A two-way asymmetric dsa sequence cerebral vessel segmentation method and system

Abstract

The application discloses a kind of two-way asymmetric DSA sequence cerebral vessel segmentation method and system, the method includes: acquisition original contrast image pre-processing generates DSA sequence;Based on the preset double-branch encoding mechanism, the complementary space-time characteristics of DSA sequence are extracted;To complementary space-time characteristics, based on the preset selective space state model, time series modeling is carried out, and space-time dynamic characteristics are obtained;According to space-time dynamic characteristics and the preset two-way guide enhancement mechanism, obtain space-time modulation characteristics, based on the asymmetric cross attention mechanism of space leading, cross attention fusion is carried out to space-time modulation characteristics, and generates target space-time representation;Target space-time representation is decoded, and the cerebral vessel segmentation result of DSA sequence is obtained.It can be seen that the application can realize efficient space-time fusion, significantly improve the accuracy of vessel segmentation, computational efficiency and robustness.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, and in particular to a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method and system. Background Technology

[0002] Digital subtraction angiography (DSA) is a medical imaging technique that assists in observing the internal structure of blood vessels. By injecting a contrast agent, a DSA sequence of continuous images is obtained. Identical parts of the images are removed, leaving only the vascular structure, in order to complete the task of vascular segmentation. It plays an important role, especially in the diagnosis of cerebrovascular diseases, interventional treatment planning, and surgical navigation.

[0003] Currently, vessel segmentation faces numerous challenges, including the strong temporal dependence of DSA sequences and the complexity of vascular structures. Therefore, existing techniques have proposed CNNs, Transformers, or Mamba models for vessel segmentation modeling; however, these methods still have limitations. First, vessel segmentation in DSA sequences relies on the simultaneous fusion of spatial features of vascular structure and temporal features of contrast agent perfusion. The locality of convolutional operations in traditional convolutional neural networks limits the ability to capture long-range dependencies and results in computational redundancy. While Transformer-based models can capture global context, the secondary computational complexity of the self-attention mechanism is significantly reduced in efficiency for long-sequence DSA data, making it difficult to meet clinical demands for real-time performance. Existing technologies suffer from insufficient spatiotemporal coupling modeling capabilities.

[0004] Secondly, while the linear complexity modeling capability of the Mamba model can solve the problem of long sequence modeling, it is mostly applicable to 2D / 3D static images and cannot make full use of the unique characteristics of DSA sequences, which are "subtle changes between frames and critical global temporal sequence". Furthermore, it fails to integrate the strong vascular structure prior provided by minimum density projection (MinIP) images, resulting in insufficient actual segmentation accuracy for low-contrast small blood vessels, which can easily lead to missed detections or false positives. Existing technologies suffer from the problem of missing structural priors and temporal dynamic fusion.

[0005] Finally, existing technologies often employ simple feature splicing or symmetric attention for spatiotemporal feature fusion to address the aforementioned issues, without further emphasizing the dominance of spatial structure in the segmentation task. This results in fused features being susceptible to background noise interference, leading to a decrease in the accuracy of the blood vessel boundaries delineated by the actual segmentation, which in turn affects the topological continuity of the segmentation structure. Existing technologies suffer from an imbalance between the efficiency and accuracy of feature fusion.

[0006] It is evident that existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a bidirectional asymmetric DSA sequence brain blood vessel segmentation method and system, which can achieve efficient spatiotemporal fusion and significantly improve the accuracy, computational efficiency and robustness of blood vessel segmentation.

[0008] To address the aforementioned technical problems, the first aspect of this invention discloses a bidirectional asymmetric DSA sequence method for cerebral blood vessel segmentation, the method comprising: Acquire raw contrast images and preprocess them to generate DSA sequences; Based on a preset dual-branch coding mechanism, complementary spatiotemporal features of the DSA sequence are extracted. These complementary spatiotemporal features include mutually independent and heterogeneous temporal and spatial features. For the complementary spatiotemporal features, temporal modeling is performed based on a preset selective spatial state model to obtain spatiotemporal dynamic features, which include temporal dynamic features and spatial dynamic features of the same dimension. Based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, spatiotemporal control features are obtained. Based on the space-dominated asymmetric cross-attention mechanism, the spatiotemporal control features are fused by cross-attention to generate a target spatiotemporal representation. The spatiotemporal representation of the target is decoded to obtain the cerebral blood vessel segmentation result of the DSA sequence.

[0009] As an optional implementation, in the first aspect of the present invention, complementary spatiotemporal features of the DSA sequence are extracted based on a preset dual-branch coding mechanism, including: The DSA sequence is input into a time encoder to capture the filling and / or diffusion process of contrast agent in the vascular network over time for hemodynamic modeling, and to extract time features of the DSA sequence that characterize the continuous dynamic properties of the blood vessels from the time branch. The minimum intensity projection image of the DSA sequence is input into the spatial encoder. During the downsampling stage, the receptive field of multi-level feature extraction is expanded frame by frame and the global spatial context is captured. The spatial features of the vascular structure prior enhanced by the DSA sequence are extracted in the spatial branch. The minimum intensity projection image is generated by taking the minimum intensity value at each pixel coordinate in the time dimension of the DSA sequence image.

[0010] As an optional implementation, in the first aspect of the present invention, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, including: The temporal features are flattened in terms of spatial dimension to obtain a temporal feature sequence; The temporal feature sequence is mapped to a low-dimensional latent space based on the target embedding dimension through linear projection to obtain a low-dimensional feature sequence of temporal features; The temporal dynamics of the low-dimensional feature sequence are captured through linear computation using a selective spatial state model, and the temporal dynamic features are output. The cyclical process of the linear computation is represented as follows: ; ; in, This represents the hidden state at time t. This represents the low-dimensional feature input at time t in the low-dimensional feature sequence. The output value is obtained by linearly combining the hidden state at time t with the low-dimensional feature input. A, E, C, and D are learnable parameters of the selective spatial state model.

[0011] As an optional implementation, in the first aspect of the present invention, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, and the method further includes: The spatial features are tokenized and embedded into the target embedding dimension to obtain spatial dynamic features. The tokenization and embedding process is represented as follows: ; ; in, Indicates spatial characteristics, This represents the dimensional structure for adjusting spatial features. Represents spatial feature sequences, In the expression, the tensor belongs to the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. Represents a linear transformation matrix. Indicates the target embedding dimension. It represents the dynamic characteristics of space.

[0012] As an optional implementation, in the first aspect of the present invention, the preset bidirectional guidance enhancement mechanism specifically includes: The first guidance sets the temporal dynamic features as temporal queries and the spatial dynamic features as spatial keys and spatial values, constructs the spatial structure's constraint on temporal evolution, and outputs the temporally modulated spatial modulation features. The first guidance is a temporally guided spatial optimization process. The second guidance sets the spatial dynamic features as temporal queries and the temporal dynamic features as spatial keys and spatial values, constructs the constraints of temporal information on spatial structure enhancement, and outputs temporal modulation features enhanced by spatial structure. The second guidance is a space-dominated temporal enhancement process. The expression for the first guiding term is: ; ; ; ; In the formula, Represents temporal dynamic characteristics, The time-series query transformation matrix represents the learnable parameters. Represents the time-series query matrix. Represents spatial dynamic characteristics, The space key transformation matrix represents the learnable parameters. Represents the space bond matrix, This represents the spatial value transformation matrix of the learnable parameters. Represents a spatial value matrix, This represents the similarity calculation of the dot product attention between the temporal query and the spatial key. This indicates the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation; The expression for the second guideline is: ; ; ; ; In the formula, The spatial query transformation matrix represents the learnable parameters. Represents a spatial query matrix. The time-series key transformation matrix represents the learnable parameters. Represents the time-series value transformation matrix. Represents the time-series key matrix. Represents the time series value matrix. This represents the similarity calculation of the dot product attention between the spatial query and the temporal key. This indicates the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement.

[0013] As an optional implementation, in the first aspect of the invention, the space-dominated asymmetric cross-attention mechanism specifically includes: The temporal modulation feature in the aforementioned temporal control features is set as a query from space to time based on the spatial structure as the dominant factor; Set the spatial modulation feature in the time-space control feature as a key and value from time to space; Calculate the matching degree matrix between queries in the spatial-temporal direction and keys in the temporal-spatial direction, and solve for the attention weights in the matching degree matrix; Based on the weighted summation of the attention weights and the values ​​from time to space, the target spatiotemporal features of the cross-attention fusion spatiotemporal information are output. The spatiotemporal features of the target are reconstructed according to the dimensions of the DSA sequence, and the spatiotemporal representation of the target is output. The calculation process for the spatiotemporal representation of the target is expressed as follows: ; ; ; ; ; In the formula, A query matrix representing the spatial to temporal direction. The bond matrix represents the time-to-space direction. A matrix representing the values ​​from time to space. This represents the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement. This represents the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation. The learnable linear transformation matrix representing the query direction. The learnable linear transformation matrix representing the bond direction. The learnable linear transformation matrix representing the value direction. This represents the matching degree matrix between queries and key dot products that represent cross-interest in time and space. The attention weights are represented by scaling factors and normalization, and F represents the spatiotemporal features of the target. Represents feature reconstruction, Represents the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. This represents the spatiotemporal characterization of the target.

[0014] As an optional implementation, in the first aspect of the invention, decoding the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence includes... The first decoding process involves doubling the spatial size of the input target spatiotemporal representation through transposed convolution and halving the number of channels. The second decoding involves combining the spatial features extracted by the dual-branch coding mechanism with the target spatiotemporal representation from the first decoding through a skip connection mechanism across all upsampled channel dimensions. The third decoding process involves fusing the spliced ​​features into the residual convolutional block to output a target feature map of the same size as the original angiographic image, and a segmentation probability map corresponding to the target feature map output by the activation function. The residual convolutional block sequentially includes two 3×3 convolutional layers, batch normalization, and an activation function.

[0015] As an optional implementation, in the first aspect of the present invention, before the acquisition of the original contrast-enhanced image and preprocessing to generate the DSA sequence, the following steps are included: The target loss function for cerebral vessel segmentation is constructed using BCE loss and Dice loss, specifically expressed as follows: ; ; ; in, This represents the BCE loss, where N represents the total number of pixels. This represents the true label of the i-th sample. This represents the predicted probability of the i-th sample. This indicates Dice's loss. Indicates the true label, This represents the probability predicted by the model. Represents the smoothing factor. This represents the category index of tissues / structures in cerebral blood vessels, where C represents the number of categories of tissues / structures in cerebral blood vessels.

[0016] A second aspect of this invention discloses a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system, the system comprising: The data acquisition module is used to acquire raw contrast images and preprocess them to generate DSA sequences; A bidirectional encoding module is used to extract complementary spatiotemporal features of the DSA sequence based on a preset bi-branch encoding mechanism. The complementary spatiotemporal features include mutually independent and heterogeneous temporal features and spatial features. The temporal modeling module is used to perform temporal modeling on the complementary spatiotemporal features based on a preset selective spatial state model to obtain spatiotemporal dynamic features, wherein the spatiotemporal dynamic features include temporal dynamic features and spatial dynamic features of the same dimension. The bidirectional asymmetric fusion module is used to obtain spatiotemporal control features based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, and to perform cross-attention fusion on the spatiotemporal control features based on the space-dominated asymmetric cross-attention mechanism to generate a target spatiotemporal representation. The decoding module is used to decode the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence.

[0017] As an optional implementation, in a second aspect of the invention, complementary spatiotemporal features of the DSA sequence are extracted based on a preset dual-branch coding mechanism, including: The DSA sequence is input into a time encoder to capture the filling and / or diffusion process of contrast agent in the vascular network over time for hemodynamic modeling, and to extract time features of the DSA sequence that characterize the continuous dynamic properties of the blood vessels from the time branch. The minimum intensity projection image of the DSA sequence is input into the spatial encoder. During the downsampling stage, the receptive field of multi-level feature extraction is expanded frame by frame and the global spatial context is captured. The spatial features of the vascular structure prior enhanced by the DSA sequence are extracted in the spatial branch. The minimum intensity projection image is generated by taking the minimum intensity value at each pixel coordinate in the time dimension of the DSA sequence image.

[0018] As an optional implementation, in a second aspect of the invention, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, including: The temporal features are flattened in terms of spatial dimension to obtain a temporal feature sequence; The temporal feature sequence is mapped to a low-dimensional latent space based on the target embedding dimension through linear projection to obtain a low-dimensional feature sequence of temporal features; The temporal dynamics of the low-dimensional feature sequence are captured through linear computation using a selective spatial state model, and the temporal dynamic features are output. The cyclical process of the linear computation is represented as follows: ; ; in, This represents the hidden state at time t. This represents the low-dimensional feature input at time t in the low-dimensional feature sequence. The output value is obtained by linearly combining the hidden state at time t with the low-dimensional feature input. A, E, C, and D are learnable parameters of the selective spatial state model.

[0019] As an optional implementation, in a second aspect of the invention, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, further comprising: The spatial features are tokenized and embedded into the target embedding dimension to obtain spatial dynamic features. The tokenization and embedding process is represented as follows: ; ; in, Indicates spatial characteristics, This represents the dimensional structure for adjusting spatial features. Represents spatial feature sequences, In the expression, the tensor belongs to the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. Represents a linear transformation matrix. Indicates the target embedding dimension. It represents the dynamic characteristics of space.

[0020] As an optional implementation, in the second aspect of the present invention, the preset bidirectional guidance enhancement mechanism specifically includes: The first guidance sets the temporal dynamic features as temporal queries and the spatial dynamic features as spatial keys and spatial values, constructs the spatial structure's constraint on temporal evolution, and outputs the temporally modulated spatial modulation features. The first guidance is a temporally guided spatial optimization process. The second guidance sets the spatial dynamic features as temporal queries and the temporal dynamic features as spatial keys and spatial values, constructs the constraints of temporal information on spatial structure enhancement, and outputs temporal modulation features enhanced by spatial structure. The second guidance is a space-dominated temporal enhancement process. The expression for the first guiding term is: ; ; ; ; In the formula, Represents temporal dynamic characteristics, The time-series query transformation matrix represents the learnable parameters. Represents the time-series query matrix. Represents spatial dynamic characteristics, The space key transformation matrix represents the learnable parameters. Represents the space bond matrix, This represents the spatial value transformation matrix of the learnable parameters. Represents a spatial value matrix, This represents the similarity calculation of the dot product attention between the temporal query and the spatial key. This indicates the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation; The expression for the second guideline is: ; ; ; ; In the formula, The spatial query transformation matrix represents the learnable parameters. Represents a spatial query matrix. The time-series key transformation matrix represents the learnable parameters. Represents the time-series value transformation matrix. Represents the time-series key matrix. Represents the time series value matrix. This represents the similarity calculation of the dot product attention between the spatial query and the temporal key. This indicates the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement.

[0021] As an optional implementation, in the second aspect of the invention, based on a space-dominated asymmetric cross-attention mechanism, specifically: The temporal modulation feature in the aforementioned temporal control features is set as a query from space to time based on the spatial structure as the dominant factor; Set the spatial modulation feature in the time-space control feature as a key and value from time to space; Calculate the matching degree matrix between queries in the spatial-temporal direction and keys in the temporal-spatial direction, and solve for the attention weights in the matching degree matrix; Based on the weighted summation of the attention weights and the values ​​from time to space, the target spatiotemporal features of the cross-attention fusion spatiotemporal information are output. The spatiotemporal features of the target are reconstructed according to the dimensions of the DSA sequence, and the spatiotemporal representation of the target is output. The calculation process for the spatiotemporal representation of the target is expressed as follows: ; ; ; ; ; In the formula, A query matrix representing the spatial to temporal direction. The bond matrix represents the time-to-space direction. A matrix representing the values ​​from time to space. This represents the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement. This represents the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation. The learnable linear transformation matrix representing the query direction. The learnable linear transformation matrix representing the bond direction. The learnable linear transformation matrix representing the value direction. This represents the matching degree matrix between queries and key dot products that represent cross-interest in time and space. The attention weights are represented by scaling factors and normalization, and F represents the spatiotemporal features of the target. Represents feature reconstruction, Represents the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. This represents the spatiotemporal characterization of the target.

[0022] As an optional implementation, in a second aspect of the invention, decoding the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence includes... The first decoding process involves doubling the spatial size of the input target spatiotemporal representation through transposed convolution and halving the number of channels. The second decoding involves combining the spatial features extracted by the dual-branch coding mechanism with the target spatiotemporal representation from the first decoding through a skip connection mechanism across all upsampled channel dimensions. The third decoding process involves fusing the spliced ​​features into the residual convolutional block to output a target feature map of the same size as the original angiographic image, and a segmentation probability map corresponding to the target feature map output by the activation function. The residual convolutional block sequentially includes two 3×3 convolutional layers, batch normalization, and an activation function.

[0023] As an optional implementation, in a second aspect of the invention, before acquiring the original contrast-enhanced image and preprocessing it to generate the DSA sequence, the following steps are included: The target loss function for cerebral vessel segmentation is constructed using BCE loss and Dice loss, specifically expressed as follows: ; ; ; in, This represents the BCE loss, where N represents the total number of pixels. This represents the true label of the i-th sample. This represents the predicted probability of the i-th sample. This indicates Dice's loss. Indicates the true label, This represents the probability predicted by the model. Represents the smoothing factor. This represents the category index of tissues / structures in cerebral blood vessels, where C represents the number of categories of tissues / structures in cerebral blood vessels.

[0024] A third aspect of this invention discloses another bidirectional asymmetric DSA sequence cerebral vessel segmentation system, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method disclosed in the first aspect of the present invention.

[0025] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method disclosed in the first aspect of the present invention.

[0026] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention independently extracts temporal and spatial features from DSA sequences using a dual-branch coding architecture, forming heterogeneous and complementary spatiotemporal feature representations. Temporal features capture the dynamic evolution of blood vessels, while spatial features retain static structural information, avoiding information redundancy caused by the "temporal-spatial" coupling in traditional single-branch models and improving the diversity of feature representation. Leveraging the linear time complexity of the state-space model, the computational cost of long sequences is significantly reduced. A bidirectional guided enhancement mechanism modulates temporal and spatial dynamic features to generate temporal-spaced features. Temporal features guide spatial features to focus on dynamically changing regions, while spatial features constrain temporal features to focus on structural consistency, achieving a bidirectional enhancement of time-guided spatial evolution and spatial constraint on temporal continuity. This enhances the adaptability of features to complex vascular topologies. A space-dominant asymmetric cross-attention mechanism is employed, using spatial features as the primary driver to aggregate temporal information through attention weights, avoiding redundant computations in traditional symmetric attention. The asymmetric design reduces computational complexity while maintaining high sensitivity to vascular edges. Utilizing the fused dynamic and static information in the spatiotemporal representation, vascular boundaries are accurately located. Through complementary features and bidirectional enhancement, the accuracy of vascular edge segmentation is significantly superior to existing methods. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0028] Figure 1 This is a schematic flowchart of a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method disclosed in an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the structure of a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system disclosed in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of another bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the BTSAnet architecture proposed in a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method disclosed in an embodiment of the present invention. Figure 5 for Figure 4 A schematic diagram of the bidirectional asymmetric spatiotemporal attention (BTSA) module. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0032] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0033] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0034] Before detailing the specific embodiments of the technical solution of this application, we will first reveal the model architecture and application scenarios suitable for supporting the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method in the technical solution of this application.

[0035] The technical solution of this application is applicable to the field of medical image segmentation technology, particularly to the image segmentation of cerebral vascular DSA sequences. In this context, the technical solution of this application can be applied to the BTSAnet model architecture it discloses, such as... Figure 4-5 As shown, the system consists of a dual-branch encoder, a Mamba timing module, and a bidirectional asymmetric attention fusion (BTAS) module.

[0036] Digital subtraction angiography (DSA) sequences are a key imaging technique for diagnosing cerebrovascular diseases, but their vessel segmentation task faces the following challenges: (1) Strong time-dependent: DSA sequence dynamically displays the flow process of contrast agent in blood vessels, while traditional single-frame segmentation methods cannot capture the complete blood vessel topology.

[0037] (2) Complex vascular structure: small blood vessels have low contrast with the background, making them easy to miss or missegment.

[0038] Limitations of existing methods: (1) Insufficient spatiotemporal coupling modeling capability: The segmentation of blood vessels in DSA sequences needs to integrate the features of "spatial vascular structure" and "temporal contrast agent perfusion". Traditional convolutional neural networks (CNNs) (such as CAVE and VSS-Net) rely on ConvLSTM / GRU to model temporal information, but the locality of convolution operations limits the ability to capture long-range dependencies and there is computational redundancy; although Transformer-based models (such as TransUNet and SwinUNet) can capture global context, the secondary computational complexity (O(N²)) of the self-attention mechanism is extremely inefficient on long-sequence DSA data, which is difficult to meet the clinical real-time requirements.

[0039] (2) Lack of structural prior and temporal dynamic fusion: Existing Mamba-based segmentation models (such as U-Mamba and Mamba-UNet) have the ability to model long sequences with linear complexity (O(N)), but they are mostly for 2D / 3D static images. They do not make full use of the unique characteristics of DSA sequences, which are "subtle changes between frames but key global temporal patterns". Furthermore, they do not fuse the strong vascular structure prior provided by minimum density projection (MinIP) images, resulting in insufficient segmentation accuracy for low-contrast small blood vessels, which can easily lead to missed detections or false positives.

[0040] (3) Imbalance between feature fusion efficiency and accuracy: Existing methods mostly use simple splicing or symmetric attention to fuse spatiotemporal features, without emphasizing the dominance of spatial structure in the segmentation task. This makes the fused features susceptible to background noise interference, the blood vessel boundaries (especially terminal branches) are not accurately delineated, and the topological continuity of the segmentation results is poor.

[0041] To address the aforementioned issues, this application proposes a spatiotemporal segmentation network based on a dual-branch encoder, Mamba architecture, BTSA module, and decoder. The dual-branch encoder captures structural priors and temporal dynamics respectively, and the BTSA module achieves efficient spatiotemporal fusion, while balancing segmentation accuracy and computational efficiency, providing a technical path for intelligent analysis of DSA sequences.

[0042] Based on the aforementioned spatiotemporal segmentation network architecture, this invention discloses a bidirectional asymmetric DSA sequence cerebral vessel segmentation method and system. Through a dual-branch coding architecture, it independently extracts temporal and spatial features of the DSA sequence, forming heterogeneous and complementary spatiotemporal feature representations. Temporal features capture the dynamic evolution of blood vessels, while spatial features retain static structural information, avoiding information redundancy caused by the "temporal-spatial" coupling in traditional single-branch models and improving the diversity of feature representation. Utilizing the linear time complexity of the state-space model, it significantly reduces the computational overhead of long sequences. Through a bidirectional guidance enhancement mechanism, temporal dynamic features and spatial dynamic features are bidirectionally modulated to generate temporal modulation features, with temporal features guiding spatial features. This approach focuses on dynamically changing regions, using spatial features to constrain temporal features and ensure structural consistency. It achieves bidirectional enhancement of temporal continuity by guiding spatial evolution through temporal constraints, improving the adaptability of features to complex vascular topologies. Employing a space-dominant asymmetric cross-attention mechanism, it aggregates temporal information through attention weights, avoiding redundant computations in traditional symmetric attention. The asymmetric design reduces computational complexity while maintaining high sensitivity to vascular edges. Utilizing the fused dynamic and static information from spatiotemporal representations, it accurately locates vascular boundaries. Through complementary features and bidirectional enhancement, it significantly outperforms existing methods in vascular segmentation accuracy, computational efficiency, and robustness, particularly suitable for complex multi-clinical diagnostic and treatment scenarios. These will be explained in detail below.

[0043] Example 1 Please see Figure 1 , Figure 1 This is a schematic flowchart of a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method disclosed in an embodiment of the present invention. Wherein, Figure 1 The described bidirectional asymmetric DSA sequence cerebral vessel segmentation method can be applied to data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 1 As shown, this bidirectional asymmetric DSA sequence cerebral vessel segmentation method may include the following operations: 101. Acquire raw contrast images and preprocess them to generate DSA sequences.

[0044] It should be noted that the original angiography images are multiple frames of raw images continuously captured by imaging equipment during angiography, including mask images and angiography images. They contain a large amount of background noise (such as bone, soft tissue, motion artifacts, etc.) and cannot be directly used for accurate vascular analysis.

[0045] Specifically, the DSA sequence is a dynamic image sequence that retains only the vascular structure after the original angiography image has undergone digital subtraction. First, the mask image and the angiography image are spatially aligned to solve the problem of poor registration artifacts. Then, the DSA frame = angiography image - mask image is calculated to eliminate background tissue (bone / soft tissue) and retain only the contrast-filled vascular area for digital subtraction. Finally, the subtracted image is denoised and enhanced (such as contrast adjustment) to optimize the sequence, improve the clarity of the vascular edges, and provide high-quality input for subsequent segmentation.

[0046] Optionally, the data of the original contrast images (such as the model training set) can be obtained from publicly available medical image databases.

[0047] 102. Based on a preset dual-branch coding mechanism, extract complementary spatiotemporal features of the DSA sequence, wherein the complementary spatiotemporal features include mutually independent and heterogeneous temporal features and spatial features.

[0048] Optionally, in order to effectively capture complementary spatiotemporal information in the DSA sequence, a dual-branch encoder is designed to run a dual-branch coding mechanism, wherein the dual branches are a temporal branch and a spatial branch, and the encoder includes a temporal encoder and a spatial encoder.

[0049] Furthermore, in the dual-branch coding architecture, the temporal encoder can process the original DSA sequence to model hemodynamics, while the spatial encoder processes the minimum intensity projection image generated from the sequence to obtain enhanced vascular structure priors, thus achieving synergistic utilization of the temporal dynamic information of the original sequence and the enhanced spatial structure priors provided by the MinIP image.

[0050] 103. For the complementary spatiotemporal features, perform temporal modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, wherein the spatiotemporal dynamic features include temporal dynamic features and spatial dynamic features of the same dimension.

[0051] Optionally, the Mamba module can be used in the network architecture for temporal modeling based on the Selective State Space Model (SSM), which can efficiently capture temporal dynamics with linear computational complexity and perform efficient long-range temporal dependency modeling on the features output by the encoder.

[0052] 104. Based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, obtain spatiotemporal control features, and perform cross-attention fusion on the spatiotemporal control features based on the space-dominated asymmetric cross-attention mechanism to generate a target spatiotemporal representation.

[0053] 105. Decode the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence.

[0054] As can be seen, the above-mentioned embodiments of the invention independently extract temporal and spatial features of DSA sequences through a dual-branch coding architecture, forming heterogeneous and complementary spatiotemporal feature representations. Temporal features capture the dynamic evolution of blood vessels, while spatial features retain static structural information, avoiding information redundancy caused by the "temporal-spatial" coupling in traditional single-branch models, and improving the diversity of feature representation. Through the linear time complexity characteristics of the state-space model, the computational cost of long sequences is significantly reduced. Through a bidirectional guided enhancement mechanism, temporal dynamic features and spatial dynamic features are bidirectionally modulated to generate temporal control features. Temporal features guide spatial features to focus on dynamically changing regions, while spatial features constrain temporal features to focus on structural consistency, achieving bidirectional enhancement of time-guided spatial evolution and space-constrained temporal continuity, improving the adaptability of features to complex vascular topologies. A spatially dominated asymmetric cross-attention mechanism is adopted, with spatial features as the main driver to aggregate temporal information through attention weights, avoiding redundant computations in traditional symmetric attention. The asymmetric design reduces computational complexity while maintaining high sensitivity to vascular edges. By utilizing the fused dynamic and static information in the spatiotemporal representation, the vascular boundary is accurately located. Through complementary features and bidirectional enhancement, the vascular edge segmentation accuracy is significantly better than existing methods.

[0055] As an optional embodiment, the step above, extracting complementary spatiotemporal features of the DSA sequence based on a preset dual-branch coding mechanism, includes: The DSA sequence is input into a time encoder to capture the filling and / or diffusion process of contrast agent in the vascular network over time for hemodynamic modeling, and to extract time features of the DSA sequence that characterize the continuous dynamic properties of the blood vessels from the time branch. The minimum intensity projection image of the DSA sequence is input into the spatial encoder. During the downsampling stage, the receptive field of multi-level feature extraction is expanded frame by frame and the global spatial context is captured. The spatial features of the vascular structure prior enhanced by the DSA sequence are extracted in the spatial branch. The minimum intensity projection image is generated by taking the minimum intensity value at each pixel coordinate in the time dimension of the DSA sequence image.

[0056] Specifically, the timing encoder uses the original DSA sequence As input, frame-by-frame feature extraction is performed in the temporal branch using residual convolutional blocks. The goal is to capture the filling and diffusion process of contrast agent in the vascular network over time, i.e., dynamic temporal information, and then output as temporal features. ,in, These represent the batch size, time duration, number of input channels, input image height, and input image width, respectively. These are the number of channels, height, and width of the output feature, respectively.

[0057] Next, the input to the spatial encoder is the minimum intensity projection image (MinIP). The imaging principle of MinIP is to take the minimum intensity value at each pixel coordinate in the time dimension. The mathematical definition of is: ; In the formula, The MinIP image is located at... pixel values, The position of the DSA image in frame t is... The intensity value, It is the total number of frames in the DSA sequence.

[0058] Furthermore, the spatial encoder extracts multi-scale hierarchical features from the MinIP image through residual convolutional blocks, each consisting of two convolutional layers, a batch normalization layer, and a nonlinear activation function. Downsampling is performed between different stages to progressively expand the receptive field and capture the global spatial context. This spatial branch outputs spatial features. .

[0059] As can be seen, the above optional embodiments overcome the limitations of a single information source through a dual-branch design. The dynamic information extracted by temporal encoding helps distinguish between real blood vessels (with continuous dynamic characteristics) and static artifacts, while the MinIP prior provided by spatial encoding serves as a spatial "navigation map," significantly improving the network model's ability to perceive vascular topology, especially the identification of small branches. By processing these two types of information independently during the encoding stage, premature mixing of heterogeneous features is avoided, laying a solid data foundation for more accurate feature fusion later.

[0060] As an optional embodiment, the step above, performing time-series modeling on the complementary spatiotemporal features based on a preset selective spatial state model to obtain spatiotemporal dynamic features, includes: The temporal features are flattened in terms of spatial dimension to obtain a temporal feature sequence; The temporal feature sequence is mapped to a low-dimensional latent space based on the target embedding dimension through linear projection to obtain a low-dimensional feature sequence of temporal features; The temporal dynamics of the low-dimensional feature sequence are captured through linear computation using a selective spatial state model, and the temporal dynamic features are output. The cyclical process of the linear computation is represented as follows: ; ; in, This represents the hidden state at time t. This represents the low-dimensional feature input at time t in the low-dimensional feature sequence. The output value is obtained by linearly combining the hidden state at time t with the low-dimensional feature input. A, E, C, and D are learnable parameters of the selective spatial state model.

[0061] Specifically, in order to effectively integrate the spatiotemporal features in the DSA sequence and make full use of the vascular structure prior provided by the MinIP image, a Bidirectional Temporal-Spatial Attention (BTSA) module is proposed. This module is placed at the network bottleneck and aims to achieve efficient interaction of spatiotemporal features with low computational cost.

[0062] Furthermore, although the original DSA sequence contains important temporal information, the vascular morphology changes between frames are relatively small. Performing dense temporal modeling in shallow networks is not only computationally redundant but may also interfere with the learning of spatial structural features. Therefore, a Mamba module is introduced at the bottleneck to perform efficient long-range temporal dependency modeling on the features output by the temporal encoder.

[0063] Furthermore, the features extracted by the temporal encoder are First, the spatial dimension is flattened to obtain the temporal feature sequence, which is represented as follows: ; Then, it is mapped to a low-dimensional latent space through linear projection, the process of which is represented as follows: Where d is the target embedding dimension, The Mamba module is based on the Selective State-Space Model (SSM), which can efficiently capture temporal dynamics with linear computational complexity.

[0064] As can be seen, through the above optional embodiments, the local continuity of temporal features is preserved by spatial flattening and low-dimensional projection; combined with selective modeling, the dynamic evolution law of blood vessels is accurately extracted; the linear complexity design improves the efficiency of the model in processing ultra-long sequences, meets the real-time clinical needs, and the selective mechanism dynamically filters irrelevant background information, which is especially suitable for multiple clinical scenarios with differences in DSA data from different patients and hospitals.

[0065] As an optional embodiment, the step above, performing time-series modeling on the complementary spatiotemporal features based on a preset selective spatial state model to obtain spatiotemporal dynamic features, further includes: The spatial features are tokenized and embedded into the target embedding dimension to obtain spatial dynamic features. The tokenization and embedding process is represented as follows: ; ; in, Indicates spatial characteristics, This represents the dimensional structure for adjusting spatial features. Represents spatial feature sequences, In the expression, the tensor belongs to the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. Represents a linear transformation matrix. Indicates the target embedding dimension. It represents the dynamic characteristics of space.

[0066] As can be seen, through the above optional embodiments, spatial features are serialized and recombined by adjusting the dimensional structure and serializing the features. This not only preserves the spatial location information of the features (i.e., the combination of height h and width w), but also provides a structurally adapted feature representation for subsequent linear transformations. By dimensional alignment and feature semantic enhancement, the channel dimension of the spatial features is unified with the embedding dimension required for subsequent temporal modeling, solving the problem of feature dimension mismatch and improving the model's compatibility in processing spatiotemporal features. By capturing the correlation between channels through linear combination, the feature's ability to express spatial information is enhanced, making the features more in line with the semantic requirements of downstream tasks. This realizes the transformation of spatial features into dynamic features, and optimizes feature adaptability, semantic expressiveness, and spatiotemporal modeling synergy, laying a feature foundation for improving the performance of spatiotemporal tasks.

[0067] As an optional embodiment, the preset bidirectional boot enhancement mechanism in the above steps includes: The first guidance sets the temporal dynamic features as temporal queries and the spatial dynamic features as spatial keys and spatial values, constructs the spatial structure's constraint on temporal evolution, and outputs the temporally modulated spatial modulation features. The first guidance is a temporally guided spatial optimization process. The second guidance sets the spatial dynamic features as temporal queries and the temporal dynamic features as spatial keys and spatial values, constructs the constraints of temporal information on spatial structure enhancement, and outputs temporal modulation features enhanced by spatial structure. The second guidance is a space-dominated temporal enhancement process. The expression for the first guiding term is: ; ; ; ; In the formula, Represents temporal dynamic characteristics, The time-series query transformation matrix represents the learnable parameters. Represents the time-series query matrix. Represents spatial dynamic characteristics, The space key transformation matrix represents the learnable parameters. Represents the space bond matrix, This represents the spatial value transformation matrix of the learnable parameters. Represents a spatial value matrix, This represents the similarity calculation of the dot product attention between the temporal query and the spatial key. This indicates the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation; The expression for the second guideline is: ; ; ; ; In the formula, The spatial query transformation matrix represents the learnable parameters. Represents a spatial query matrix. The time-series key transformation matrix represents the learnable parameters. Represents the time-series value transformation matrix. Represents the time-series key matrix. Represents the time series value matrix. This represents the similarity calculation of the dot product attention between the spatial query and the temporal key. This indicates the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement.

[0068] It is worth noting that the bidirectional asymmetric cross-attention mechanism provided by the BTSA module in the BTSAnet architecture proposed in this application can achieve mutual guidance and enhancement of spatial and temporal features.

[0069] Specifically, in the temporally guided spatial optimization process, temporal features are used as queries and spatial features as keys to constrain the temporal evolution of the spatial structure. The first guiding path reflects the role of temporal guidance. Hemodynamic dynamics are used as queries to lock the relevant regions in the spatial structure, enabling the model to further optimize and confirm spatial features based on the filling process. This is particularly helpful in identifying weak vascular segments due to uneven contrast agent filling and suppressing background noise and spurious responses from non-vascular regions.

[0070] Following this, in the space-dominated temporal enhancement process, spatial dynamic features extracted from the MinIP images are used as queries, while temporal dynamic features modeled from the DSA sequence via the Mamba module are used as keys and values, thus enhancing the spatial structure with temporal information. The second guiding path embodies the asymmetry of space dominance, using the vascular spatial structure as the query to retrieve relevant information within the temporal dynamics. This ensures that any enhanced dynamic changes occur within the context of the vascular structure, fundamentally avoiding misinterpreting background noise as vascular dynamics.

[0071] Furthermore, the BTSA module employs a bidirectional asymmetric cross-attention mechanism, emphasizing the dominant role of spatial structure in the fusion process. Unlike traditional self-attention mechanisms, this application incorporates both bidirectional and asymmetric design. By constructing two paths—"space-dominated temporal enhancement" and "temporally guided spatial optimization"—it achieves mutual calibration of spatiotemporal features, resulting in a bidirectional effect in information flow. This provides a data foundation for the subsequent asymmetric cross-attention fusion process.

[0072] As can be seen, through the above optional embodiments, by using temporal dynamic features as queries and spatial dynamic features as keys and values, an attention mechanism is used to construct constraints on the temporal evolution of spatial structure. Dynamic constraints on spatial changes suppress irrelevant background noise and improve the adaptability of spatial structure to dynamic processes. By using spatial dynamic features as queries and temporal dynamic features as keys and values, an attention mechanism is used to construct enhanced constraints on spatial structure of temporal information. Structured enhancement of temporal representation eliminates redundant temporal information, improves the robustness of temporal modeling to complex structures, and bidirectional guidance and collaboration form a bidirectional calibration mechanism for spatiotemporal features, avoiding deviations caused by unidirectional modeling, adapting to spatiotemporal coupling modes in different scenarios, and enhancing the model's adaptability to complex data.

[0073] As an optional embodiment, the space-dominated asymmetric cross-attention mechanism in the above steps includes: The temporal modulation feature in the aforementioned temporal control features is set as a query from space to time based on the spatial structure as the dominant factor; Set the spatial modulation feature in the time-space control feature as a key and value from time to space; Calculate the matching degree matrix between queries in the spatial-temporal direction and keys in the temporal-spatial direction, and solve for the attention weights in the matching degree matrix; Based on the weighted summation of the attention weights and the values ​​from time to space, the target spatiotemporal features of the cross-attention fusion spatiotemporal information are output. The spatiotemporal features of the target are reconstructed according to the dimensions of the DSA sequence, and the spatiotemporal representation of the target is output. The calculation process for the spatiotemporal representation of the target is expressed as follows: ; ; ; ; ; In the formula, A query matrix representing the spatial to temporal direction. The bond matrix represents the time-to-space direction. A matrix representing the values ​​from time to space. This represents the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement. This represents the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation. The learnable linear transformation matrix representing the query direction. The learnable linear transformation matrix representing the bond direction. The learnable linear transformation matrix representing the value direction. This represents the matching degree matrix between queries and key dot products that represent cross-interest in time and space. The attention weights are represented by scaling factors and normalization, and F represents the spatiotemporal features of the target. Represents feature reconstruction, Represents the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. This represents the spatiotemporal characterization of the target.

[0074] It should be noted that the asymmetry in this application reflects that the source of the query in the second attention computation differs from the bidirectional cross-computation in the first, always proceeding from space to time. The mechanism design, which prioritizes the fusion process over treating it equally, ensures that the fusion process is always anchored by a clear structural prior. This allows the temporal dynamic information to be adaptively aligned and enhance the spatial representation, thereby effectively suppressing background noise and improving the segmentation accuracy of blood vessel boundaries (especially small branches).

[0075] As can be seen, through the above optional embodiments, and by utilizing the "bidirectional and asymmetric" design concept of the BTSA module in the BTSAnet architecture, a highly efficient spatiotemporal feature fusion is achieved. Compared to the self-attention mechanism in existing technologies that involves symmetrical interaction among all features, this embodiment, by granting the spatial structure dominance (as a Query) and establishing bidirectional, clearly defined interaction paths, achieves precise control over the information flow. This ensures that the fused features maintain the integrity of the vascular topology while incorporating discriminative hemodynamic information, thereby effectively improving the accuracy of vascular segmentation while maintaining the integrity of the spatial structure, particularly excelling in suppressing background noise and delineating low-contrast small vessels.

[0076] As an optional embodiment, the above steps, including decoding the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence, include... The first decoding process involves doubling the spatial size of the input target spatiotemporal representation through transposed convolution and halving the number of channels. The second decoding involves combining the spatial features extracted by the dual-branch coding mechanism with the target spatiotemporal representation from the first decoding through a skip connection mechanism across all upsampled channel dimensions. The third decoding process involves fusing the spliced ​​features into the residual convolutional block to output a target feature map of the same size as the original angiographic image, and a segmentation probability map corresponding to the target feature map output by the activation function. The residual convolutional block sequentially includes two 3×3 convolutional layers, batch normalization, and an activation function.

[0077] Specifically, in the BTSAnet architecture, the decoder undertakes the crucial task of reconstructing high-resolution segmentation maps and integrating multi-scale contextual information. The input to the decoder is the spatiotemporal representation of the target output from the bottleneck layer, which has been fused and reshaped by the BTSA modules. The decoder employs a symmetrical upsampling structure, consisting of multiple stages.

[0078] Furthermore, each upsampling stage first uses transposed convolution to double the spatial size of the feature map while halving the number of channels. Then, multi-scale features from the corresponding layer in the spatial encoder (MinIP branch) are introduced through skip connections. These features from the shallow layer of the encoder are rich in details such as the edges and textures of blood vessels. The features from the skip connections are cropped or bilinearly interpolated to match the current resolution, and then concatenated with the upsampled features in the channel dimension.

[0079] Furthermore, the concatenated features are fed into a residual convolution block for fusion and refinement. Each residual block consists of two 3×3 convolutional layers, batch normalization, and ReLU activation functions in sequence to enhance the model's non-linear expressive power.

[0080] Finally, the decoder outputs a feature map with the same spatial size as the original input image. This feature map is mapped from the number of channels to the number of categories (blood vessels and background) through a 1×1 convolutional layer, and the final segmentation probability map is generated through the Sigmoid activation function, thus outputting the segmentation result.

[0081] As can be seen, through the above optional embodiments, the decoder outputs the vessel segmentation results, ensuring that the fusion features rich in high-level spatiotemporal semantics from the BTSA module can be effectively integrated with the multi-level spatial details provided by the spatial encoder. This process gradually restores the spatial resolution and significantly optimizes the accurate segmentation of vessel boundaries and the continuity of topological structure.

[0082] As an optional embodiment, the above steps, before preprocessing the acquired raw contrast images to generate the DSA sequence, include: The target loss function for cerebral vessel segmentation is constructed using BCE loss and Dice loss, specifically expressed as follows: ; ; ; in, This represents the BCE loss, where N represents the total number of pixels. This represents the true label of the i-th sample. This represents the predicted probability of the i-th sample. This indicates Dice's loss. Indicates the true label, This represents the probability predicted by the model. Represents the smoothing factor. This represents the category index of tissues / structures in cerebral blood vessels, where C represents the number of categories of tissues / structures in cerebral blood vessels.

[0083] Specifically, in order to optimize the performance of the blood vessel segmentation model, this embodiment uses a combination of Binary Cross-Entropy Loss (BCE) and Dice loss as the objective function. BCE loss is beneficial for pixel-level classification, while Dice loss can effectively alleviate the class imbalance problem between blood vessels and background. The combination of the two can balance boundary accuracy and region consistency.

[0084] Test case For the spatiotemporal segmentation network based on a dual-branch encoder, Mamba architecture, BTSA module, and decoder, which supports the bidirectional asymmetric DSA sequence cerebral vessel segmentation method proposed in this application, model training, evaluation, and testing were performed on the publicly available DIAS (DSA-sequence Intracranial Artery Segmentation dataset). This dataset contains 60 DSA sequences with pixel-level fine annotations. To ensure the fairness and generalization ability of the evaluation, we adopted a patient-randomized partitioning strategy, dividing the 60 data points into 48 for training and 12 for testing, strictly ensuring no patient overlap between the training and testing sets.

[0085] Specifically, this test case employs a multi-dimensional, comprehensive, and impartial evaluation metric to assess the model's vessel segmentation performance, primarily considering two aspects: the overlap of segmented regions and the accuracy of pixel-level classification. These include: Dice Similarity Coefficient (DSC) and Intersection over Union (IoU): These are used to evaluate the overall consistency between the segmented region and the true label.

[0086] Sensitivity (SEN): Measures the model’s ability to correctly identify blood vessel pixels and avoid missed detections.

[0087] Specificity (Spec): Measures the model’s ability to correctly identify background pixels and suppress false positives.

[0088] Precision (Pre): Measures the proportion of pixels predicted as blood vessels that are actually blood vessels.

[0089] Accuracy (ACC): Reflects the overall classification accuracy of all pixels.

[0090] Furthermore, the model training optimization strategy is optimized using the stochastic gradient descent (SGD) algorithm, with momentum set to 0.99, weight decay coefficient of 1×10-2, initial learning rate set to 0.01 and polynomial learning rate decay (power=0.9).

[0091] The training cycle consists of 500 epochs, with each epoch containing 250 iterations.

[0092] Training and Validation: During the training phase, a five-fold cross-validation strategy was employed on the 48 sequences in the training set. This involved randomly dividing the data into five parts, using four parts sequentially for training, and reserving the remaining part for validation. This approach maximized data utilization and ensured model stability. Data Preprocessing and Augmentation: Each batch consisted of one input sample. The input image size was 512×512 pixels, randomly cropped from the original image. Several data augmentation strategies were employed to enhance model robustness, including: random rotation (-15° to +15°), elastic deformation, random scaling (0.8 to 1.2x), gamma augmentation (coefficient range 0.7 to 1.5), and mirror flipping.

[0093] Furthermore, no data augmentation was used during the testing phase. We cropped 512×512 pixel image patches from the test images using a sliding window, input them into the network for prediction, and then stitched the prediction results together to obtain the final prediction image. The final performance metric for the test set was the average of the quantitative segmentation results of the five models obtained through five-fold cross-validation on the test set, to ensure the statistical reliability and robustness of the results. The results are shown in Table 1 (all metrics are expressed as percentages (%)) comparing the vessel segmentation performance on the DIAS dataset.

[0094] Table 1 - Comparison of vessel segmentation performance on the DIAS dataset (mean ± standard deviation)

[0095] Finally, the advantages of this application compared to the prior art are: (1) The segmentation accuracy is significantly improved, especially for small blood vessels and low-contrast areas: Experimental Validation (based on the DIAS public dataset): As shown in Table 1, the BTSANet proposed in this application achieved the best performance on all evaluation metrics, significantly outperforming other comparative models. The core metrics of the bidirectional asymmetric DSA sequence brain vessel segmentation method proposed in this application, namely the Dice similarity coefficient of 94.39% and the intersection-over-union ratio (IoU) of 89.41%, represent absolute improvements of 5.55% and 9.09% respectively compared to the existing best Mamba model (U-Mamba), and relative improvements of 18.14% and 27.79% respectively compared to the traditional CNN model (UNet). The sensitivity (SEN) reached 93.96%, effectively detecting low-contrast small vessels (diameter <1mm) with a false negative rate reduced by more than 40%. The "spatial-dominated" fusion mechanism of the BTSA module ensures that the temporal dynamics revolve around the vascular structure, and the MinIP structure prior enhances the saliency of the vascular region, avoiding background noise interference, and improving the accuracy of vascular boundary delineation by 15%~20%.

[0096] (2) High computational efficiency, meeting the real-time needs of clinical practice: Advantages in linear complexity: The time-series modeling complexity of the Mamba architecture is O(N), which is an order of magnitude lower than that of the Transformer (O(N²). Deploying Mamba only at the bottleneck reduces the computational cost by 30% to 40% compared to the "full encoder time-series modeling" scheme (FLOPs=124.97G).

[0097] Parameter number optimization: The model has 24.14M parameters, which is only 26% of TransUNet (93.23M) and 58% of SwinUNet (41.34M).

[0098] (3) Highly robust and adaptable to complex clinical scenarios: By using a dual-branch encoder and a BTSA module, structural prior and temporal dynamic information are effectively fused to enhance spatial structure representation, suppress background noise, and improve the stability and robustness of segmentation results. The performance is stable under different sequence lengths (4-8 frames).

[0099] (4) It has strong clinical applicability and supports the diagnosis and treatment needs of multiple scenarios: Automated segmentation results can assist physicians in quickly and accurately interpreting DSA sequences, reducing subjectivity and providing technical support for the diagnosis of cerebrovascular diseases, surgical navigation, and efficacy evaluation, with broad clinical application prospects.

[0100] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system disclosed in an embodiment of the present invention. Figure 2 The described bidirectional asymmetric DSA sequence cerebral vessel segmentation system can be applied to data processing systems / data processing devices / data processing servers (wherein, the server includes a local processing server or a cloud processing server). For example... Figure 2 As shown, this bidirectional asymmetric DSA sequence cerebral vessel segmentation system may include: The data acquisition module 201 is used to acquire raw contrast images and preprocess them to generate DSA sequences.

[0101] The bidirectional encoding module 202 is used to extract complementary spatiotemporal features of the DSA sequence based on a preset bi-branch encoding mechanism. The complementary spatiotemporal features include mutually independent and heterogeneous temporal features and spatial features.

[0102] The temporal modeling module 203 is used to perform temporal modeling on the complementary spatiotemporal features based on a preset selective spatial state model to obtain spatiotemporal dynamic features, wherein the spatiotemporal dynamic features include temporal dynamic features and spatial dynamic features of the same dimension.

[0103] The bidirectional asymmetric fusion module 204 is used to obtain spatiotemporal control features based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, and to perform cross-attention fusion on the spatiotemporal control features based on the space-dominated asymmetric cross-attention mechanism to generate a target spatiotemporal representation.

[0104] The decoding module 205 is used to decode the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence.

[0105] As can be seen, the above-mentioned embodiments of the invention independently extract temporal and spatial features of DSA sequences through a dual-branch coding architecture, forming heterogeneous and complementary spatiotemporal feature representations. Temporal features capture the dynamic evolution of blood vessels, while spatial features retain static structural information, avoiding information redundancy caused by the "temporal-spatial" coupling in traditional single-branch models, and improving the diversity of feature representation. Through the linear time complexity characteristics of the state-space model, the computational cost of long sequences is significantly reduced. Through a bidirectional guided enhancement mechanism, temporal dynamic features and spatial dynamic features are bidirectionally modulated to generate temporal control features. Temporal features guide spatial features to focus on dynamically changing regions, while spatial features constrain temporal features to focus on structural consistency, achieving bidirectional enhancement of time-guided spatial evolution and space-constrained temporal continuity, improving the adaptability of features to complex vascular topologies. A spatially dominated asymmetric cross-attention mechanism is adopted, with spatial features as the main driver to aggregate temporal information through attention weights, avoiding redundant computations in traditional symmetric attention. The asymmetric design reduces computational complexity while maintaining high sensitivity to vascular edges. By utilizing the fused dynamic and static information in the spatiotemporal representation, the vascular boundary is accurately located. Through complementary features and bidirectional enhancement, the vascular edge segmentation accuracy is significantly better than existing methods.

[0106] As an optional embodiment, based on a preset dual-branch coding mechanism, complementary spatiotemporal features of the DSA sequence are extracted, including: The DSA sequence is input into a time encoder to capture the filling and / or diffusion process of contrast agent in the vascular network over time for hemodynamic modeling, and to extract time features of the DSA sequence that characterize the continuous dynamic properties of the blood vessels from the time branch. The minimum intensity projection image of the DSA sequence is input into the spatial encoder. During the downsampling stage, the receptive field of multi-level feature extraction is expanded frame by frame and the global spatial context is captured. The spatial features of the vascular structure prior enhanced by the DSA sequence are extracted in the spatial branch. The minimum intensity projection image is generated by taking the minimum intensity value at each pixel coordinate in the time dimension of the DSA sequence image.

[0107] As can be seen, the above optional embodiments overcome the limitations of a single information source through a dual-branch design. The dynamic information extracted by temporal encoding helps distinguish between real blood vessels (with continuous dynamic characteristics) and static artifacts, while the MinIP prior provided by spatial encoding serves as a spatial "navigation map," significantly improving the network model's ability to perceive vascular topology, especially the identification of small branches. By processing these two types of information independently during the encoding stage, premature mixing of heterogeneous features is avoided, laying a solid data foundation for more accurate feature fusion later.

[0108] As an optional embodiment, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, including: The temporal features are flattened in terms of spatial dimension to obtain a temporal feature sequence; The temporal feature sequence is mapped to a low-dimensional latent space based on the target embedding dimension through linear projection to obtain a low-dimensional feature sequence of temporal features; The temporal dynamics of the low-dimensional feature sequence are captured through linear computation using a selective spatial state model, and the temporal dynamic features are output. The cyclical process of the linear computation is represented as follows: ; ; in, This represents the hidden state at time t. This represents the low-dimensional feature input at time t in the low-dimensional feature sequence. The output value is obtained by linearly combining the hidden state at time t with the low-dimensional feature input. A, E, C, and D are learnable parameters of the selective spatial state model.

[0109] As can be seen, through the above optional embodiments, the local continuity of temporal features is preserved by spatial flattening and low-dimensional projection; combined with selective modeling, the dynamic evolution law of blood vessels is accurately extracted; the linear complexity design improves the efficiency of the model in processing ultra-long sequences, meets the real-time clinical needs, and the selective mechanism dynamically filters irrelevant background information, which is especially suitable for multiple clinical scenarios with differences in DSA data from different patients and hospitals.

[0110] As an optional embodiment, the complementary spatiotemporal features are subjected to temporal series modeling based on a preset selective spatial state model to obtain spatiotemporal dynamic features, which further includes: The spatial features are tokenized and embedded into the target embedding dimension to obtain spatial dynamic features. The tokenization and embedding process is represented as follows: ; ; in, Indicates spatial characteristics, This represents the dimensional structure for adjusting spatial features. Represents spatial feature sequences, In the expression, the tensor belongs to the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. Represents a linear transformation matrix. Indicates the target embedding dimension. It represents the dynamic characteristics of space.

[0111] As can be seen, through the above optional embodiments, spatial features are serialized and recombined by adjusting the dimensional structure and serializing the features. This not only preserves the spatial location information of the features (i.e., the combination of height h and width w), but also provides a structurally adapted feature representation for subsequent linear transformations. By dimensional alignment and feature semantic enhancement, the channel dimension of the spatial features is unified with the embedding dimension required for subsequent temporal modeling, solving the problem of feature dimension mismatch and improving the model's compatibility in processing spatiotemporal features. By capturing the correlation between channels through linear combination, the feature's ability to express spatial information is enhanced, making the features more in line with the semantic requirements of downstream tasks. This realizes the transformation of spatial features into dynamic features, and optimizes feature adaptability, semantic expressiveness, and spatiotemporal modeling synergy, laying a feature foundation for improving the performance of spatiotemporal tasks.

[0112] As an optional embodiment, the preset bidirectional boot enhancement mechanism is specifically as follows: The first guidance sets the temporal dynamic features as temporal queries and the spatial dynamic features as spatial keys and spatial values, constructs the spatial structure's constraint on temporal evolution, and outputs the temporally modulated spatial modulation features. The first guidance is a temporally guided spatial optimization process. The second guidance sets the spatial dynamic features as temporal queries and the temporal dynamic features as spatial keys and spatial values, constructs the constraints of temporal information on spatial structure enhancement, and outputs temporal modulation features enhanced by spatial structure. The second guidance is a space-dominated temporal enhancement process. The expression for the first guiding term is: ; ; ; ; In the formula, Represents temporal dynamic characteristics, The time-series query transformation matrix represents the learnable parameters. Represents the time-series query matrix. Represents spatial dynamic characteristics, The space key transformation matrix represents the learnable parameters. Represents the space bond matrix, This represents the spatial value transformation matrix of the learnable parameters. Represents a spatial value matrix, This represents the similarity calculation of the dot product attention between the temporal query and the spatial key. This indicates the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation; The expression for the second guideline is: ; ; ; ; In the formula, The spatial query transformation matrix represents the learnable parameters. Represents a spatial query matrix. The time-series key transformation matrix represents the learnable parameters. Represents the time-series value transformation matrix. Represents the time-series key matrix. Represents the time series value matrix. This represents the similarity calculation of the dot product attention between the spatial query and the temporal key. This indicates the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement.

[0113] As can be seen, through the above optional embodiments, by using temporal dynamic features as queries and spatial dynamic features as keys and values, an attention mechanism is used to construct constraints on the temporal evolution of spatial structure. Dynamic constraints on spatial changes suppress irrelevant background noise and improve the adaptability of spatial structure to dynamic processes. By using spatial dynamic features as queries and temporal dynamic features as keys and values, an attention mechanism is used to construct enhanced constraints on spatial structure of temporal information. Structured enhancement of temporal representation eliminates redundant temporal information, improves the robustness of temporal modeling to complex structures, and bidirectional guidance and collaboration form a bidirectional calibration mechanism for spatiotemporal features, avoiding deviations caused by unidirectional modeling, adapting to spatiotemporal coupling modes in different scenarios, and enhancing the model's adaptability to complex data.

[0114] As an optional implementation, a space-dominated asymmetric cross-attention mechanism is used, specifically: The temporal modulation feature in the aforementioned temporal control features is set as a query from space to time based on the spatial structure as the dominant factor; Set the spatial modulation feature in the time-space control feature as a key and value from time to space; Calculate the matching degree matrix between queries in the spatial-temporal direction and keys in the temporal-spatial direction, and solve for the attention weights in the matching degree matrix; Based on the weighted summation of the attention weights and the values ​​from time to space, the target spatiotemporal features of the cross-attention fusion spatiotemporal information are output. The spatiotemporal features of the target are reconstructed according to the dimensions of the DSA sequence, and the spatiotemporal representation of the target is output. The calculation process for the spatiotemporal representation of the target is expressed as follows: ; ; ; ; ; In the formula, A query matrix representing the spatial to temporal direction. The bond matrix represents the time-to-space direction. A matrix representing the values ​​from time to space. This represents the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement. This represents the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation. The learnable linear transformation matrix representing the query direction. The learnable linear transformation matrix representing the bond direction. The learnable linear transformation matrix representing the value direction. This represents the matching degree matrix between queries and key dot products that represent cross-interest in time and space. The attention weights are represented by scaling factors and normalization, and F represents the spatiotemporal features of the target. Represents feature reconstruction, Represents the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. This represents the spatiotemporal characterization of the target.

[0115] As can be seen, through the above optional embodiments, and by utilizing the "bidirectional and asymmetric" design concept of the BTSA module in the BTSAnet architecture, a highly efficient spatiotemporal feature fusion is achieved. Compared to the self-attention mechanism in existing technologies that involves symmetrical interaction among all features, this embodiment, by granting the spatial structure dominance (as a Query) and establishing bidirectional, clearly defined interaction paths, achieves precise control over the information flow. This ensures that the fused features maintain the integrity of the vascular topology while incorporating discriminative hemodynamic information, thereby effectively improving the accuracy of vascular segmentation while maintaining the integrity of the spatial structure, particularly excelling in suppressing background noise and delineating low-contrast small vessels.

[0116] As an optional embodiment, the spatiotemporal representation of the target is decoded to obtain the cerebral blood vessel segmentation result of the DSA sequence, including... The first decoding process involves doubling the spatial size of the input target spatiotemporal representation through transposed convolution and halving the number of channels. The second decoding involves combining the spatial features extracted by the dual-branch coding mechanism with the target spatiotemporal representation from the first decoding through a skip connection mechanism across all upsampled channel dimensions. The third decoding process involves fusing the spliced ​​features into the residual convolutional block to output a target feature map of the same size as the original angiographic image, and a segmentation probability map corresponding to the target feature map output by the activation function. The residual convolutional block sequentially includes two 3×3 convolutional layers, batch normalization, and an activation function.

[0117] As can be seen, through the above optional embodiments, the decoder outputs the vessel segmentation results, ensuring that the fusion features rich in high-level spatiotemporal semantics from the BTSA module can be effectively integrated with the multi-level spatial details provided by the spatial encoder. This process gradually restores the spatial resolution and significantly optimizes the accurate segmentation of vessel boundaries and the continuity of topological structure.

[0118] As an optional embodiment, the system is also used to perform the following steps before acquiring the raw contrast-enhanced images and preprocessing them to generate the DSA sequence: The target loss function for cerebral vessel segmentation is constructed using BCE loss and Dice loss, specifically expressed as follows: ; ; ; in, This represents the BCE loss, where N represents the total number of pixels. This represents the true label of the i-th sample. This represents the predicted probability of the i-th sample. This indicates Dice's loss. Indicates the true label, This represents the probability predicted by the model. Represents the smoothing factor. This represents the category index of tissues / structures in cerebral blood vessels, where C represents the number of categories of tissues / structures in cerebral blood vessels.

[0119] Specifically, in order to optimize the performance of the blood vessel segmentation model, this embodiment uses a combination of Binary Cross-Entropy Loss (BCE) and Dice loss as the objective function. BCE loss is beneficial for pixel-level classification, while Dice loss can effectively alleviate the class imbalance problem between blood vessels and background. The combination of the two can balance boundary accuracy and region consistency.

[0120] Example 3 Please see Figure 3 , Figure 3 This is another bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system disclosed in the embodiments of the present invention. Figure 3 The described bidirectional asymmetric DSA sequence cerebral vessel segmentation system is applied in a data processing system / data processing device / data processing server (wherein, the server includes a local processing server or a cloud processing server). For example... Figure 3 As shown, this bidirectional asymmetric DSA sequence cerebral vessel segmentation system may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method described in Embodiment 1.

[0121] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to perform the steps of the bidirectional asymmetric DSA sequence cerebral vessel segmentation method described in Embodiment 1.

[0122] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the bidirectional asymmetric DSA sequence cerebral vessel segmentation method described in Embodiment 1.

[0123] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0124] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0125] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0126] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0131] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0132] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0133] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0134] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0135] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0136] Finally, it should be noted that the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method and system disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bidirectional asymmetric DSA sequence method for cerebral blood vessel segmentation, characterized in that, The method includes: Acquire raw contrast images and preprocess them to generate DSA sequences; Based on a preset dual-branch coding mechanism, complementary spatiotemporal features of the DSA sequence are extracted. These complementary spatiotemporal features include mutually independent and heterogeneous temporal and spatial features. For the complementary spatiotemporal features, temporal modeling is performed based on a preset selective spatial state model to obtain spatiotemporal dynamic features, which include temporal dynamic features and spatial dynamic features of the same dimension. Based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, spatiotemporal control features are obtained. Based on the space-dominated asymmetric cross-attention mechanism, the spatiotemporal control features are fused by cross-attention to generate a target spatiotemporal representation. The spatiotemporal representation of the target is decoded to obtain the cerebral blood vessel segmentation result of the DSA sequence.

2. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, Based on a preset dual-branch coding mechanism, complementary spatiotemporal features of the DSA sequence are extracted, including: The DSA sequence is input into a time encoder to capture the filling and / or diffusion process of contrast agent in the vascular network over time for hemodynamic modeling, and to extract time features of the DSA sequence that characterize the continuous dynamic properties of the blood vessels from the time branch. The minimum intensity projection image of the DSA sequence is input into the spatial encoder. During the downsampling stage, the receptive field of multi-level feature extraction is expanded frame by frame and the global spatial context is captured. The spatial features of the vascular structure prior enhanced by the DSA sequence are extracted in the spatial branch. The minimum intensity projection image is generated by taking the minimum intensity value at each pixel coordinate in the time dimension of the DSA sequence image.

3. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, For the complementary spatiotemporal features, temporal series modeling is performed based on a preset selective spatial state model to obtain spatiotemporal dynamic features, including: The temporal features are flattened in terms of spatial dimension to obtain a temporal feature sequence; The temporal feature sequence is mapped to a low-dimensional latent space based on the target embedding dimension through linear projection to obtain a low-dimensional feature sequence of temporal features; The temporal dynamics of the low-dimensional feature sequence are captured through linear computation using a selective spatial state model, and the temporal dynamic features are output. The cyclical process of the linear computation is represented as follows: ； ； in, This represents the hidden state at time t. This represents the low-dimensional feature input at time t in the low-dimensional feature sequence. The output value is obtained by linearly combining the hidden state at time t with the low-dimensional feature input. A, E, C, and D are learnable parameters of the selective spatial state model.

4. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 3, characterized in that, For the complementary spatiotemporal features, time-series modeling is performed based on a preset selective spatial state model to obtain spatiotemporal dynamic features, which also includes: The spatial features are tokenized and embedded into the target embedding dimension to obtain spatial dynamic features. The tokenization and embedding process is represented as follows: ； ； in, Indicates spatial characteristics, This represents the dimensional structure for adjusting spatial features. Represents spatial feature sequences, In the expression, the tensor belongs to the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. Represents a linear transformation matrix. Indicates the target embedding dimension. It represents the dynamic characteristics of space.

5. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, The pre-defined two-way guidance enhancement mechanism is as follows: The first guidance sets the temporal dynamic features as temporal queries and the spatial dynamic features as spatial keys and spatial values, constructs the spatial structure's constraint on temporal evolution, and outputs the temporally modulated spatial modulation features. The first guidance is a temporally guided spatial optimization process. The second guidance sets the spatial dynamic features as temporal queries and the temporal dynamic features as spatial keys and spatial values, constructs the constraints of temporal information on spatial structure enhancement, and outputs temporal modulation features enhanced by spatial structure. The second guidance is a space-dominated temporal enhancement process. The expression for the first guiding term is: ； ； ； ； In the formula, Represents temporal dynamic characteristics, The time-series query transformation matrix represents the learnable parameters. Represents the time-series query matrix. Represents spatial dynamic characteristics, The space key transformation matrix represents the learnable parameters. Represents the space bond matrix, This represents the spatial value transformation matrix of the learnable parameters. Represents a spatial value matrix, This represents the similarity calculation of the dot product attention between the temporal query and the spatial key. This indicates the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation; The expression for the second guideline is: ； ； ； ； In the formula, The spatial query transformation matrix represents the learnable parameters. Represents a spatial query matrix. The time-series key transformation matrix represents the learnable parameters. Represents the time-series value transformation matrix. Represents the time-series key matrix. Represents the time series value matrix. This represents the similarity calculation of the dot product attention between the spatial query and the temporal key. This indicates the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement.

6. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, The space-dominated asymmetric cross-attention mechanism is as follows: The temporal modulation feature in the aforementioned temporal control features is set as a query from space to time based on the spatial structure as the dominant factor; Set the spatial modulation feature in the time-space control feature as a key and value from time to space; Calculate the matching degree matrix between queries in the spatial-temporal direction and keys in the temporal-spatial direction, and solve for the attention weights in the matching degree matrix; Based on the weighted summation of the attention weights and the values ​​from time to space, the target spatiotemporal features of the cross-attention fusion spatiotemporal information are output. The spatiotemporal features of the target are reconstructed according to the dimensions of the DSA sequence, and the spatiotemporal representation of the target is output. The calculation process for the spatiotemporal representation of the target is expressed as follows: ; ； ； ； ； In the formula, A query matrix representing the spatial to temporal direction. The bond matrix represents the time-to-space direction. A matrix representing the values ​​from time to space. This represents the temporal modulation characteristics from the spatial to the temporal direction after spatial structure enhancement. This represents the spatial modulation characteristics from the temporal to the spatial direction after temporal modulation. The learnable linear transformation matrix representing the query direction. The learnable linear transformation matrix representing the bond direction. The learnable linear transformation matrix representing the value direction. This represents the matching degree matrix between queries and key dot products that represent cross-interest in time and space. The attention weights are represented by scaling factors and normalization, and F represents the spatiotemporal features of the target. Represents feature reconstruction, Represents the set of real numbers. Indicates batch size, Indicates the height of the output feature, Indicates the width of the output feature. This represents the number of channels in the output feature. This represents the spatiotemporal characterization of the target.

7. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, Decoding the spatiotemporal representation of the target yields the cerebral blood vessel segmentation results of the DSA sequence, including... The first decoding process involves doubling the spatial size of the input target spatiotemporal representation through transposed convolution and halving the number of channels. The second decoding involves combining the spatial features extracted by the dual-branch coding mechanism with the target spatiotemporal representation from the first decoding through a skip connection mechanism across all upsampled channel dimensions. The third decoding process involves fusing the spliced ​​features into the residual convolutional block to output a target feature map of the same size as the original angiographic image, and a segmentation probability map corresponding to the target feature map output by the activation function. The residual convolutional block sequentially includes two 3×3 convolutional layers, batch normalization, and an activation function.

8. The bidirectional asymmetric DSA sequence cerebral vessel segmentation method according to claim 1, characterized in that, Before the acquisition of raw contrast images and preprocessing to generate DSA sequences, the process includes: The target loss function for cerebral vessel segmentation is constructed using BCE loss and Dice loss, specifically expressed as follows: ； ； ； in, This represents the BCE loss, where N represents the total number of pixels. This represents the true label of the i-th sample. This represents the predicted probability of the i-th sample. This indicates Dice's loss. Indicates the true label, This represents the probability predicted by the model. Represents the smoothing factor. This represents the category index of tissues / structures in cerebral blood vessels, where C represents the number of categories of tissues / structures in cerebral blood vessels.

9. A bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system, characterized in that, The system includes: The data acquisition module is used to acquire raw contrast images and preprocess them to generate DSA sequences; A bidirectional encoding module is used to extract complementary spatiotemporal features of the DSA sequence based on a preset bi-branch encoding mechanism. The complementary spatiotemporal features include mutually independent and heterogeneous temporal features and spatial features. The temporal modeling module is used to perform temporal modeling on the complementary spatiotemporal features based on a preset selective spatial state model to obtain spatiotemporal dynamic features, wherein the spatiotemporal dynamic features include temporal dynamic features and spatial dynamic features of the same dimension. The bidirectional asymmetric fusion module is used to obtain spatiotemporal control features based on the spatiotemporal dynamic features and the preset bidirectional guidance enhancement mechanism, and to perform cross-attention fusion on the spatiotemporal control features based on the space-dominated asymmetric cross-attention mechanism to generate a target spatiotemporal representation. The decoding module is used to decode the spatiotemporal representation of the target to obtain the cerebral blood vessel segmentation result of the DSA sequence.

10. A bidirectional asymmetric DSA sequence cerebral blood vessel segmentation system, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the bidirectional asymmetric DSA sequence cerebral blood vessel segmentation method as described in any one of claims 1-8.

Images