Coronary artery CTA stenosis degree intelligent assessment method based on reinforcement learning

Abstract

The invention relates to the technical field of medical image processing, and particularly discloses a coronary CTA stenosis degree intelligent evaluation method based on reinforcement learning, which comprises the following steps: firstly, carrying out preprocessing and blood vessel segmentation on a coronary CTA three-dimensional image to obtain three-dimensional segmentation data containing blood vessel probability distribution; then, constructing a blood vessel tree topological graph based on the segmentation data, and reasoning, identifying and correcting connection errors through a topological relation to obtain an optimized topological structure; iteratively optimizing the segmented data by using the optimized topological structure to obtain a three-dimensional blood vessel structure with correct topology, and calculating a connection matrix and a geometric attribute matrix of a blood vessel tree; and finally, inputting the matrix data into a stenosis evaluation model constructed based on a graph attention network and a recurrent neural network, and outputting a stenosis degree evaluation result of each blood vessel segment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an intelligent assessment method for the degree of coronary CTA stenosis based on reinforcement learning. Background Technology

[0002] In intelligent assessment methods for vascular stenosis based on coronary CTA, existing technologies typically rely on deep learning models to directly segment and analyze stenosis in 3D images. However, the quality of coronary CTA images is often limited by factors such as scanning parameters, patient movement, and physiological characteristics, especially in complex anatomical structures like coronary artery bifurcations, where poor local contrast filling and interference from calcified plaques can easily lead to blurred vessel boundaries and weakened grayscale features. Existing segmentation models, when processing such areas, often focus on local pixel accuracy while lacking an understanding of the overall vascular tree topology, easily producing segmentation results with broken or incorrectly connected vessels. This topological error is directly passed to subsequent centerline extraction and stenosis analysis steps, resulting in structural biases in the assessment results and even failing to effectively assess the true stenosis of certain vessel segments.

[0003] The technical problem solved by this invention is: at the bifurcation of the coronary artery, the contrast agent concentration in the upstream vessel is locally lower than that in the downstream branch, causing the segmentation network trained based on standard data to misjudge the low-concentration vessel segment that should be connected as background noise and truncate it, thus causing a fundamental error in the topological connection relationship of the vascular tree. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent assessment method for the degree of coronary CTA stenosis based on reinforcement learning, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: The intelligent assessment method for coronary CTA stenosis based on reinforcement learning includes the following steps: Step 1: Acquire coronary CTA 3D image data, and obtain standardized 3D image data after preprocessing; Step 2: Perform blood vessel segmentation processing on the standardized 3D image data to obtain 3D segmentation data containing the probability distribution of blood vessels; Step 3: Construct an initial vascular tree topology based on the 3D segmentation data, and identify and correct connection errors in the initial vascular tree topology through topological relationship reasoning to obtain an optimized topology structure; Step 4: Iteratively optimize the 3D segmentation data using the optimized topology graph structure to obtain a topologically correct 3D vascular structure, and calculate the connection matrix and geometric attribute matrix of the vascular tree based on the 3D vascular structure; Step 5: Input the connection matrix and geometric attribute matrix into the trained stenosis assessment model of the blood vessel segment, and output the assessment results of the stenosis degree of each blood vessel segment.

[0006] As a further aspect of the present invention: obtaining the three-dimensional segmentation data containing the probability distribution of blood vessels specifically includes: A three-dimensional convolutional network is used to construct the encoding path, and blood vessel feature maps at different resolutions are obtained through multi-scale feature extraction. Feature maps are upsampled level by level in the decoding path, and the encoded features at the corresponding scale are transmitted and fused across layers. During feature fusion, the transmitted coded features are weighted. Based on the fused multi-scale features, the three-dimensional spatial relationships are reconstructed through convolution operations, generating the probability value of each voxel belonging to a blood vessel, thus forming three-dimensional segmentation data.

[0007] As a further aspect of the present invention: the construction of the initial vascular tree topology map based on three-dimensional segmentation data specifically includes: Extract the vascular centralline skeleton from the 3D segmentation data, and use the centralline branch points and endpoints as candidate nodes; Verify the connectivity of adjacent candidate nodes, retain node pairs that satisfy the continuity of blood vessel direction, and establish edge connections; Calculate the blood vessel diameter and direction vector at each node, and integrate the geometric features with the topological connectivity; A weighted topological graph structure is constructed based on node connectivity and geometric features.

[0008] As a further aspect of the present invention: the identification and correction of connection errors in the initial vascular tree topology graph through topological relationship reasoning specifically includes: Multiple rounds of message passing are performed on the topology graph, and node representations are updated through feature interactions between adjacent nodes. Based on the updated node representation, the connection confidence between nodes is calculated, and erroneous connection edges with low confidence are identified. Based on the anatomical features of the coronary artery, virtual connection paths are inferred for the broken nodes, and supplementary connection edges are generated. The supplementary connecting edges are integrated with the original topology graph to eliminate topological contradictions and form an optimized topology graph structure with a complete anatomical structure.

[0009] As a further aspect of the present invention: the iterative optimization of the 3D segmentation data using the optimized topological graph structure specifically includes: A spatial constraint field is constructed based on the optimized topology graph structure, and the node connection relationship is transformed into a connectivity constraint condition in three-dimensional space; Establish a feature propagation path to diffuse the connection information in the topology map into three-dimensional space along the direction of the blood vessel centerline; During feature propagation, feature enhancement is performed on topologically connected regions in the segmented data, while feature suppression is performed on non-connected regions. Through multiple rounds of feature adjustment and data reconstruction, broken and incorrectly connected regions in the segmented data are gradually corrected to obtain a three-dimensional vascular structure that conforms to the anatomical structure.

[0010] As a further aspect of the present invention: the calculation of the connection matrix and geometric attribute matrix of the vascular tree based on the three-dimensional vascular structure specifically includes: Traverse all edges in the topological graph and construct a connection matrix based on the actual connection relationships between blood vessel segments; The diameter sequence of each blood vessel segment is extracted along the centerline of the blood vessel, and the average diameter and minimum diameter are calculated as geometric attributes; The curvature characteristics of the blood vessel are obtained by calculating the curvature of the centerline based on the spatial coordinates of the blood vessel segment. By integrating connection states, diameter features, and curvature features, a complete connection matrix and geometric attribute matrix representation are formed.

[0011] As a further aspect of the present invention: the construction process of the stenosis assessment model for the vascular segment specifically includes: An evaluation network is constructed based on an architecture that combines graph attention networks and recurrent neural networks. The graph attention network is used to process the connection matrix of the vascular tree and to aggregate the information of adjacent nodes by calculating the attention coefficients between nodes. Recurrent neural networks are used to process the geometric property sequences of blood vessel segments and capture the temporal features of blood vessel morphology. The outputs of the graph attention network and the recurrent neural network are fused to form a comprehensive evaluation network, which is denoted as the stenosis evaluation model of the vascular segment.

[0012] As a further aspect of the present invention: the output of the stenosis assessment results for each vascular segment specifically includes: The connection matrix and geometric attribute matrix are input into the comprehensive evaluation network, and the feature representation of each blood vessel segment is calculated through the graph attention layer. Analyzing vascular segment feature sequences using recurrent neural network layers; The stenosis probability distribution of each vascular segment is calculated using a fully connected layer; The narrowing level is determined based on the probability distribution, and an evaluation report is generated based on the confidence score of the network output.

[0013] As a further aspect of the present invention: the generation of the evaluation report specifically includes: Generate a stenosis level classification for each vessel segment, including five levels: normal, mild stenosis, moderate stenosis, severe stenosis, and complete occlusion; Output the specific location information of each stenotic lesion in the vascular tree, including the name of the vessel to which it belongs and the distance from the origin of the vessel; Provide a confidence score for each narrowing assessment result.

[0014] The beneficial effects of this invention are: (1) By introducing topological relationship reasoning and iterative optimization mechanism, it can actively identify and correct vascular connection errors caused by image quality problems. In particular, it can effectively restore the vascular connection that was broken at the bifurcation due to uneven contrast agent, ensuring the correctness and continuity of the three-dimensional vascular model in anatomical structure, and providing a reliable structural basis for subsequent analysis.

[0015] (2) Based on the connection matrix and geometric attribute matrix extracted from the modified topology, and combined with the evaluation model that integrates graph attention and recurrent neural network, the global topological information and local sequence morphological features of the vascular tree can be utilized more fully. Thus, even when the initial segmentation is imperfect or there is local image interference, the stenosis level and location information of each vascular segment can still be output stably and accurately, reducing the risk of evaluation failure or misjudgment caused by errors in the preceding steps. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0019] Please see Figure 1 As shown, this invention is an intelligent assessment method for the degree of coronary CTA stenosis based on reinforcement learning, comprising the following steps: Step 1: Acquire coronary CTA 3D image data, and obtain standardized 3D image data after preprocessing; Step 2: Perform blood vessel segmentation processing on the standardized 3D image data to obtain 3D segmentation data containing the probability distribution of blood vessels; Step 3: Construct an initial vascular tree topology based on the 3D segmentation data, and identify and correct connection errors in the initial vascular tree topology through topological relationship reasoning to obtain an optimized topology structure; Step 4: Iteratively optimize the 3D segmentation data using the optimized topology graph structure to obtain a topologically correct 3D vascular structure, and calculate the connection matrix and geometric attribute matrix of the vascular tree based on the 3D vascular structure; Step 5: Input the connection matrix and geometric attribute matrix into the trained stenosis assessment model of the blood vessel segment, and output the assessment results of the stenosis degree of each blood vessel segment.

[0020] In step one, coronary CTA three-dimensional image data is acquired, and after preprocessing, standardized three-dimensional image data is obtained, specifically including: Raw coronary CTA sequence data conforming to DICOM standards were acquired through a medical imaging archiving and communication system. This data originated from routinely used clinical computed tomography (CT) scanners, employing ECG-gated technology to complete thin-slice scans of the cardiac region within a single breath-hold cycle. The raw data contains continuous cross-sectional image sequences with slice thicknesses ranging from 0.5 mm to 0.625 mm and a pixel matrix size of 512 × 512, comprehensively covering the anatomical structures from the coronary artery ostium to distal branches.

[0021] Image quality optimization was performed on the raw data. Anisotropic diffusion filtering was used for noise reduction, which effectively suppresses noise while maintaining the sharpness of blood vessel boundaries. During the filtering process, the diffusion coefficient was adaptively adjusted according to local image features, with stronger smoothing in uniform areas and weaker smoothing in blood vessel edge areas, ensuring that important anatomical details were not lost.

[0022] Image intensity normalization was performed. By analyzing the grayscale distribution of tissue within the mediastinal window and using the grayscale value of adipose tissue in the same slice as a reference, the image grayscale values ​​were mapped to a normalized numerical range. This process eliminated grayscale inconsistencies caused by differences in scanning protocols, ensuring that image data from different sources have comparable grayscale characteristics and providing a consistent image basis for subsequent vessel segmentation.

[0023] In step two, the standardized 3D image data is processed for blood vessel segmentation to obtain 3D segmentation data containing the probability distribution of blood vessels, specifically including: A 3D convolutional encoding path is constructed to extract multi-scale vascular features from the input image. The encoding path comprises four downsampling stages, each consisting of two 3D convolutional layers and one max-pooling layer. The 3D convolutional layers use 3×3×3 kernels, generating 64, 128, 256, and 512 feature maps at each stage, respectively. The max-pooling layer uses a 2×2×2 pooling window with a stride of 2, halving the spatial size of the feature maps at each stage. Through this multi-scale feature extraction process, the encoding path captures feature information ranging from local vascular texture to global vascular structure at different resolution levels.

[0024] In the decoding path, the spatial resolution of the feature map is restored through successive upsampling operations. The decoding path also contains four upsampling stages. In each stage, the feature map size is first doubled using a 3D deconvolution operation. Then, the upsampled result is concatenated across layers with the feature map of the corresponding scale in the encoding path. During the cross-layer transmission, the feature map of the encoding path is first adjusted by a 1×1×1 convolution to make its number of channels consistent with the feature map of the current stage of the decoding path. Then, the adjusted feature map and the upsampled result are concatenated along the channel dimension to complete feature fusion.

[0025] During feature fusion, the transmitted encoded features are weighted. This weighting process consists of two steps: first, a spatial weight map is calculated based on the morphological complexity of the vessel boundary, assigning higher weights to features in the vessel boundary region; second, channel attention vectors are calculated based on the information entropy of the feature channels, assigning higher weights to information-rich feature channels. Through this dual weighting mechanism, the crucial boundary information and structural features for vessel segmentation are strengthened during feature fusion.

[0026] Based on the fused multi-scale features, the distribution of blood vessels in 3D space is reconstructed through 3D convolution operations. A 1×1×1 convolutional layer is used to map the number of channels of the fused features to 2, corresponding to the background and blood vessel categories respectively. Then, the probability value of each voxel belonging to a blood vessel is calculated using the softmax function to generate the final 3D segmentation data. The value of each voxel in this segmentation data represents the probability that the location belongs to a blood vessel structure, ranging from 0 to 1, where a higher probability value indicates that the location is more likely to belong to blood vessel tissue.

[0027] In step three, an initial vascular tree topology is constructed based on the 3D segmentation data. Connection errors in the initial vascular tree topology are identified and corrected through topological relationship reasoning to obtain an optimized topology structure. Specifically, this includes: First, the vascular centerline skeleton is extracted from the 3D segmentation data. A 3D thinning algorithm is used to iteratively process the segmented data. This algorithm detects the 26-neighborhood connectivity of voxels and progressively removes boundary voxels that meet the deletion criteria until a centerline skeleton of a single voxel width is obtained. During the thinning process, branch points and endpoints are identified by detecting the number of adjacent points of a voxel: when the number of adjacent points is 1, it is marked as an endpoint; when it is greater than 2, it is marked as a branch point. Each centerline point records its spatial coordinates, the corresponding vessel radius value, and the path distance to the nearest endpoint.

[0028] Secondly, an initial vascular tree topology graph is constructed. Branch points and endpoints in the central axis skeleton are used as nodes in the graph, and central axis segments between adjacent nodes are used as edges. For each node, its spatial coordinates, vascular radius, and direction vector are calculated; for each edge, the node numbers it connects to, the central axis length, the average radius, and the curvature information are recorded. The curvature is calculated as follows: the central axis is discretized into a series of points, a circle is determined by three adjacent points, the reciprocal of the circle's radius is calculated as the curvature value at that point, and then the average curvature over the entire edge is taken.

[0029] Then, multiple rounds of message passing are performed on the topology graph to update node representations. During each round of message passing, each node collects feature information from its immediate neighbors, including spatial location, vessel radius, and orientation vector. By calculating the similarity between the features of neighboring nodes and the features of the current node, a weight coefficient is assigned to each neighboring node; the higher the similarity, the greater the weight. The weighted features of neighboring nodes are then fused with the features of the current node to update the representation of the current node. This process is repeated three times, ensuring that node features are effectively propagated throughout the topology graph.

[0030] Based on the updated node representations, the connection confidence between nodes is calculated. For each existing edge in the topology graph, the distance between the two connected nodes in the feature space is calculated, considering both the angle between the direction vectors of the two nodes and the continuity of the vessel radius. Specifically, the connection confidence calculation comprehensively considers three factors: the cosine similarity of the feature vectors of the two nodes, the dot product of the direction vectors of the two nodes, and the absolute value of the difference between the radii of the two nodes. When the confidence is below 0.5, the edge is identified as a potentially faulty connection.

[0031] Based on the anatomical features of coronary arteries, a virtual connection path is inferred for fractured nodes. When two nodes are detected to be spatially close but not connected in the topological graph, their anatomical positional relationship is analyzed. First, it is determined whether the two nodes belong to the continuation of the same vessel, which is determined by comparing the consistency of the direction vectors and the continuity of the radii of the two nodes. Then, according to the natural course of the vessel, a smooth Bézier curve is generated between the two nodes as a virtual connection path, which needs to satisfy the conditions of continuous curvature and smooth radius change.

[0032] The supplementary connecting edges are integrated with the original topology to eliminate topological inconsistencies. When adding new connecting edges, it is necessary to ensure that loops that do not conform to the anatomical structure are not formed. By detecting loops in the graph, connecting edges that would form abnormal loops are removed. At the same time, for dangling edges in the topology graph, i.e., edges that are connected at one end and isolated at the other end, their spatial location and orientation are used to determine whether they should be retained or removed. The final optimized topology graph structure ensures that the connectivity of the vascular tree conforms to the anatomical characteristics of the coronary arteries.

[0033] A spatial constraint field is constructed based on the optimized topology graph structure. The node connections in the topology graph are mapped to three-dimensional space, and a connectivity constraint value is assigned to each voxel location. Specifically, voxels located on the path connecting two adjacent nodes are assigned a higher connectivity constraint value, while voxels far from the path are assigned a lower connectivity constraint value. The calculation of the connectivity constraint value considers the distance of the voxel to the nearest centerline and the connection strength of that centerline in the topology graph.

[0034] A feature propagation path is established to diffuse the connection information in the topology graph into three-dimensional space along the vessel centerline. Starting from a node in the topology graph, the connection relationship features of the node are propagated to the surrounding space along the vessel centerline. During propagation, the feature intensity decreases with increasing distance, and the attenuation rate is proportional to the local vessel radius. At vessel bifurcation, the features continue to propagate along the directions of each branch, forming a complete feature propagation network.

[0035] During feature propagation, feature enhancement is performed on topologically connected regions in the segmented data, while feature suppression is performed on disconnected regions. For regions with high connectivity constraints, the probability value of that region in the segmented data is increased to strengthen the vascular feature; for regions with low connectivity constraints, their probability value is decreased accordingly. The intensity of feature enhancement is proportional to the connectivity constraint value, while also considering the probability values ​​in the original segmented data to avoid over-adjusting high-probability regions.

[0036] Through multiple rounds of feature adjustment and data reconstruction, broken and erroneously connected regions in the segmented data are gradually corrected. Each round of adjustment includes three steps: first, updating the spatial constraint field based on the current topology; second, adjusting the feature distribution of the segmented data according to the spatial constraint field; and finally, recalculating the vessel probability values ​​based on the adjusted features. This process is repeated until the change in the segmented data is less than a preset threshold or the maximum number of iterations is reached. The final obtained 3D vascular structure maintains the original details while ensuring the correct topological connections of the vascular tree.

[0037] In step four, the optimized topological graph structure is used to iteratively optimize the 3D segmentation data to obtain a topologically correct 3D vascular structure. Based on the 3D vascular structure, the connection matrix and geometric attribute matrix of the vascular tree are calculated, specifically including: The process of constructing a spatial constraint field based on the optimized topology graph structure first requires mapping the connectivity relationships between nodes into three-dimensional space. For each node in the topology graph, its spatial coordinates and connectivity relationships are determined, and then the influence range of that node on surrounding voxels is calculated. The size of the influence range is proportional to the radius of the blood vessel at that node, typically set to three times the radius of the blood vessel at that node. Within the influence range, the connectivity constraint value of a voxel is calculated using a Gaussian function based on its distance to the node; voxels that are closer to the node receive higher constraint values. For blood vessel segments connecting two nodes, voxels along their path receive higher constraint values, which is achieved through linear interpolation along the centerline path.

[0038] When establishing the feature propagation path, feature information is propagated level by level from the root node of the topology graph along the connection direction of the blood vessels. A breadth-first strategy is used for feature propagation, processing directly connected nodes first, and then proceeding to more distant nodes. During propagation, each node receives feature information from upstream nodes, combines it with local features for calculation, and then passes the calculation result to downstream nodes. The attenuation coefficient of feature information is inversely proportional to the propagation distance, while also considering the curvature of the blood vessel path; regions with higher curvature experience faster feature attenuation. This propagation method ensures that the topology connection information effectively covers the entire blood vessel tree structure.

[0039] The specific operations for feature enhancement and suppression during feature propagation include: for voxels in topologically connected regions, increasing their probability value in the segmented data proportionally to the magnitude of their connectivity constraint value, with the increase being the product of the original probability value and the constraint value. For voxels in non-connected regions, decreasing their probability value in inverse square of the distance to the nearest vascular node. This adjustment needs to ensure the continuity of the overall vascular structure while eliminating artifacts caused by topological errors. A threshold for probability value changes is set during the adjustment process to ensure that the magnitude of a single adjustment does not exceed 0.2, avoiding structural distortion caused by over-adjustment.

[0040] The segmentation data is optimized through multiple rounds of feature adjustment and data reconstruction. Each round of adjustment includes three steps: first, the connectivity constraint values ​​of all voxels are updated based on the current topology graph; second, the probability distribution of the segmented data is adjusted according to the constraint values; and finally, a three-dimensional Gaussian filter is used to smooth the adjusted data. After each round of adjustment, the connectivity index of the vascular structure is recalculated, and the iteration stops when the connectivity change is less than 5% after two consecutive rounds of adjustment. The entire iteration process is limited to a maximum of 10 rounds to ensure stable optimization results are obtained within a reasonable timeframe.

[0041] After obtaining the topologically correct 3D vascular structure, the connection matrix of the vascular tree is calculated. First, a unique index number is assigned to each node in the topology graph. Then, an n×n matrix is ​​constructed, where n is the total number of nodes. For each element a[i][j] in the matrix, where i represents the starting node in the connection relationship in the topology graph and j represents the target node, the element is set to 1 if the corresponding two nodes are directly connected in the topology graph, and 0 otherwise. For vascular segments with branches, the connection relationships of the branch nodes need to be specifically recorded to ensure that the connection matrix accurately reflects the topology of the vascular tree. The construction of the connection matrix also needs to consider the directionality of the vascular segments; directed connections are distinguished using specific markers.

[0042] When extracting the diameter sequence of a vessel segment along its centerline, samples are taken at fixed intervals along each segment, with the sampling interval set to half the average diameter of that segment. The diameter of the vessel at each sampling point is calculated by drawing a plane perpendicular to the centerline through that point, finding the intersection points of the plane with the segmented data boundary, and calculating the distance between the two intersection points as the diameter value at that point. After obtaining the diameter values ​​for all sampling points, the arithmetic mean of these diameter values ​​is calculated as the average diameter of the vessel segment, and the minimum diameter value is recorded as the minimum diameter.

[0043] When calculating the curvature characteristics of a vascular segment, the vascular centerline is first parametrically represented by discretizing it into a series of continuous points. For each point on the centerline, a local arc is fitted using five points: two points before and two points after it. The radius of this arc is calculated using the spatial coordinates of these five points, and the curvature value is the reciprocal of the arc radius. For a vascular segment, the average curvature value of all points is calculated as the average curvature of the segment, and the maximum curvature value is recorded as the curvature extremum of the segment. During the calculation, asymmetric neighborhood points are used for calculations at the endpoints of the centerline.

[0044] When integrating connectivity, diameter, and curvature features to form a complete matrix representation, a connectivity matrix is ​​first constructed. This symmetric matrix records the connection relationships between all nodes. Then, a geometric attribute matrix is ​​constructed, containing detailed geometric features of each vessel segment. Each row corresponds to a vessel segment and includes the following: start node number, end node number, vessel segment length, average diameter, minimum diameter, average curvature, and maximum curvature. The two matrices are linked through node numbers, jointly providing a complete description of the vessel tree's topology and geometric features. All values ​​are retained to three significant figures to ensure data accuracy and consistency.

[0045] In step five, the connection matrix and geometric attribute matrix are input into the trained stenosis assessment model for each vascular segment, and the stenosis assessment results for each segment are output, specifically including: The construction of a vascular stenosis assessment model first requires establishing an architecture based on a combination of graph attention networks and recurrent neural networks. The graph attention network is used to process the connection matrix of the vascular tree. Its implementation includes: for each node in the topology graph, calculating its attention coefficient with all neighboring nodes. The calculation of the attention coefficient is based on the similarity of node features. First, the features of neighboring nodes are concatenated with the features of the current node. Then, an attention score is calculated using a trainable weight vector. Finally, the score is normalized using a softmax function to obtain a standardized attention coefficient. The normalized attention coefficient is used as a weight to perform a weighted summation of the features of neighboring nodes, thereby updating the feature representation of the current node. This process is repeated multiple times, allowing the node features to fully aggregate neighborhood information.

[0046] The recurrent neural network (RNN) portion is used to process the geometric attribute sequences of vascular segments. Its implementation involves unfolding the vascular tree into sequential data according to anatomical structure. At each time step, the input is the geometric attribute of a vascular segment, including features such as diameter and curvature. The RNN employs a gated recurrent unit (ROU) structure, controlling the flow of information through update and reset gates. The update gate determines how much information from previous states is retained, while the reset gate determines how the new input is combined with the previous state. The hidden state at each time step is calculated based on the current input and the hidden state from the previous time step, thus capturing the continuous changes in vascular morphology along the blood flow direction. The network has two layers, and the hidden state dimension is 128.

[0047] When fusing features from the outputs of graph attention networks and recurrent neural networks, the feature vectors from both networks are first concatenated to form a composite feature vector. Then, a non-linear transformation is applied to the concatenated features through two fully connected layers: the first layer uses 128 neurons, and the second layer uses 64 neurons, both employing the ReLU activation function. Finally, an output layer maps the feature dimension to the number of classes with varying degrees of narrowness. During training, a cross-entropy loss function is used, and the Adam optimizer is employed for parameter updates. The initial learning rate is set to 0.001, and it decays to 0.5 times its original value every 50 training epochs.

[0048] When outputting the stenosis assessment results for each vessel segment, the connection matrix and geometric attribute matrix are first input into the comprehensive evaluation network. The feature representation of each vessel segment is calculated through a graph attention layer. This process involves four independent attention heads, each using different weight parameters to calculate attention coefficients. Finally, the outputs of the four heads are concatenated to obtain the final feature representation. The output dimension of the attention layer is 256, corresponding to 64-dimensional feature vectors for each of the four attention heads.

[0049] When analyzing vascular segment feature sequences using recurrent neural network (RNN) layers, the vascular segments are sorted according to the blood flow direction and input into the RNN sequentially. For each vascular segment, the RNN outputs a 128-dimensional feature vector, which contains morphological feature information of the current vascular segment and its upstream segments. During sequence processing, a bidirectional RNN is used to extract features from both the upstream and downstream blood flow directions. Finally, the outputs from the two directions are concatenated to form a 256-dimensional comprehensive feature representation.

[0050] When calculating the stenosis probability distribution for each vascular segment using fully connected layers, a three-layer fully connected network is employed. The first layer maps the 256-dimensional input features to 128 dimensions using the ReLU activation function; the second layer maps the 128-dimensional features to 64 dimensions, also using the ReLU activation function; the third layer maps the 64-dimensional features to five output nodes, corresponding to the five categories of normal, mild stenosis, moderate stenosis, severe stenosis, and complete occlusion. Finally, the output is converted into a probability distribution using the softmax function, with the sum of the probabilities of all categories being 1.

[0051] When determining the stenosis level based on probability distribution, the category with the highest probability value is selected as the stenosis level for that vessel segment. The probability value of this category is also recorded as a confidence score. Assessment results with a confidence score below 0.8 are marked as requiring manual review in the final report. The criteria for determining the stenosis level are as follows: normal corresponds to 0%–25% stenosis, mild stenosis to 25%–50%, moderate stenosis to 50%–70%, severe stenosis to 70%–99%, and complete occlusion to 100% stenosis.

[0052] When generating the assessment report, the stenosis level classification is first generated for each vessel segment. The classification is based on the ratio of the minimum diameter to the reference diameter of the vessel segment, where the reference diameter is the average diameter of the normal proximal portion of the segment. The formula for calculating the stenosis percentage is: (1 - minimum diameter / reference diameter) × 100%. Based on the calculated stenosis percentage, the vessel is classified into the corresponding stenosis level according to a preset threshold range.

[0053] When outputting the specific location information of each stenotic lesion in the vascular tree, the anatomical location of the lesion segment in the vascular tree is first determined. The path from the coronary artery ostium to the lesion segment is recorded by traversing the vascular tree topology. Location information includes the name of the vessel (e.g., left anterior descending artery, left circumflex artery, right coronary artery, etc.) and the path distance from the vessel's origin. The path distance is calculated by accumulating the length of each segment along the vessel centerline starting from the coronary artery ostium until the target lesion location is reached.

[0054] When providing a confidence score for each stenosis assessment result, the confidence score is derived from the probability distribution output by the assessment model. For the assessment result of each vessel segment, the confidence score is the probability value corresponding to the stenosis level. When generating the final assessment report, the assessment results of all vessel segments are sorted according to anatomical location, listed sequentially from proximal to distal. For each vessel segment, the report includes: vessel segment number, anatomical location description, stenosis level, stenosis percentage, confidence score, and a mark requiring manual review. The report format uses a hierarchical display, with severe and above stenosis highlighted to facilitate clinicians' rapid location of key lesions. The report also provides an overall stenosis distribution across the vascular tree, including the number and location distribution statistics of stenotic segments.

[0055] The working principle of this invention is as follows: First, the coronary CTA 3D image data is preprocessed to obtain standardized image data. Then, a 3D convolutional network is used for vessel segmentation to obtain 3D segmentation data containing probability distributions. Next, a vascular tree topology graph is constructed based on the segmentation data. Connection errors are identified and corrected through topological relationship reasoning to obtain an optimized topology structure. Then, the segmentation data is iteratively optimized using the optimized topology structure to obtain a topologically correct 3D vascular structure, and the connection matrix and geometric attribute matrix of the vascular tree are calculated. Finally, the matrix data is input into a stenosis assessment model built based on graph attention networks and recurrent neural networks, outputting an assessment report containing stenosis level, location information, and confidence score. This method, through the combination of topology correction and reinforcement learning techniques, achieves accurate and automated assessment of the degree of coronary artery stenosis.

[0056] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent assessment of coronary CTA stenosis based on reinforcement learning, characterized in that, Includes the following steps: Step 1: Acquire coronary CTA 3D image data, and obtain standardized 3D image data after preprocessing; Step 2: Perform blood vessel segmentation processing on the standardized 3D image data to obtain 3D segmentation data containing the probability distribution of blood vessels; Step 3: Construct an initial vascular tree topology based on the 3D segmentation data, and identify and correct connection errors in the initial vascular tree topology through topological relationship reasoning to obtain an optimized topology structure; Step 4: Iteratively optimize the 3D segmentation data using the optimized topology graph structure to obtain a topologically correct 3D vascular structure, and calculate the connection matrix and geometric attribute matrix of the vascular tree based on the 3D vascular structure; Step 5: Input the connection matrix and geometric attribute matrix into the trained stenosis assessment model of the blood vessel segment, and output the assessment results of the stenosis degree of each blood vessel segment.

2. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, Obtaining the three-dimensional segmentation data containing the probability distribution of blood vessels specifically includes: A three-dimensional convolutional network is used to construct the encoding path, and blood vessel feature maps at different resolutions are obtained through multi-scale feature extraction. Feature maps are upsampled level by level in the decoding path, and the encoded features at the corresponding scale are transmitted and fused across layers. During feature fusion, the transmitted coded features are weighted. Based on the fused multi-scale features, the three-dimensional spatial relationships are reconstructed through convolution operations, generating the probability value of each voxel belonging to a blood vessel, thus forming three-dimensional segmentation data.

3. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The construction of the initial vascular tree topology map based on 3D segmentation data specifically includes: Extract the vascular centralline skeleton from the 3D segmentation data, and use the centralline branch points and endpoints as candidate nodes; Verify the connectivity of adjacent candidate nodes, retain node pairs that satisfy the continuity of blood vessel direction, and establish edge connections; Calculate the blood vessel diameter and direction vector at each node, and integrate the geometric features with the topological connectivity; A weighted topological graph structure is constructed based on node connectivity and geometric features.

4. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The process of identifying and correcting connection errors in the initial vascular tree topology graph through topological relationship reasoning specifically includes: Multiple rounds of message passing are performed on the topology graph, and node representations are updated through feature interactions between adjacent nodes. Based on the updated node representation, the connection confidence between nodes is calculated, and erroneous connection edges with low confidence are identified. Based on the anatomical features of the coronary artery, virtual connection paths are inferred for the broken nodes, and supplementary connection edges are generated. The supplementary connecting edges are integrated with the original topology graph to eliminate topological contradictions and form an optimized topology graph structure with a complete anatomical structure.

5. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The iterative optimization of the 3D segmentation data using the optimized topology graph structure specifically includes: A spatial constraint field is constructed based on the optimized topology graph structure, and the node connection relationship is transformed into a connectivity constraint condition in three-dimensional space; Establish a feature propagation path to diffuse the connection information in the topology map into three-dimensional space along the direction of the blood vessel centerline; During feature propagation, feature enhancement is performed on topologically connected regions in the segmented data, while feature suppression is performed on non-connected regions. Through multiple rounds of feature adjustment and data reconstruction, broken and incorrectly connected regions in the segmented data are gradually corrected to obtain a three-dimensional vascular structure that conforms to the anatomical structure.

6. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The connection matrix and geometric attribute matrix for calculating the vascular tree based on the three-dimensional vascular structure specifically include: Traverse all edges in the topological graph and construct a connection matrix based on the actual connection relationships between blood vessel segments; The diameter sequence of each blood vessel segment is extracted along the centerline of the blood vessel, and the average diameter and minimum diameter are calculated as geometric attributes; The curvature characteristics of the blood vessel are obtained by calculating the curvature of the centerline based on the spatial coordinates of the blood vessel segment. By integrating connection states, diameter features, and curvature features, a complete connection matrix and geometric attribute matrix representation are formed.

7. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The process of constructing the stenosis assessment model for the aforementioned vascular segment specifically includes: An evaluation network is constructed based on an architecture that combines graph attention networks and recurrent neural networks. The graph attention network is used to process the connection matrix of the vascular tree and to aggregate the information of adjacent nodes by calculating the attention coefficients between nodes. Recurrent neural networks are used to process the geometric property sequences of blood vessel segments and capture the temporal features of blood vessel morphology. The outputs of the graph attention network and the recurrent neural network are fused to form a comprehensive evaluation network, which is denoted as the stenosis evaluation model of the vascular segment.

8. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 1, characterized in that, The output of the stenosis assessment results for each vascular segment specifically includes: The connection matrix and geometric attribute matrix are input into the comprehensive evaluation network, and the feature representation of each blood vessel segment is calculated through the graph attention layer. Analyzing vascular segment feature sequences using recurrent neural network layers; The stenosis probability distribution of each vascular segment is calculated using a fully connected layer; The narrowing level is determined based on the probability distribution, and an evaluation report is generated based on the confidence score of the network output.

9. The intelligent assessment method for coronary CTA stenosis based on reinforcement learning according to claim 8, characterized in that, The generation of the evaluation report specifically includes: Generate a stenosis level classification for each vessel segment, including five levels: normal, mild stenosis, moderate stenosis, severe stenosis, and complete occlusion; Output the specific location information of each stenotic lesion in the vascular tree, including the name of the vessel to which it belongs and the distance from the origin of the vessel; Provide a confidence score for each narrowing assessment result.

Images