the current node relative to each key blood vessel region by a preset attention mechanism function based on the distance and morphological parameters includes:

[0039] The distance and morphological parameters are digitized and normalized. Through the preset attention mechanism function, the inverse of the distance and the score of the morphological parameters are combined to calculate a raw attention score for each key vascular area for the current node.

[0040] The Softmax function is applied to the original attention scores of all key vascular regions to obtain the final attention weight distribution of the current node to each key vascular region.

[0041] Optionally, the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function for multiple consecutive times, including:

[0042] During the iterative expansion process of the EPA algorithm, the distance between the current search polyhedron and the target point is obtained, and the number of iterations is recorded;

[0043] If the geometric distance reduction of the polyhedron to the target point is less than a predetermined threshold within a preset number of consecutive iterations, it means that the target point has not been effectively approached.

[0044] For each path segment generated by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches a preset threshold, it is considered as a multiple consecutive violation of the high-priority constraint.

[0045] Optionally, adjusting the weight of the objective function or temporarily relaxing the secondary constraint condition according to the pre-stored importance level of the blood vessel region includes:

[0046] Obtaining pre-stored importance level information of each blood vessel region;

[0047] If the path deviates from the target area of ​​high importance due to the trend towards the low-importance area, the weight of the path characteristic item that guides the path to the high-importance target area in the multi-objective optimization function is increased; if the path stagnates near the high-importance area due to strict secondary constraints, the threshold of one or more secondary constraints is temporarily relaxed according to the importance level of the area;

[0048] Apply the adjusted weights or relaxed constraints to subsequent iterations of EPA.

[0049] Optionally, restarting the expansion process from the adjacent candidate node includes:

[0050] identifying a set of leading-edge nodes of the current search polyhedron and / or backtracking nodes on a path segment that causes problems;

[0051] Screening at least one candidate restart node based on the node's path feature vector and its connectivity to the target area;

[0052] Based on the adjusted objective function weights or relaxed constraints, the success probability of restarting the search from each candidate restart node is predicted, and the candidate restart node with the highest success probability is used as the starting extension point of the new EPA algorithm;

[0053] Continue the path search from the starting extension point.

[0054] This application combines the distance between the node and the key vascular area and the vascular morphological parameters to dynamically calculate the attention weight. It can learn and abstract the complex path features that guide blood flow to the target key vascular area from the 3D cerebral vascular imaging data in a deeper and more accurate manner, thereby generating a high-quality node feature vector that contains rich clinical prior knowledge and path preferences. In addition, during the iterative search process, if the EPA of this application fails to effectively approach the target point within the preset number of iterations, or violates any high-priority constraint in the preset multi-objective optimization function for multiple consecutive times, it can adjust the weight distribution of each item in the target optimization function according to the pre-stored vascular area importance level, or temporarily relax the secondary constraints, and select the optimal node from the neighboring candidate nodes at the current search front to restart the expansion process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of Example 1;

[0056] Figure 2 Schematic diagram of the initial vascular network map;

[0057] Figure 3 A schematic diagram of the obstacles encountered in the search;

[0058] Figure 4 A diagram showing the search path from a new node. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0060] The terms "first", "second" and corresponding terminology numbers in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances. This is merely a way of distinguishing when describing objects with the same properties in the embodiments of the present application. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, so that a process, method, system, product or apparatus that includes a series of units is not necessarily limited to those units, but may include other units that are not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0061] In addition, in the description of this application, unless otherwise specified, "plurality" means two or more. The term "and / or" or the character " / " in this application is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B, or A / B, can mean: A exists alone, A and B exist at the same time, or B exists alone.

[0062] In a specific embodiment, a cerebrovascular path planning method based on artificial intelligence is provided. Figure 1 Shown, including:

[0063] Step 1: Acquire 3D cerebral vascular imaging data and generate an initial vascular network map; perform regional division and hierarchical analysis on the initial vascular network map to identify trunks, branches, and key vascular regions;

[0064] Acquire three-dimensional imaging data of the patient's cerebral vasculature, such as a high-resolution three-dimensional dataset obtained through medical imaging techniques such as computed tomography angiography (CTA), magnetic resonance angiography (MRA), or digital subtraction angiography (DSA). Process the acquired three-dimensional imaging data to generate an initial vascular network atlas. The vascular network atlas shown is a structured representation of the complex cerebral vascular system, a graph structure consisting of nodes and edges, where nodes are vascular bifurcation points or key points on vascular segments, and edges are vascular segments connecting these nodes, such as Figure 2 The process of generating this atlas includes image segmentation algorithms to extract vascular structures, centerline extraction to determine the course of blood vessels, and topological analysis to establish the connection relationship between nodes. This converts continuous image information into discrete graph data.

[0065] The generated initial vascular network map is subjected to regional division and hierarchical analysis. The regional division is to divide the entire vascular network into several areas with specific significance based on anatomical knowledge or the geometric characteristics of the blood vessels, such as the anterior cerebral artery blood supply area, the middle cerebral artery blood supply area, etc. The hierarchical analysis is to distinguish the primary and secondary relationships of blood vessels, such as identifying the main arterial trunks, primary branches, secondary branches, etc.

[0066] Identify main vessels, major branch vessels, and key vascular areas that are crucial for subsequent pathway planning. These key vascular areas include target lesion locations for interventional treatment, important anatomical landmarks, or risk areas that require special attention or avoidance.

[0067] In one embodiment, feature vectors of nodes in the atlas are generated based on region division and hierarchy, such as trunk, branch vessels, and / or key vessel regions.

[0068] Step 2: Graph Transformer model is used to aggregate the node features of the vascular network graph. In each iteration, the model calculates the weight in the attention calculation based on the distance between the node and the key vascular area and the morphological parameters, and outputs the node path feature vector;

[0069] After obtaining a structured vascular network map and key region information, the Graph Transformer model performs deep learning and aggregation on the features of each node in the map. During the iterative process, the Graph Transformer model integrates the distance between each node and each identified key vascular region, such as the shortest path length or number of hops in the map, as well as the morphological parameters of the node, such as vessel diameter, curvature, and branching angle.

[0070] Based on distance and morphological parameters, the attention mechanism is used to calculate the contribution weights of different information sources to the current node feature update. Specifically, the attention weights tend to give higher attention to neighboring nodes or paths that are topologically closer to the key area and morphologically more suitable for the passage of instruments. Through the iterative aggregation and update of multi-layer Graph Transformer, the initial simple features of each node will gradually evolve into a high-dimensional, information-intensive path feature vector. The output is the path feature vector of each node in the vascular network relative to one or more key vascular areas. The path feature vector encodes comprehensive information about the potential path starting from the node and flowing to a specific key area.

[0071] Step 3: Based on the path feature vector, an extended polyhedron algorithm (EPA) is used to perform a path search. The EPA includes a multi-objective optimization function for path length, tortuosity, minimum vessel diameter, and safety distance. During the iterative expansion process of the EPA algorithm, if the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or repeatedly violates any high-priority constraint in the multi-objective optimization function, the objective function weight is adjusted or secondary constraints are temporarily relaxed based on pre-stored vessel region importance levels, and the expansion process is restarted from adjacent candidate nodes.

[0072] The path search uses the extended polyhedron algorithm (EPA), which can use the path feature vector learned by the Graph Transformer model to guide the search direction. Specifically, the path feature vector contains the relevant characteristics of the path between nodes in the vascular network. The EPA algorithm uses these feature vectors to effectively search and plan a cerebrovascular navigation path that meets clinical requirements. For example, the path feature vector contains information such as target accessibility, vessel diameter, and straightness. EPA obtains the characteristics of the node from the path feature vector, obtains the target accessibility score and safety score through MLP, etc., and then uses the extended polyhedron algorithm (EPA) to search for the path. During the search process, EPA will simultaneously optimize a multi-objective optimization function. The multi-objective optimization function contains multiple key indicators for evaluating the quality of the path, such as the total length of the path, the overall curvature of the path, the minimum lumen diameter of the blood vessels on the path, and the safe distance between the path and certain important anatomical structures or risk areas.

[0073] During the iterative expansion of the path search, the search status is continuously monitored. The current searched polyhedron fails to effectively approach the target lesion point after reaching the preset upper limit of the number of iterations; or the path segments generated by multiple attempts during the search process violate a high-priority constraint preset in the multi-objective optimization function, for example, Figure 3 As shown, if an obstacle is encountered, or the minimum diameter of the path segment is less than the safety threshold, or the curvature is too large, the path search is in trouble or heading in an infeasible direction. At this time, obtain the pre-stored information about the importance levels of different vascular regions. For example, some main blood supply arteries are more critical than their distal small branches. According to the current problem encountered and the importance of the vascular region, adjust the weights of each indicator in the multi-objective optimization function. For example, near the critical area, increase the weight of the safety of the vascular diameter, or slightly relax the requirements for the path length in the non-critical area, or temporarily relax some minor constraints while ensuring the core safety, such as not damaging the vascular wall. After the adjustment is completed, the EPA algorithm will select the most promising point from a new set of candidate nodes adjacent to the current search front to restart its expansion process, thereby trying to jump out of the local dilemma and find a new feasible path. Figure 4A feasible path from the new one is shown.

[0074] Step 4: Calculate and output a cerebral vascular navigation path that meets clinical requirements.

[0075] After the EPA search, one or more candidate paths from the starting point to the target point are generated. These candidate paths are then evaluated and screened, and a final cerebrovascular navigation path that meets clinical application requirements is calculated. The output navigation path is not just a collection of spatial points; it may also contain rich information such as vessel diameter, curvature, and relationships with important structures at each point along the path. The path is displayed in a 3D visualization environment for physician confirmation.

[0076] In an optional embodiment, the model calculates the weight in the attention calculation based on the distance between the node and the key blood vessel area and the morphological parameters at each iteration, and outputs the path feature vector of the node, including:

[0077] Obtain node information of the vascular network map and identified key vascular area information;

[0078] For each node in the graph, in each iteration, the distance between the current node and each key vascular region is calculated, and the morphological parameters of the current node are extracted. Based on the distance and morphological parameters, the attention weight of the current node relative to each key vascular region is calculated using a preset attention mechanism function. The calculated attention weight is used to aggregate the features of neighboring nodes and update the features of the current node.

[0079] Through the iterative processing of multi-layer Graph Transformer until the preset maximum number of iterations is reached, the path feature vector from each node to the key vascular area is output.

[0080] The distance between the current node and each identified critical vascular region is calculated. In one embodiment, the distance is the distance on the vascular network graph, for example, the number of edges included in the graph-theoretic shortest path from the current node to the nearest node in the critical vascular region, or a weighted shortest path length. Local morphological parameters associated with the current node are extracted, including but not limited to the average diameter of the vascular segment, the local maximum curvature, and hemodynamic parameters.

[0081] Calculate the attention weight of the current node for each key vascular region. The attention weight is determined by a preset attention mechanism function based on the distance calculated above and the extracted morphological parameters. Preferably, the distance is taken as its inverse or passes through a decay function so that the closer the distance, the greater the contribution value. The morphological parameters are scored according to clinical significance, for example, the larger the diameter, the better, and the smaller the curvature, the better. In an optional embodiment, the attention weight of the current node relative to each key vascular region is calculated by a preset attention mechanism function based on the distance and morphological parameters, including:

[0082] The distance and morphological parameters are digitized and normalized. Through the preset attention mechanism function, the inverse of the distance and the score of the morphological parameters are combined to calculate a raw attention score for each key vascular area for the current node.

[0083] Apply the Softmax function to the original attention scores of all key vascular regions to obtain the final attention weight distribution of the current node for each key vascular region. In one embodiment, the following formula is used for update: Among them, h(N i ) is the current node N i The feature vector at the beginning of this iteration, σ is the nonlinear activation function, W self is a learnable weight matrix, Neighbors(N i ) is the set of all neighbor nodes of the current node, N j is the current node N i A neighbor node of h(N j ) is the neighbor node N j In the feature vector of this iteration, M is the total number of key vascular regions, KR k is the kth key vascular region, α(N i ,KR k ) is the current node N i Relative to the key vascular region KR k The attention weight, W k,neigh is a learnable weight matrix used to learn from neighbor nodes N j Extraction of features from specific key vascular regions KR k Related information.

[0084] In an optional embodiment, the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function for multiple consecutive times, including:

[0085] During the iterative expansion process of the EPA algorithm, the distance between the current search polyhedron and the target point is obtained, and the number of iterations is recorded;

[0086] If the geometric distance reduction of the polyhedron to the target point is less than a predetermined threshold within a preset number of consecutive iterations, it means that the target point has not been effectively approached.

[0087] For each path segment generated by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches a preset threshold, it is considered as a multiple consecutive violation of the high-priority constraint.

[0088] After each iteration ends or a new polyhedron state is expanded, the representative point of the current search polyhedron is calculated, for example, the point on the polyhedron closest to the target point, or the geometric distance between the center point of the polyhedron and the preset target point, where the geometric distance is Euclidean distance, etc.

[0089] Obtain a window of consecutive iterations, for example, 5 iterations, and a minimum distance reduction threshold. When the current iteration number of the EPA algorithm is greater than or equal to the preset number of consecutive iterations, obtain the distance between the polyhedron and the target point at the current iteration. Calculate the actual geometric distance reduction from the polyhedron to the target point within the past preset number of consecutive iterations. If the geometric distance reduction from the polyhedron to the target point is less than the predetermined threshold, it means that the target point has not been effectively approached.

[0090] A set of high-priority constraints is obtained, including but not limited to minimum vessel diameter, maximum path curvature, and safety distance; and a threshold for the number of consecutive violations is obtained. For each newly generated path segment in each expansion by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches the preset threshold, it is considered a multiple consecutive violation of the high-priority constraint.

[0091] In an optional embodiment, adjusting the weight of the objective function or temporarily relaxing the secondary constraint condition according to the pre-stored importance level of the blood vessel region includes:

[0092] Obtaining pre-stored importance level information of each blood vessel region;

[0093] If the path deviates from the target area of ​​high importance due to the trend towards the low-importance area, the weight of the path characteristic item that guides the path to the high-importance target area in the multi-objective optimization function is increased; if the path stagnates near the high-importance area due to strict secondary constraints, the threshold of one or more secondary constraints is temporarily relaxed according to the importance level of the area;

[0094] Apply the adjusted weights or relaxed constraints to subsequent iterations of EPA.

[0095] Specifically, a vascular region importance level database or mapping table is constructed and stored, and different regions or specific vascular segments in the cerebral vascular network are assigned importance levels, and the importance levels can be numerical values ​​or category labels.

[0096] The current search polyhedron is located in an area of ​​lower importance, while the target point is located in an area of ​​significantly higher importance. This indicates that the current search direction is excessively biased towards optimizing certain local metrics, such as extremely short path length, at the expense of overall guidance toward the highly important target area. Identify the terms in the multi-objective optimization function that are related to the path characteristics that guide the search toward the highly important target area and increase the weight corresponding to these path characteristics. For example, if the current search polyhedron is in an area of ​​importance level 4 and the target is in an area of ​​importance level 1, the original objective function is F = 0.5 * Length + 0.5 * Curvature, and the adjusted objective function is F' = 0.7 * Length + 0.3 * Curvature.

[0097] The current search polyhedron is close to or located in an area with a higher importance level. For example, the target lesion is in this area or its vicinity. The main reason for the stagnation is the continuous violation of certain secondary constraints. The secondary constraints are relative to the core safety constraints. In one embodiment, the secondary constraints include curvature and path length. Without these secondary constraints, it is difficult to find a path in the high-importance area. According to the importance level of the current area, the preset constraint relaxation strategy table is searched, and the secondary constraints of specific types are relaxed while ensuring the core safety. For example, the original maximum radian is 0.4rad / mm, and it is temporarily adjusted to 0.45rad / mm. The adjusted objective function weight or the relaxed constraint threshold is updated to the parameter setting of the EPA algorithm. Based on these new parameters, the EPA algorithm will continue its iterative expansion process starting from the previously selected candidate restart node.

[0098] In an optional embodiment, restarting the expansion process from the adjacent candidate node includes:

[0099] identifying a set of leading-edge nodes of the current search polyhedron and / or backtracking nodes on a path segment that causes problems;

[0100] Screening at least one candidate restart node based on the node's path feature vector and its connectivity to the target area;

[0101] Based on the adjusted objective function weights or relaxed constraints, the success probability of restarting the search from each candidate restart node is predicted, and the candidate restart node with the highest success probability is used as the starting extension point of the new EPA algorithm;

[0102] Continue the path search from the starting extension point.

[0103] Obtain all nodes in the vascular network graph that constitute or are near the current EPA algorithm search frontier, i.e., the polyhedron surface. If the stagnation or violation is directly caused by a specific segment or segments, such as a segment with excessively large curvature or a segment with excessively small diameter, backtrack to this problematic segment and use the nodes on it as candidate restart node sources.

[0104] For each node in the source node set, evaluate the connectivity to the target area, for example, determine the distance to the target area, the expected path quality, or the ability to avoid known obstacles based on the current graph state; in another embodiment, the connectivity is a score of the number of paths between the candidate node and the target area. Filter out at least one preliminary candidate restart node, for example, select the node in the path feature vector that best indicates the path to the target area and whose connectivity score is higher than a certain threshold. Predict the success probability of restarting the search from each candidate restart node, preferably through a learning model such as a classifier or regressor, which takes node features, adjusted optimization parameters, and local graph structure as input, and outputs the success probability or expected benefit of restarting from the node. Determine the highest calculated node as the starting extension point for the new EPA algorithm.

[0105] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some feature data can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0106] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0107] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0108] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Claims

1. A cerebrovascular path planning method based on artificial intelligence, characterized in that: include: Acquire 3D cerebral vascular imaging data and generate an initial vascular network map; perform regional division and hierarchical analysis on the initial vascular network map to identify trunks, branch vessels, and key vascular regions; The Graph Transformer model is used to aggregate the node features of the vascular network graph. At each iteration, the model calculates the weight in the attention calculation based on the distance between the node and the key vascular area and the morphological parameters, and outputs the node's path feature vector; Based on the path feature vector, an extended polyhedron algorithm (EPA) is used to perform path search. The EPA includes a multi-objective optimization function for path length, tortuosity, minimum vessel diameter, and safety distance. During the iterative expansion process of the EPA algorithm, if the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, the objective function weight is adjusted or secondary constraints are temporarily relaxed based on pre-stored vessel region importance levels, and the expansion process is restarted from a neighboring candidate node. Calculate and output cerebral vascular navigation paths that meet clinical requirements.

2. The method according to claim 1, characterized in that At each iteration of the model, the weights in the attention calculation are calculated based on the distance between the node and the key vascular area and the morphological parameters, and the path feature vector of the node is output, including: Obtain node information of the vascular network map and identified key vascular area information; For each node in the graph, in each iteration, the distance between the current node and each key vascular region is calculated, and the morphological parameters of the current node are extracted. Based on the distance and morphological parameters, the attention weight of the current node relative to each key vascular region is calculated using a preset attention mechanism function. The calculated attention weight is used to aggregate the features of neighboring nodes and update the features of the current node. Through the iterative processing of multi-layer Graph Transformer until the preset maximum number of iterations is reached, the path feature vector from each node to the key vascular area is output.

3. The method according to claim 2, characterized in that The method calculates the attention weight of the current node relative to each key blood vessel region based on the distance and morphological parameters through a preset attention mechanism function, including: The distance and morphological parameters are digitized and normalized. Through the preset attention mechanism function, the inverse of the distance and the score of the morphological parameters are combined to calculate a raw attention score for each key vascular area for the current node. The Softmax function is applied to the original attention scores of all key vascular regions to obtain the final attention weight distribution of the current node to each key vascular region.

4. The method according to claim 1, wherein The currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, including: During the iterative expansion process of the EPA algorithm, the distance between the current search polyhedron and the target point is obtained, and the number of iterations is recorded; If the geometric distance reduction of the polyhedron to the target point is less than a predetermined threshold within a preset number of consecutive iterations, it means that the target point has not been effectively approached. For each path segment generated by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches a preset threshold, it is considered as a multiple consecutive violation of the high-priority constraint.

5. The method according to claim 1, wherein The adjusting the weight of the objective function or temporarily relaxing the secondary constraint condition according to the pre-stored importance level of the blood vessel region includes: Obtaining pre-stored importance level information of each blood vessel region; If the path deviates from the target area of ​​high importance due to the trend towards the low-importance area, the weight of the path characteristic item that guides the path to the high-importance target area in the multi-objective optimization function is increased; if the path stagnates near the high-importance area due to strict secondary constraints, the threshold of one or more secondary constraints is temporarily relaxed according to the importance level of the area; Apply the adjusted weights or relaxed constraints to subsequent iterations of EPA.

6. The method according to claim 1, characterized in that The process of restarting the expansion from the adjacent candidate node includes: identifying a set of leading-edge nodes of the current search polyhedron and / or backtracking nodes on a path segment that causes problems; Screening at least one candidate restart node based on the node's path feature vector and its connectivity to the target area; Based on the adjusted objective function weights or relaxed constraints, the success probability of restarting the search from each candidate restart node is predicted, and the candidate restart node with the highest success probability is used as the starting extension point of the new EPA algorithm; Continue the path search from the starting extension point.

7. A cerebrovascular path planning system based on artificial intelligence, characterized in that: include: The recognition unit is used to obtain 3D imaging data of cerebral blood vessels and generate an initial vascular network map; perform regional division and hierarchical analysis on the initial vascular network map to identify trunk vessels, branch vessels, and key vascular regions; A path feature extraction unit is used to aggregate the node features of the vascular network graph using the Graph Transformer model. The model calculates the weight in the attention calculation based on the distance between the node and the key vascular area and the morphological parameters at each iteration, and outputs the node's path feature vector; a path search unit configured to perform a path search based on the path feature vector using an extended polyhedron algorithm (EPA), wherein the EPA includes a multi-objective optimization function for path length, tortuosity, minimum vessel diameter, and safety distance; and wherein during an iterative extension process of the EPA algorithm, if the currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, then adjusting the weight of the objective function or temporarily relaxing secondary constraints based on pre-stored vessel region importance levels, and restarting the extension process from a neighboring candidate node. The path output unit is used to calculate and output a cerebral vascular navigation path that meets clinical requirements.

8. The system according to claim 7, characterized in that At each iteration of the model, the weights in the attention calculation are calculated based on the distance between the node and the key vascular area and the morphological parameters, and the path feature vector of the node is output, including: Obtain node information of the vascular network map and identified key vascular area information; For each node in the graph, in each iteration, the distance between the current node and each key vascular region is calculated, and the morphological parameters of the current node are extracted. Based on the distance and morphological parameters, the attention weight of the current node relative to each key vascular region is calculated using a preset attention mechanism function. The calculated attention weight is used to aggregate the features of neighboring nodes and update the features of the current node. Through the iterative processing of multi-layer Graph Transformer until the preset maximum number of iterations is reached, the path feature vector from each node to the key vascular area is output.

9. The system according to claim 8, characterized in that The method calculates the attention weight of the current node relative to each key blood vessel region based on the distance and morphological parameters through a preset attention mechanism function, including: The distance and morphological parameters are digitized and normalized. Through the preset attention mechanism function, the inverse of the distance and the score of the morphological parameters are combined to calculate a raw attention score for each key vascular area for the current node. The Softmax function is applied to the original attention scores of all key vascular regions to obtain the final attention weight distribution of the current node to each key vascular region.

10. The system according to claim 7, wherein: The currently searched polyhedron fails to effectively approach the target point within a preset number of iterations or violates any high-priority constraint in the multi-objective optimization function multiple times in a row, including: During the iterative expansion process of the EPA algorithm, the distance between the current search polyhedron and the target point is obtained, and the number of iterations is recorded; If the geometric distance reduction of the polyhedron to the target point is less than a predetermined threshold within a preset number of consecutive iterations, it means that the target point has not been effectively approached. For each path segment generated by the EPA algorithm, if the number of consecutive violations of any high-priority constraint in the multi-objective optimization function reaches a preset threshold, it is considered as a multiple consecutive violation of the high-priority constraint.

Images