Intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement

Abstract

The application discloses an intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement, which comprises the following steps: constructing a static neighborhood graph based on the center points of each region after dividing the point cloud into local regions; performing geometric feature coding on the points in each local region to generate initial local features, respectively calculating the rotation invariant position coding of each local region and the rotation invariant direction coding between adjacent regions in the static neighborhood graph, and embedding the initial local features; after multi-layer aggregation and update of each initial local feature on the static neighborhood graph through a neighborhood saliency enhancement encoder, performing geometric perception decoding on the static neighborhood graph through a geometric perception decoder, and then aggregating the global context information between regions through a sequence processor to obtain decoding features; and predicting and outputting the intracranial aneurysm segmentation result based on the decoding features. The method can effectively improve the fine segmentation capability for complex blood vessel structures and take into account the calculation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric structure enhancement. Background Technology

[0002] Intracranial aneurysms (IAS) are localized dilations of cerebral arteries caused by thinning of the vessel walls. Rupture of an IAS can lead to subarachnoid hemorrhage and neurological impairment, making timely treatment crucial. However, accurate aneurysm localization is challenging and highly dependent on surgical experience. Current localization methods are primarily based on magnetic resonance angiography (MRA) and computed tomography angiography (CTA), but their accuracy is insufficient in cases involving complex vascular structures or small aneurysms.

[0003] Compared to traditional methods, point cloud methods can directly describe the geometric structure of blood vessel surfaces, offering advantages such as lower memory overhead and sensitivity to curvature and boundary information, making them particularly suitable for fine modeling of the junction between aneurysms and blood vessels. In recent years, research on automated analysis of intracranial aneurysms (IAs) has seen the emergence of segmentation methods based on 3D surface point clouds. For example, Qi et al. proposed PointNet, introducing a permutation-invariant point-level multilayer perceptron combined with symmetric pooling. Wang et al. proposed DGCNN, implementing dynamic graph construction and edge feature modeling through EdgeConv. Zhao et al. proposed Point Transformer, introducing a local-global context modeling method based on an attention mechanism. However, existing point cloud segmentation methods still have the following shortcomings when directly applied to intracranial aneurysm segmentation.

[0004] (1) Insufficient utilization of local boundary features: In intracranial aneurysm segmentation tasks, the information with real discriminative value is concentrated at the junction of the aneurysm body and neck, the high curvature change area, and the vascular attachment area. Traditional point cloud methods often treat points in the local neighborhood as approximately equal weights, or use a relatively coarse neighborhood aggregation method, which makes it difficult to highlight key points with task significance.

[0005] (2) Limited ability to model complex vascular structures: Intracranial vessels are characterized by their slenderness, tortuosity, numerous bifurcations, and complex topology, especially near aneurysms where local morphological changes are drastic. Although existing methods can extract certain local geometric features, they are still insufficient in modeling mid-range vessel orientation, structural continuity between adjacent regions, and complex geometric contexts. This can lead to missegmentation, incorrect connections, or even topological defects in models in complex vascular pathways, narrow neck regions, or regions where multiple vessels intersect.

[0006] (3) Sensitive to pose changes and registration perturbations: Existing point cloud segmentation methods rely on Cartesian coordinates, normal vectors, or conventional position encoding to represent spatial relationships. These representations change significantly with the overall rotation of the point cloud or changes in scanning pose, leading to instability in the feature space and affecting the model's generalization ability and segmentation consistency. For medical data, different patients, different acquisition poses, and slight deviations during reconstruction can amplify this problem. Encoding methods based on coordinates, normal vectors, or sinusoidal position functions are still rotationally sensitive and easily affected by registration and pose changes.

[0007] Furthermore, while pursuing improved accuracy, existing 3D medical segmentation models often incur significant parameter counts and computational overhead, hindering rapid deployment in real-world clinical environments. Therefore, further research is needed on lightweight geometric perception methods for point cloud segmentation of intracranial aneurysms. These methods aim to fully exploit the local salient features of the aneurysm neck boundary and high-curvature regions, enhance mid-range context modeling capabilities for complex vascular topologies, possess robustness to pose changes, and achieve high-precision aneurysm segmentation with lower model complexity to meet the needs of clinical auxiliary diagnosis and rapid deployment. Summary of the Invention

[0008] In view of the above, the purpose of this invention is to provide a point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement. This method improves the representation ability of key boundary regions and complex vascular geometry, enhances the contextual modeling ability of mid-range vascular continuity and multi-regional structural relationships, and reduces the impact of pose changes and registration perturbations on feature representation, thereby improving the accuracy and stability of intracranial aneurysm point cloud segmentation. Simultaneously, while ensuring segmentation performance, it also considers model computational efficiency and practical clinical deployment needs, thus improving the applicability of the method in complex medical point cloud scenarios.

[0009] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement, comprising the following steps: The input point cloud is divided into local regions to construct several local regions, and a static neighborhood graph is constructed based on the center point of each local region. Geometric feature encoding is performed on points within each local region to generate initial local features. Rotation-invariant position encoding based on the local covariance matrix is ​​calculated for each local region, and rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map is calculated and embedded into the initial local features. After the neighborhood saliency enhancement encoder performs multi-level aggregation and update of each initial local feature on the static neighborhood graph, the geometric perception decoder performs geometric perception decoding on the static neighborhood graph, and then the sequence processor aggregates the global context information between each region to obtain the decoded features. The decoded features are reconstructed to the original point cloud level, and the intracranial aneurysm segmentation result for each point is predicted and output by the segmentation head.

[0010] Preferably, the step of dividing the input point cloud into local regions to construct several local regions, and constructing a static neighborhood graph based on the center point of each local region, includes: After normalizing the input point cloud, several region centers are selected by sampling the farthest point. A preset number of neighboring points are selected around each center to form a local region. Then, a static k-nearest neighbor graph is constructed on all region center points according to Euclidean distance as a static neighborhood graph.

[0011] Preferably, the calculation of rotation-invariant position encoding of each local region based on the local covariance matrix includes: For each local region, calculate the local covariance matrix of its local point set, perform principal component analysis on the local covariance matrix to extract the three principal component directions to form a local rotation matrix, and generate the rotation-invariant position code of the local region through shared mapping using the local rotation matrix.

[0012] Preferably, the calculation of rotation-invariant direction encoding between adjacent regions in the static neighborhood graph based on a relative rotation matrix includes: For each edge in the static neighborhood graph, obtain the local rotation matrix of two adjacent local regions, multiply the transpose of one matrix by the other matrix to obtain the relative rotation matrix, and generate the rotation-invariant direction code corresponding to the edge through a shared mapping using the relative rotation matrix.

[0013] Preferably, the step of performing multi-level aggregation update of each initial local feature on the static neighborhood graph through the neighborhood saliency enhancement encoder includes: For each coding layer, for each central region and its neighboring regions, the features of the central region and the difference between the features of the central region and the features of the neighboring regions are concatenated along the channel dimension to construct a pairwise representation of the edges. Pairwise representations are input into a compressed multilayer perceptron to obtain neighborhood relationship features. Max pooling is then used to perform order-independent saliency aggregation of all neighborhood edge features with the same center to obtain local aggregation results. The local convergence results are input into an extended multilayer perceptron to generate feature update values, and then combined with residual connections, layer normalization, and a feedforward network to update the features of the current layer.

[0014] Preferably, the geometric perception decoding on the static neighborhood graph via the geometric perception decoder includes: For each decoding layer, for each central region and its neighboring regions, calculate the relative displacement vector between the features of the central region and the features of the neighboring regions; The relative displacement vector is concatenated with the features of the central region and the features of the neighboring region and then input into the message multilayer perceptron to generate a geometric perception message. At the same time, the relative displacement vector is concatenated with the features of the central region and then input into a gated multilayer perceptron and processed by an activation function to generate gated weights. The geometric perception messages of each neighborhood are weighted by gating weights and then aggregated in an order-independent manner to obtain the implicit communication results of each central region. After layer normalization and feature expansion, the updated decoding features are obtained.

[0015] Preferably, the aggregation of global context information between regions through a sequence processor includes: A Mamba-based sequence processor is used to arrange the regional features output by the geometry-aware decoder in spatial order according to the regional center points to form a feature sequence. Global context modeling is performed along the sequence direction to output an enhanced decoded feature sequence.

[0016] Secondly, the present invention provides an intracranial aneurysm point cloud segmentation device based on neighborhood saliency and geometric enhancement, which is implemented using the above-mentioned intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement, including: The module includes a local neighborhood construction module, a geometric structure encoding module, a global saliency aggregation module, and a point cloud reconstruction and segmentation module. The local neighborhood construction module is used to divide the input point cloud into local regions to construct several local regions, and to construct a static neighborhood graph based on the center point of each local region. The geometric structure encoding module is used to encode the geometric features of points in each local region to generate initial local features. It calculates the rotation-invariant position encoding based on the local covariance matrix for each local region and the rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map, and embeds the initial local features. The global saliency aggregation module is used to perform multi-level aggregation and update of each initial local feature on the static neighborhood graph through the neighborhood saliency enhancement encoder, perform geometric perception decoding on the static neighborhood graph through the geometric perception decoder, and then aggregate the global context information between each region through the sequence processor to obtain the decoded features. The point cloud reconstruction and segmentation module is used to reconstruct the decoded features to the original point cloud level, and output the intracranial aneurysm segmentation result for each point through the segmentation head prediction.

[0017] Thirdly, an electronic device provided by an embodiment of the present invention includes a memory and one or more processors. The memory is used to store a computer program, and the processor is used to implement the above-described method for intracranial aneurysm point cloud segmentation based on neighborhood saliency and geometric enhancement when executing the computer program.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the above-described method for intracranial aneurysm point cloud segmentation based on neighborhood saliency and geometric enhancement.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) This invention constructs a pairwise representation of the edges of the central region and the neighboring region by constructing a neighborhood saliency enhancement encoder, and adopts a compression-expansion mapping and a sequence-independent saliency convergence mechanism to highlight key geometric information such as the aneurysm neck boundary, high curvature region and lesion attachment area, which solves the problem of insufficient utilization of local discriminative features in traditional methods and improves the precision of aneurysm boundary segmentation.

[0020] (2) This invention achieves selective geometric communication between neighborhoods through message generation and gating weighting mechanism based on relative displacement in the geometric perception decoder, and combines global context modeling along the regional sequence direction with Mamba-based sequence processor, which enhances the ability to model the mid-range structural continuity and cross-regional relationship in complex vascular topology, multi-branch and multi-vascular intersection scenarios, and reduces missegmentation and topological defects.

[0021] (3) By embedding rotation-invariant position encoding based on local covariance matrix and rotation-invariant orientation encoding based on relative rotation matrix into each local region, this invention effectively eliminates the influence of overall point cloud rotation, pose change and slight registration perturbation on feature expression, and significantly improves the segmentation stability and generalization ability of the model under different acquisition conditions and registration errors.

[0022] (4) While improving the segmentation accuracy and stability, this invention adopts static neighborhood graph reuse, lightweight multilayer perceptron and linear complexity Mamba sequence processing, which effectively controls the number of model parameters and computational overhead, takes into account the requirements of computational efficiency and real-time performance for clinical deployment, and improves the applicability of the method in complex medical point cloud scenarios. It is expected to provide a more reliable technical means for the automated fine segmentation of intracranial aneurysms, thereby providing support for subsequent lesion analysis, treatment evaluation and clinical auxiliary decision-making. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement provided in the embodiments of the present invention. Figure 3 This is a schematic diagram of the implicit communication module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the intracranial aneurysm point cloud segmentation device based on neighborhood saliency and geometric structure enhancement provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0026] Current point cloud-based intracranial aneurysm segmentation methods generally suffer from insufficient utilization of key boundary information, limited ability to model complex vascular structures, and low stability of segmentation results when dealing with the aneurysm-neck junction, areas with high curvature changes, and complex vascular attachment regions. Especially when dealing with slender, tortuous vessels with numerous bifurcations, complex topological relationships, and variations in pose and slight registration perturbations, existing methods often struggle to simultaneously preserve local details and represent the overall structure, easily leading to overly smoothed boundaries, local missegmentation, incorrect connections, or decreased segmentation performance in complex scenarios. These problems not only affect the accuracy and consistency of intracranial aneurysm segmentation results but also limit the application value of these methods in refined analysis and clinical decision support.

[0027] In view of this, the inventive concept of this invention is as follows: A point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement is provided. This method constructs a static neighborhood graph to stabilize local topological relationships, embeds rotation-invariant position and orientation codes based on covariance and relative rotation matrices into each local region to eliminate the influence of pose changes. Furthermore, a neighborhood saliency enhancement encoder is used to compress, converge, and expand the paired representations of center and neighborhood features, highlighting key geometric information such as the aneurysm neck boundary and high curvature areas. Selective local geometric communication is achieved through gated message aggregation in a geometry-aware decoder, and global context dependencies are modeled along the regional sequence direction using a Mamba-based sequence processor. Finally, the aneurysm segmentation result is predicted based on the decoded features, thereby improving the ability to finely segment complex vascular structures while maintaining computational efficiency.

[0028] like Figure 1 and Figure 2 As shown in the embodiment, an intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement is provided, including the following steps: S1, divide the input point cloud into local regions to construct several local regions, and construct a static neighborhood graph based on the center point of each local region.

[0029] In this embodiment, the input 3D surface point cloud is first preprocessed to obtain a normalized point set. Then, farthest point sampling (FPS) is used to select several region centers from the point cloud, and sampling is performed around each center point. Select The nth neighboring point forms the nth... This divides the original point cloud into several local regions, enabling subsequent feature modeling at the region level.

[0030] At all regional center points Above, a static k-nearest neighbor (KNN) graph is constructed based on Euclidean distance, where for any two center points... and Its distance Defined as: , in, This represents the Euclidean distance.

[0031] Based on this distance, define the center point. static neighborhood for: , Here, TopK represents selecting the k nearest neighbors. This allows the formation of a static neighborhood graph at the region center, which can be reused throughout the encoding and decoding stages without being recalculated at each layer, thus ensuring stable neighborhood relationships and reducing computational overhead.

[0032] S2, geometric feature encoding is performed on points in each local region to generate initial local features. The rotation-invariant position encoding based on the local covariance matrix of each local region and the rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map are calculated and embedded into the initial local features.

[0033] In this embodiment, for each local region, the local geometric information of its internal points is input into a lightweight multilayer perceptron for encoding. Specifically, the coordinate information of each point in the local region is input into a coordinate-based multilayer perceptron (Coord-MLP), and then the resulting features are concatenated, linearly projected, and max-pooled to obtain the initial local features (represented by a token) corresponding to that region, denoted as . .

[0034] Furthermore, to reduce the impact of input pose variations on feature representation, rotation-invariant orientation encoding (RI-OE) and rotation-invariant position encoding (RI-PE) layers are embedded to process the input. For the Let there be a region, and let its local point set be... First, calculate the local covariance matrix. : , in, Indicates a local point index. This represents the total number of local points. This represents the local mean of the region, indicated by the superscript. This indicates transpose.

[0035] For the local covariance matrix Principal component analysis was performed to obtain the local rotation matrix. : , in, It represents a three-dimensional special orthogonal group.

[0036] center point Projecting onto a locally rotated coordinate system yields : , Then obtain the location code through shared mapping. : , in, This represents the positional encoding mapping function. This indicates the center point after local rotation and alignment. Indicates all A set of 3D real vectors.

[0037] In addition to positional encoding, this invention also constructs relative direction encoding between regions. For edges in a static neighborhood graph... Constructing a relative rotation matrix using principal component analysis of two regions as local reference frames. : , Among them, superscript This indicates transpose.

[0038] Will After vectorization, the vectorized data is fed into a shared mapping to obtain the direction code. : , in, Indicates a shared mapping function. This indicates that the matrix can be flattened into a vector. Indicates all A set of 3D real vectors.

[0039] Finally, the obtained position code and direction encoding The initial local features are then incorporated into a multi-layer perceptron (MLP) to obtain the final features. : , in, Represents aggregate functions, This represents the total number of layers in a multilayer perceptron. This indicates the level index.

[0040] S3 involves updating each initial local feature in a multi-layer aggregation on a static neighborhood graph through a neighborhood saliency enhancement encoder, followed by geometric perception decoding on the static neighborhood graph through a geometric perception decoder, and then aggregating the global context information between regions through a sequence processor to obtain the decoded features.

[0041] In the embodiment, after embedding rotation-invariant position encoding and rotation-invariant direction encoding, the resulting initial local features are... The input neighborhood saliency enhancement encoder (NSE) is updated layer by layer. For the coding layer... Central region features and the difference between the characteristics of the central region and the characteristics of the neighboring regions. Pairwise representation of constructing edges (The following is omitted) : , Subsequently, the edge representations are input into a compressed multilayer perceptron (C-MLP) to obtain compact neighborhood relationship features. : , For the same center point Order-independent aggregation (ONF) is performed on all neighborhood edge features to obtain the local saliency aggregation result. : , In this implementation, ONF employs max pooling to highlight the strongest and most discriminative geometric relationships within local regions. Next, the local saliency convergence results are... The data is fed into an Extruded Multilayer Perceptron (E-MLP) and mapped back to the token dimension to obtain... The token update for the current layer is completed by combining residuals, layer normalization (LN), and a feedforward network (FFN), as shown below: , in, Indicates the first The input token sequence of the layer NSE encoder, Indicates the first The token sequence output by the layer NSE encoder. This indicates the total number of layers in the NSE.

[0042] After L layers of NSE encoding, the features output by the encoder are obtained. : , in, This indicates the total number of local areas.

[0043] S4 reconstructs the decoded features to the original point cloud level, and outputs the intracranial aneurysm segmentation result for each point through segmentation head prediction.

[0044] In this embodiment, the encoder output is... As input to the decoder, geometry-aware decoding is performed on each local region token on the static neighborhood graph. First, it will be processed by the Implicit Communication Module (ICM), the process is as follows: Figure 3 As shown, let the first... Layer decoding time center The corresponding token is ,Neighbor The corresponding token is And denote the relative displacement as Geometric sensing messages are calculated using information MLP and gated MLP respectively. and gating weights : , , The weighted neighborhood messages are aggregated in an order-independent manner to obtain the center. Implicit communication results: , Then, the token is obtained through aggregation, layer normalization (LN), and eigenvalue expansion: , , in, Indicates the first The token input to the layer ICM module, This represents the token after aggregation, LN, and eigenvalue expansion processing. This represents the token after residual join and FFM processing. This indicates that the domain-aggregated messages will be mapped back to the token dimension. This indicates that the token is processed using a feedforward network.

[0045] After the implicit communication module completes the local geometric communication between adjacent local blocks, the final tokens of each local region are stacked to obtain the feature. The data is then input into a Mamba-based sequence processor to further model mid-range dependencies along the token sequence, resulting in more globally context-aware decoding features. ,in, This indicates the number of sequence processing layers in the decoding stage.

[0046] Finally, the decoded features at the local region level, processed by ICM and Mamba, are input into the reconstruction head. First, the token features are linearly projected, and then upsampling and propagation are performed from the region center point level to the original point level to recover dense point-level feature representations. Afterward, the reconstructed point-level features are input into the segmentation head, which outputs the category prediction result for each point, resulting in segmentation maps of blood vessels and aneurysms.

[0047] In summary, the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement provided in this embodiment of the invention can effectively improve the boundary delineation ability, complex vascular structure modeling ability, and robustness to pose changes in the intracranial aneurysm point cloud segmentation task. The specific effects are as follows.

[0048] Table 1. Segmentation results of different methods on IntrA

[0049] First, regarding technical performance, experimental results on the publicly available IntrA dataset demonstrate that our method outperforms existing representative methods in the intracranial aneurysm segmentation task. As shown in Table 1, V_IoU, A_IoU, V_DSC, and A_DSC represent the vessel crossover ratio, aneurysm crossover ratio, vessel prediction and annotation overlap, and aneurysm prediction and annotation overlap, respectively. Bold values ​​indicate the best performance for each metric. Wilcoxon's test was performed on the model results; * indicates a p-value less than 0.0001 compared to PMMNet, and † indicates a p-value less than 0.0001 compared to M-PointNet. With a 1024-point setting, our method achieved vessel IoU and DSC of 95.45% and 97.62%, respectively, and aneurysm IoU and DSC of 84.35% and 91.03%, respectively. In comparison, M-PointNet achieved 83.68% and 90.05% IoU and DSC for aneurysms under the same settings, PMMNet achieved 77.48% and 85.68%, respectively, PCT achieved 78.12% and 86.77%, and PointCNN achieved only 74.11% and 81.74%, respectively. Our proposed method improved aneurysm IoU and DSC by 0.50% and 0.78% compared to M-PointNet, respectively, and the comparisons with both M-PointNet and PMMNet were statistically significant (p<0.0001). This indicates that our method can more effectively handle the fine segmentation problems of the aneurysm-neck junction, high-curvature regions, and complex vascular attachment regions.

[0050] Secondly, in terms of classification and generalization ability, this method is not only applicable to intracranial aneurysm segmentation, but also demonstrates good classification ability and cross-dataset adaptability. In the IntrA classification task, the aneurysm classification accuracy of this method reaches 95.82% with a 1024-point setting, which is higher than PMMNet's 94.32%, 3DMedPT's 93.26%, and M-PointNet's 91.86%; with a 512-point setting, the F1-score reaches 0.933, which is close to or better than most of the comparison methods. On the general point cloud dataset, this method achieves an average class IoU of 84.8% on ShapeNetPart, only 0.2% lower than KPConv, and higher than PointNet++, DGCNN, PCT, and AdaptConv; on ModelNet40, the classification accuracy reaches 93.6%, which is higher than PointCNN's 91.7%, DGCNN's 92.2%, KPConv's 92.9%, and PCT's 93.2%. This indicates that the proposed method not only improves the performance of intracranial aneurysm point cloud analysis, but also has good general three-dimensional point cloud understanding capabilities and scalability.

[0051] Furthermore, regarding the performance improvements brought by key modules, this method addresses issues such as insufficient utilization of key boundary features, insufficient cross-regional contextual communication, and pose sensitivity through three mechanisms: neighborhood saliency enhancement encoding, geometry-aware decoding, and rotation-invariant encoding. Ablation experiments show that, with a 1024-point setting, the aneurysm IoU and DSC of the standard baseline Transformer are only 72.94% and 80.88%, respectively. After introducing the neighborhood saliency enhancement encoder, these indicators improve to 82.58% and 89.97%; further addition of the geometry-aware decoder improves them to 83.64% and 90.73%; and finally, the addition of rotation-invariant orientation and position encoding further improves them to 85.62% and 91.89%. Specifically, the geometry-aware decoder improves aneurysm IoU and DSC by 1.06% and 0.76%, respectively, while rotation-invariant encoding further improves them by 1.98% and 1.16%. This demonstrates that each component of this method can address different deficiencies in existing technologies and create synergistic gains.

[0052] Furthermore, regarding model complexity and deployment efficiency, this method maintains high segmentation performance while balancing computational overhead and practical application requirements. With a 1024-point configuration, the model's parameter count is 15.94M and FLOPs is 30.25G, lower than 3D U-Net's 22.58M and 59.68G, lower than V-Net's 45.60M and 91.57G, and significantly lower than TransBTS's 333.00G FLOPs and SwinUNETR's 61.98M parameters and 394.84G FLOPs. In comparison, this method still maintains high segmentation performance for blood vessels and aneurysms on the IntrA dataset, indicating that it can effectively balance accuracy and efficiency in complex medical point cloud scenarios, which is beneficial for subsequent engineering implementation and clinical auxiliary applications.

[0053] Finally, from an application perspective, this method provides more stable and reliable technical support for the automated fine segmentation of intracranial aneurysms. Because it can more accurately characterize the aneurysm neck boundary, complex vascular structures, and geometric features under pose changes, it helps provide more refined structural information for subsequent lesion analysis, treatment evaluation, and decision support. Furthermore, the good performance of this method on general point cloud classification and segmentation tasks indicates that its design concept is not only applicable to intracranial aneurysm point cloud segmentation but can also provide a reference for other complex medical point cloud analysis tasks. The above experimental results and analysis show that this method has good overall performance in terms of technical performance, computational efficiency, and application adaptability.

[0054] Based on the same inventive concept, such as Figure 4 As shown, this embodiment of the invention also provides an intracranial aneurysm point cloud segmentation device 400 based on neighborhood saliency and geometric structure enhancement, including: a local neighborhood construction module 410, a geometric structure encoding module 420, a global saliency aggregation module 430, and a point cloud reconstruction and segmentation module 440.

[0055] The local neighborhood construction module 410 is used to divide the input point cloud into local regions to construct several local regions, and to construct a static neighborhood graph based on the center point of each local region.

[0056] The geometric structure encoding module 420 is used to generate initial local features by encoding the geometric features of points in each local region. It calculates the rotation-invariant position encoding based on the local covariance matrix for each local region and the rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map, and embeds the initial local features.

[0057] The global saliency aggregation module 430 is used to perform multi-level aggregation and update of each initial local feature on the static neighborhood graph through the neighborhood saliency enhancement encoder, perform geometric perception decoding on the static neighborhood graph through the geometric perception decoder, and then aggregate the global context information between each region through the sequence processor to obtain the decoded features.

[0058] The point cloud reconstruction and segmentation module 440 is used to reconstruct the decoded features to the original point cloud level, and output the intracranial aneurysm segmentation result for each point through the segmentation head prediction.

[0059] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described method for intracranial aneurysm point cloud segmentation based on neighborhood saliency and geometric enhancement when executing the computer program.

[0060] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the above-described method for intracranial aneurysm point cloud segmentation based on neighborhood saliency and geometric enhancement.

[0061] It should be noted that the intracranial aneurysm point cloud segmentation device, electronic device, and computer-readable storage medium provided in the above embodiments all belong to the same inventive concept as the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement. For details of its implementation process, please refer to the embodiments of the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement, which will not be repeated here.

[0062] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement, characterized in that, Includes the following steps: The input point cloud is divided into local regions to construct several local regions, and a static neighborhood graph is constructed based on the center point of each local region. Geometric feature encoding is performed on points within each local region to generate initial local features. Rotation-invariant position encoding based on the local covariance matrix is ​​calculated for each local region, and rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map is calculated and embedded into the initial local features. After the neighborhood saliency enhancement encoder performs multi-level aggregation and update of each initial local feature on the static neighborhood graph, the geometric perception decoder performs geometric perception decoding on the static neighborhood graph, and then the sequence processor aggregates the global context information between each region to obtain the decoded features. The decoded features are reconstructed to the original point cloud level, and the intracranial aneurysm segmentation result for each point is predicted and output by the segmentation head.

2. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1, characterized in that, The step of dividing the input point cloud into local regions to construct several local regions, and constructing a static neighborhood graph based on the center point of each local region, includes: After normalizing the input point cloud, several region centers are selected by sampling the farthest point. A preset number of neighboring points are selected around each center to form a local region. Then, a static k-nearest neighbor graph is constructed on all region center points according to Euclidean distance as a static neighborhood graph.

3. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1, characterized in that, The calculation of rotation-invariant position encoding for each local region based on the local covariance matrix includes: For each local region, calculate the local covariance matrix of its local point set, perform principal component analysis on the local covariance matrix to extract the three principal component directions to form a local rotation matrix, and generate the rotation-invariant position code of the local region through shared mapping using the local rotation matrix.

4. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1 or 3, characterized in that, The calculation of rotation-invariant direction encoding between adjacent regions in the static neighborhood graph based on the relative rotation matrix includes: For each edge in the static neighborhood graph, obtain the local rotation matrix of two adjacent local regions, multiply the transpose of one matrix by the other matrix to obtain the relative rotation matrix, and generate the rotation-invariant direction code corresponding to the edge through a shared mapping using the relative rotation matrix.

5. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1, characterized in that, The step of performing multi-level aggregation and updating of each initial local feature on the static neighborhood graph using a neighborhood saliency enhancement encoder includes: For each coding layer, for each central region and its neighboring regions, the features of the central region and the difference between the features of the central region and the features of the neighboring regions are concatenated along the channel dimension to construct a pairwise representation of the edges. Pairwise representations are input into a compressed multilayer perceptron to obtain neighborhood relationship features. Max pooling is then used to perform order-independent saliency aggregation of all neighborhood edge features with the same center to obtain local aggregation results. The local convergence results are input into an extended multilayer perceptron to generate feature update values, and then combined with residual connections, layer normalization, and a feedforward network to update the features of the current layer.

6. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1, characterized in that, The geometric perception decoding on the static neighborhood graph via the geometric perception decoder includes: For each decoding layer, for each central region and its neighboring regions, calculate the relative displacement vector between the features of the central region and the features of the neighboring regions; The relative displacement vector is concatenated with the features of the central region and the features of the neighboring region and then input into the message multilayer perceptron to generate a geometric perception message. At the same time, the relative displacement vector is concatenated with the features of the central region and then input into a gated multilayer perceptron and processed by an activation function to generate gated weights. The geometric perception messages of each neighborhood are weighted by gating weights and then aggregated in an order-independent manner to obtain the implicit communication results of each central region. After layer normalization and feature expansion, the updated decoding features are obtained.

7. The intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement according to claim 1, characterized in that, The aggregation of global context information between regions through a sequence processor includes: A Mamba-based sequence processor is used to arrange the regional features output by the geometry-aware decoder in spatial order according to the regional center points to form a feature sequence. Global context modeling is performed along the sequence direction to output an enhanced decoded feature sequence.

8. A point cloud segmentation device for intracranial aneurysms based on neighborhood saliency and geometric enhancement, implemented using the point cloud segmentation method for intracranial aneurysms based on neighborhood saliency and geometric enhancement as described in any one of claims 1 to 7, characterized in that, include: The module includes a local neighborhood construction module, a geometric structure encoding module, a global saliency aggregation module, and a point cloud reconstruction and segmentation module. The local neighborhood construction module is used to divide the input point cloud into local regions to construct several local regions, and to construct a static neighborhood graph based on the center point of each local region. The geometric structure encoding module is used to encode the geometric features of points in each local region to generate initial local features. It calculates the rotation-invariant position encoding based on the local covariance matrix for each local region and the rotation-invariant direction encoding based on the relative rotation matrix between adjacent regions in the static neighborhood map, and embeds the initial local features. The global saliency aggregation module is used to perform multi-level aggregation and update of each initial local feature on the static neighborhood graph through the neighborhood saliency enhancement encoder, perform geometric perception decoding on the static neighborhood graph through the geometric perception decoder, and then aggregate the global context information between each region through the sequence processor to obtain the decoded features. The point cloud reconstruction and segmentation module is used to reconstruct the decoded features to the original point cloud level, and output the intracranial aneurysm segmentation result for each point through the segmentation head prediction.

9. An electronic device comprising a memory and one or more processors, the memory for storing a computer program, characterized in that, The processor is used to implement the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric enhancement as described in any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a computer, it implements the intracranial aneurysm point cloud segmentation method based on neighborhood saliency and geometric structure enhancement as described in any one of claims 1 to 7.

Images