A method and device for positioning a left circumflex branch on a left anterior descending branch of a cardiovascular, and an electronic device

Abstract

The present disclosure provides a method and device for positioning the left circumflex branch on the left anterior descending branch of the cardiovascular system and an electronic device. By acquiring a sequence of continuous cross-sectional images, separating each blood vessel branch on the left anterior descending branch, and combining the spatial structure information and image feature information of the blood vessel branch for fusion determination, the automatic positioning of the left circumflex branch position is realized. Compared with the existing method which mainly relies on single image information or manual experience, the branch recognition accuracy can be effectively improved, the adaptability to noise, occlusion and image quality fluctuation scenarios can be enhanced, the branch attribute judgment ambiguity can be reduced, and thus the accuracy, continuity and auxiliary analysis efficiency of clinical operation can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of medical image processing technology, and more specifically, to a method, apparatus, and electronic device for locating the superior left circumflex branch of the left anterior descending artery of a cardiovascular system. Background Technology

[0002] In coronary interventional procedures, intravascular imaging techniques such as optical coherence tomography (OCT) and intravascular ultrasound (IVUS) can acquire continuous cross-sectional images of the blood vessel interior, providing crucial information for physicians to observe vessel morphology, plaque distribution, and branch openings. Analyzing the numerous continuous cross-sectional images generated during the retraction process can assist physicians in determining lesion location, measuring lesion length, and identifying the opening locations of important branch vessels, thereby supporting stent implantation, bifurcation lesion assessment, and intraoperative pathway planning.

[0003] However, most current intravascular imaging analysis methods still rely on physicians observing consecutive images frame by frame and manually determining the location and attributes of branches. Because the left anterior descending artery (LAD) has a large number of branches, different branches appear similar in local cross-sectional images. Furthermore, images can be affected by noise, artifacts, occlusion, changes in imaging angle, and fluctuations in local image quality. This leads to difficulties in localization, inaccurate judgment, or low recognition efficiency for physicians during actual image interpretation. Especially when identifying the left circumflex branch (LCX) within the LAD, relying solely on a single frame or local visual experience often makes it difficult to accurately distinguish the target branch from other similar branches, thus affecting subsequent analysis and procedures. Summary of the Invention

[0004] This disclosure provides at least one method, apparatus, and electronic device for locating the left circumflex branch on the left anterior descending artery of a cardiovascular system. By acquiring a continuous sequence of cross-sectional images, the various vascular branches on the left anterior descending artery segment are separated, and the spatial structural information and image feature information of the vascular branches are fused together for determination, thereby achieving automatic location of the left circumflex branch. Compared with existing methods that mainly rely on single image information or human experience for judgment, this method can effectively improve the accuracy of branch recognition, enhance adaptability to noise, occlusion, and image quality fluctuations, reduce ambiguity in branch attribute judgment, and thus improve the accuracy, continuity, and auxiliary analysis efficiency of clinical operations.

[0005] This disclosure provides a method for locating the superior left circumflex branch of the left anterior descending artery of a cardiovascular system, including: A continuous cross-sectional image sequence of the target vascular segment is obtained, and multiple vascular branch image sequences on the left anterior descending artery segment are separated based on the continuous cross-sectional image sequence, while the spatial structure information corresponding to each vascular branch is determined. For each vascular branch, spatial feature information of the vascular branch is extracted based on the corresponding spatial structure information, and image feature information of the vascular branch is extracted based on the corresponding branch image sequence; The spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are fused to obtain the fused features of the corresponding blood vessel branches; Based on the fusion features, each vascular branch is classified and determined to determine whether each vascular branch is a left circumflex branch, and the position of the left circumflex branch is output according to the determination result.

[0006] In one optional implementation, a continuous cross-sectional image sequence of the target vessel segment is acquired, and multiple vessel branch image sequences on the left anterior descending artery segment are separated based on the continuous cross-sectional image sequence. Simultaneously, the spatial structure information corresponding to each vessel branch is determined, specifically including: Intravascular imaging was performed on the left anterior descending artery segment to obtain cross-sectional images continuously distributed along the vessel axis. The vascular lumen region and branch region are identified in each frame of cross-sectional image to obtain the branch contour information in each frame; The branch contour information in different frames is correlated, and multiple branch contours belonging to the same blood vessel branch are combined to form a corresponding blood vessel branch image sequence. The spatial structure information of the corresponding vascular branch is determined based on the vascular branch image sequence. The spatial structure information includes at least one of the following: branch location, branch cross-sectional area, and branch length.

[0007] In one optional implementation, spatial feature information of blood vessel branches is extracted based on the corresponding spatial structure information, specifically including: The spatial structural information representing the branch location, branch cross-sectional area, and branch length is input into the spatial feature extraction network; During training, the spatial feature extraction network performs similarity-bringing constraints on spatial features corresponding to the same type of vascular branches, performs similarity-difference constraints on spatial features corresponding to different types of vascular branches, and applies redundancy removal constraints to spatial features from different sources before and after fusion. The spatial feature extraction network is used to perform convolutional feature extraction and aggregation on the spatial structure information to obtain the first spatial feature. The first spatial feature is subjected to dimensionality reduction representation processing to obtain the second spatial feature; The first spatial feature and the second spatial feature are fused together to obtain the spatial feature information of blood vessel branches.

[0008] In one optional implementation, the first spatial feature is subjected to dimensionality reduction representation processing to obtain the second spatial feature, specifically including: During training, a feature distribution is constructed based on the spatial features extracted at the current training stage; Based on the feature distribution, a low-dimensional projection relationship matching the current training phase is determined; The first spatial feature is reduced in dimension using the low-dimensional projection relationship to obtain the second spatial feature.

[0009] In one optional implementation, the first spatial feature and the second spatial feature are fused to obtain vascular branch spatial feature information, specifically including: Perform a mapping transformation on the second spatial feature to make the second spatial feature correspond to the first spatial feature in the feature representation dimension; The transformed second spatial features are used as supplementary information and residually fused with the first spatial features to obtain the spatial feature information of the blood vessel branches.

[0010] In one optional implementation, vascular branch image feature information is extracted based on the corresponding branch image sequence, specifically including: Input multiple frames of images of the corresponding blood vessel branches into a 3D image feature extraction network; During the training process, the three-dimensional image feature extraction network labels each vascular branch on the left anterior descending artery to obtain labels for the left circumflex branch and non-left circumflex branch. The labels are used to supervise the training of the three-dimensional image feature extraction network, and the vascular branch image feature information is extracted from the feature output layer of the three-dimensional image feature extraction network. Temporal correlation features and spatial texture features are extracted from the multi-frame images using 3D convolution processing; The features processed by 3D convolution are compressed in dimension and extracted by 2D convolution to obtain deep image representations. The deep image representation is converged and vectorized to obtain the vascular branch image feature information.

[0011] In one optional implementation, the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch are fused to obtain the fused features of the corresponding blood vessel branch, specifically including: The spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are subjected to dimensionality-upgrading processing to construct a fusion matrix representing the interaction relationship between the two types of features; The fusion matrix is ​​convolved to enhance the expression of the combination relationship between adjacent features; The fusion matrix after convolution processing is adaptively weighted based on the vascular branch image feature information to enhance key fusion features related to the determination of the left circumflex branch; The weighted fusion matrix is ​​downsampled and flattened to obtain the fusion representation vector. The fused representation vector is concatenated with the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused feature.

[0012] In one optional implementation, each vascular branch is classified and determined based on the fusion features to obtain a determination result of whether each vascular branch is a left circumflex branch, and the position of the left circumflex branch is output according to the determination result, specifically including: The fused features are input into the classification network; The output vascular branch category results are processed through at least two fully connected layers, and it is determined whether the corresponding vascular branch is a left circumflex branch based on the category results; Based on the determination results of whether each vascular branch is a left circumflex branch, a corresponding classification label sequence is generated; The classification label sequence is mapped to the branch positions in the left anterior descending artery segment to output the position of the left circumflex artery branch.

[0013] This disclosure also provides a positioning device for the superior left circumflex branch of the left anterior descending artery, comprising: The spatial information acquisition module is used to acquire a continuous cross-sectional image sequence of the target vascular segment, separate multiple vascular branch image sequences on the left anterior descending artery segment based on the continuous cross-sectional image sequence, and determine the spatial structure information corresponding to each vascular branch. The image feature acquisition module is used to extract spatial feature information of blood vessel branches based on the corresponding spatial structure information for each blood vessel branch, and to extract image feature information of blood vessel branches based on the corresponding branch image sequence. The feature fusion module is used to fuse the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused features of the corresponding blood vessel branch. The location discrimination module is used to classify and determine each vascular branch according to the fusion features, obtain the determination result of whether each vascular branch is a left circumflex branch, and output the location of the left circumflex branch according to the determination result.

[0014] This disclosure also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform steps in any of the above-described methods for locating the left anterior descending artery superior to the left circumflex artery, or any possible implementation of the above-described methods for locating the left anterior descending artery superior to the left circumflex artery.

[0015] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs steps in any of the possible implementations of the above-described method for locating the left superior circumflex branch of the left anterior descending artery or the above-described method for locating the left superior circumflex branch of the left anterior descending artery.

[0016] This disclosure also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of any possible implementation of the above-mentioned method for locating the superior left circumflex branch of the left anterior descending artery or the above-mentioned method for locating the superior left circumflex branch of the left anterior descending artery.

[0017] This disclosure provides a method, apparatus, and electronic device for locating the left circumflex branch on the left anterior descending artery of a cardiovascular system. By acquiring a continuous sequence of cross-sectional images, it separates the vascular branches on the left anterior descending artery segment and combines the spatial structural information and image feature information of the vascular branches for fusion determination, thus achieving automatic location of the left circumflex branch. Compared to existing methods that mainly rely on single image information or manual experience, this method effectively improves branch recognition accuracy, enhances adaptability to noise, occlusion, and image quality fluctuations, reduces ambiguity in branch attribute judgment, and thereby improves the accuracy, continuity, and auxiliary analysis efficiency of clinical operations.

[0018] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0020] Figure 1 A flowchart is shown below illustrating a method for locating the superior left circumflex branch of the left anterior descending artery, as provided in an embodiment of this disclosure. Figure 2 A schematic flowchart of a spatial feature network (SFN) provided in an embodiment of this disclosure is shown. Figure 3 A schematic diagram of the structure of a spatial feature network provided in an embodiment of this disclosure is shown; Figure 4This diagram illustrates a structural schematic for calculating a PCA projection matrix according to an embodiment of the present disclosure. Figure 5 This diagram illustrates the structure of a three-dimensional target image feature extraction network (IFN) provided in an embodiment of the present disclosure. Figure 6 A flowchart illustrating a network architecture SIFN that integrates SFI and IFI provided in an embodiment of this disclosure is shown. Figure 7 This diagram illustrates a feature fusion and classification structure provided by an embodiment of the present disclosure. Figure 8 A schematic diagram of a positioning device for the superior left circumflex branch of the left anterior descending artery provided in an embodiment of this disclosure is shown. Figure 9 A schematic diagram of an electronic device provided in an embodiment of the present disclosure is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0024] Research has revealed that current cardiovascular imaging analysis methods largely rely on physicians observing consecutive images frame by frame and manually determining branch locations and attributes. Due to the large number of branches on the left anterior descending artery (LAD), different branches appear similar in local cross-sectional images. Furthermore, images can be affected by noise, artifacts, occlusion, changes in imaging angle, and fluctuations in local image quality. This leads to difficulties in localization, inaccurate judgment, and low recognition efficiency for physicians during image interpretation. Particularly when identifying the left circumflex branch (LCX) on the LAD, relying solely on a single frame or local visual experience often makes it difficult to accurately distinguish the target branch from other similar branches, thus impacting subsequent analysis and procedures.

[0025] Based on the above research, this disclosure provides a method, device, and electronic device for locating the left circumflex branch on the left anterior descending artery of a cardiovascular system. By acquiring a continuous sequence of cross-sectional images, the various vascular branches on the left anterior descending artery segment are separated, and the spatial structural information and image feature information of the vascular branches are fused and determined, thereby achieving automatic localization of the left circumflex branch location. Compared with existing methods that mainly rely on single image information or human experience for judgment, this method can effectively improve the accuracy of branch recognition, enhance adaptability to noise, occlusion, and image quality fluctuations, reduce ambiguity in branch attribute judgment, and thus improve the accuracy, continuity, and auxiliary analysis efficiency of clinical operations.

[0026] To facilitate understanding of this embodiment, a detailed description of the method for locating the superior left circumflex branch of the left anterior descending artery (LAD) disclosed in this disclosure embodiment is provided first. The execution entity of the method for locating the superior left circumflex branch of the LAD provided in this disclosure embodiment is generally a computer device with a certain computing capability. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc. In some possible implementations, the method for locating the superior left circumflex branch of the LAD can be implemented by a processor calling computer-readable instructions stored in memory.

[0027] See Figure 1 As shown, Figure 1 This is a flowchart of a method for locating the superior left circumflex branch of the left anterior descending artery, as provided in an embodiment of this disclosure. The method includes steps S101 to S104, wherein: S101. Obtain a continuous cross-sectional image sequence of the target vascular segment, separate multiple vascular branch image sequences on the left anterior descending branch vascular segment based on the continuous cross-sectional image sequence, and determine the spatial structure information corresponding to each vascular branch.

[0028] The method for locating the branch position of the left circumflex branch on the left anterior descending artery in this embodiment first obtains a continuous cross-sectional image sequence of the target vessel segment, and then separates multiple vessel branch image sequences on the left anterior descending artery segment based on the continuous cross-sectional image sequence, while determining the spatial structure information corresponding to each vessel branch.

[0029] Here, the target vessel segment is preferably the left anterior descending artery segment, and the continuous cross-sectional image sequence is preferably acquired using an intravascular imaging device, which can be an optical coherence tomography (OCT) device or an intravascular ultrasound imaging device. By continuously scanning the target vessel segment, multiple frames of sequentially arranged cross-sectional images can be obtained along the vessel's extension direction, thereby forming a continuous cross-sectional image sequence characterizing the changes in the internal structure of the vessel.

[0030] In practice, to ensure that all branch structures within the target vessel segment can be captured relatively completely, the sampling of continuous cross-sectional image sequences should cover the entire detection range of the target vessel segment. Preferably, cross-sectional images are acquired continuously along the vessel axis at preset sampling intervals to maintain stable spatial continuity between adjacent images.

[0031] The sampling interval can be set according to the actual imaging accuracy requirements, for example, it can be set to 0.05mm, 0.1mm or 0.2mm, so as to balance image resolution and processing efficiency.

[0032] Furthermore, after obtaining the continuous cross-sectional image sequence, branch recognition processing is performed on each frame of the cross-sectional image in the sequence to separate multiple vascular branch image sequences from the left anterior descending artery segment. Specifically, a vascular segmentation network can be used to process each frame of the continuous cross-sectional image sequence to identify the lumen region and the branch cross-sectional region, thereby obtaining the branch contour information in each frame of the image.

[0033] Here, branch contour information can characterize the branch's location range, morphological boundaries, and area features within the current cross-section. Since the same vascular branch usually appears consecutively in multiple adjacent cross-sectional images, the branch contours in multiple images can be associated and merged based on the spatial continuity, positional proximity, and contour change patterns between adjacent frames. This allows multiple branch contours belonging to the same vascular branch to be combined, thereby forming a branch image sequence corresponding to that vascular branch.

[0034] It should be understood that the above process is not limited to using a specific type of segmentation network. Other image processing models or algorithms capable of realizing lumen region identification, branch region identification, and cross-frame branch association can also be used. As long as it is possible to distinguish different branches on the left anterior descending artery segment from the trunk image information based on a continuous cross-sectional image sequence, and to establish corresponding multi-frame branch image sets, it can be applied to this embodiment.

[0035] It should be noted that, while separating the image sequences of each vascular branch, it is also necessary to further determine the spatial structural information corresponding to each vascular branch. The spatial structural information is used to characterize the spatial distribution and geometric structural attributes of the vascular branches in the left anterior descending artery segment, and preferably includes at least one of the following: branch location, branch cross-sectional area, and branch length.

[0036] In practical implementation, based on the established sequence of vascular branch images, the starting and ending positions of each vascular branch in the continuous cross-sectional image sequence, as well as its distribution range along the sequence direction, can be determined to obtain branch position and length information. Simultaneously, the cross-sectional area information of the vascular branch can be obtained based on the area calculation results of the corresponding branch contour in each frame. Furthermore, the area changes, contour changes, and center position changes in each frame of the branch image sequence can be statistically analyzed to form a more complete spatial structure description.

[0037] The branch position can be represented as the relative or absolute position of the vascular branch in the left anterior descending artery segment, used to characterize the location of the vascular branch in the entire target vascular segment; the branch cross-sectional area can be represented as the opening area or the area enclosed by the contour of the vascular branch on the corresponding cross-section, used to reflect the local geometric size of the branch; the branch length can be represented as the length of the range in which the same vascular branch appears continuously in a continuous cross-sectional image sequence, used to characterize the extent of the branch's extension in the scanning direction.

[0038] S102. For each vascular branch, extract the spatial feature information of the vascular branch according to the corresponding spatial structure information, and extract the image feature information of the vascular branch according to the corresponding branch image sequence.

[0039] In practice, the spatial structure information of the corresponding vascular branch is input into a spatial feature extraction network. This network encodes the structural attributes of the vascular branch and outputs corresponding spatial feature information. The spatial structure information preferably includes parameters reflecting the spatial attributes of the vascular branch, such as branch location, branch cross-sectional area, and branch length. After receiving these parameters, the spatial feature extraction network learns and abstracts the relationships between different parameters, thereby converting the original structural parameters into feature expressions suitable for subsequent classification and recognition. The resulting spatial feature information of the vascular branch, compared to directly using the original spatial structure parameters, better reflects the differences and similarities of each vascular branch at the spatial structure level.

[0040] Furthermore, the spatial feature extraction network preferably employs a multi-layer feature extraction structure to process the input spatial structure information layer by layer. Specifically, the branch positions, branch cross-sectional areas, and branch lengths can be first standardized and organized, and then fed into the network backbone structure as multi-channel inputs. Then, convolution processing, normalization processing, nonlinear activation processing, pooling processing, and residual feature extraction processing are used to abstract the spatial structure information of the vascular branches step by step, obtaining intermediate spatial feature representations. Finally, these intermediate spatial feature representations are globally aggregated to form global spatial features.

[0041] To further enhance the expressive power of spatial features, this embodiment also performs dimensionality reduction and fusion processing based on global spatial features. A spatial feature extraction network is used to extract and aggregate spatial structural information through convolutional features to obtain first spatial features. The first spatial features are then subjected to dimensionality reduction to obtain second spatial features. Finally, the first and second spatial features are fused to obtain vascular branch spatial feature information. During training, a feature distribution is constructed based on the spatial features extracted in the current training phase. A low-dimensional projection relationship matching the current training phase is determined based on the feature distribution. The first spatial features are then dimensionality-reduced using this low-dimensional projection relationship to obtain the second spatial features. A mapping transformation is performed on the second spatial features to make them correspond to the first spatial features in terms of feature representation dimension. The transformed second spatial features are then used as supplementary information and fused with the first spatial features using residual fusion to obtain vascular branch spatial feature information.

[0042] Specifically, a low-dimensional statistical representation can be constructed based on the global spatial features extracted during the current training phase, and the low-dimensional statistical representation can be used as supplementary information to the global spatial features. Then, the low-dimensional statistical representation and the global spatial features are matched in terms of expression dimension through mapping transformation, and the two are then fused to form the final spatial feature information of blood vessel branches.

[0043] In some implementations, to reduce information redundancy between spatial features from different sources and enhance the ability of spatial features to distinguish between different types of vascular branches, a feature constraint mechanism can be added during the training process of the spatial feature extraction network. The feature constraint mechanism may include: making spatial features of vascular branches of the same type tend to be closer to each other, making spatial features of vascular branches of different types tend to be farther apart, and applying redundancy removal constraints to spatial features from different sources before and after fusion.

[0044] On the other hand, for each vascular branch, extracting vascular branch image feature information based on the corresponding branch image sequence can include: inputting a multi-frame branch image sequence corresponding to the vascular branch into an image feature extraction network, whereby the image feature extraction network extracts features from the image content of the vascular branch and outputs the corresponding vascular branch image feature information. Since a vascular branch typically corresponds to multiple frames in a continuous cross-sectional image sequence, by jointly processing multiple frames, it is possible to simultaneously utilize the morphological change information and continuity information of the branch in different cross-sections. Therefore, the extracted vascular branch image feature information not only includes local texture and contour features in a single frame image but also structural change features across frames.

[0045] In specific implementation, the image feature extraction network preferably adopts a network structure that combines three-dimensional feature extraction and two-dimensional feature extraction. Specifically, multiple frames of images corresponding to a blood vessel branch can be input into the network as a three-dimensional image block. The correlation features between the multiple frames and the spatial texture features within the image are extracted through three-dimensional convolution processing. Subsequently, the features processed by three-dimensional convolution are dimensionally compressed to adapt them to subsequent two-dimensional convolution processing. Then, the compressed features are further extracted using a two-dimensional convolution network to obtain a deeper image semantic representation. Finally, the obtained deep image semantic representation is converged and vectorized to output the feature information of the blood vessel branch image.

[0046] Furthermore, the image feature extraction network can undergo supervised training during the training phase based on labels for both left and non-left circumflex branches. Specifically, each vascular branch on the left anterior descending artery can be pre-labeled, with the target left circumflex branch labeled as one category and other non-left circumflex branches labeled as another. The image feature extraction network is then trained using these labeling results. Through supervised training, the image feature extraction network learns image representation patterns that help distinguish between left and non-left circumflex branches, thereby forming highly discriminative vascular branch image feature information in the network's feature output layer.

[0047] Specifically, 3D convolution processing is used to extract temporal correlation features and spatial texture features from multiple frames of images; the features processed by 3D convolution are then subjected to dimensionality compression and 2D convolution extraction to obtain deep image representations; and the deep image representations are then converged and vectorized to obtain vascular branch image feature information.

[0048] It should be noted that in this embodiment, the extraction of spatial feature information and the extraction of image feature information of vascular branches can be performed in parallel or sequentially according to a preset order, and the present invention does not limit this. As long as two types of feature information representing the same vascular branch can be extracted from the spatial structure dimension and the image content dimension respectively, and used for subsequent fusion analysis and classification, they should all fall within the protection scope of the present invention.

[0049] S103. The spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch are fused to obtain the fused features of the corresponding blood vessel branch.

[0050] In practice, the two types of features are first aligned so that they have a corresponding feature representation basis in the subsequent joint modeling process. Then, a fusion matrix representing the interaction relationship between the two types of features is constructed and the fusion matrix is ​​enhanced. After that, the enhanced fusion matrix is ​​adaptively weighted based on image features. Then, the weighted fusion result is compressed and integrated to finally form the fusion feature of the corresponding blood vessel branch.

[0051] Furthermore, in the initial stage of the fusion process, the spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches can be subjected to dimensionality-upgrading processing in order to construct a fusion matrix that represents the pairwise correlation between the two types of features.

[0052] Here, the fusion matrix can be understood as an interactive feature representation that reflects the correspondence between different dimensions of spatial features and different dimensions of image features. Compared with simply concatenating or directly adding the two types of features, this method of constructing the fusion matrix can more fully describe the fine-grained coupling relationship between the two types of features, thereby enabling the subsequent network to learn richer cross-modal discrimination information.

[0053] After constructing the fusion matrix, convolutional enhancement processing can be applied to it. Specifically, convolutional layers can be used to extract local neighborhood features from the fusion matrix, further modeling the combination patterns between adjacent elements. Since different regions in the fusion matrix correspond to the interaction results between different spatial feature dimensions and image feature dimensions, convolutional processing can not only focus on the relationship between single feature pairs, but also further capture the joint variation law of multiple sets of feature relationships in a local range, thereby improving the ability of the fusion representation to describe the complex vascular branch attributes.

[0054] In this embodiment, to make the fusion result more prominently reflect the key information related to branch determination, adaptive weighting processing can be performed on the convolution-enhanced fusion matrix based on the vascular branch image feature information. Specifically, image feature information can be used as conditional guidance information to assign different importance weights to different regions in the fusion matrix, thereby enhancing the fusion features related to target branch identification and suppressing noise components or feature components irrelevant to the determination.

[0055] After adaptive weighting, the weighted fusion matrix can be downsampled and flattened to obtain a compressed fusion representation vector. The downsampling process can be achieved through multi-level convolutional compression, gradually reducing the feature scale of the fusion matrix while preserving the main discriminative information, thereby reducing the computational load of subsequent classification processing. The flattening process converts the compressed two-dimensional or multi-dimensional fusion result into a one-dimensional feature vector, enabling unified integration with the original vascular branch spatial feature information and vascular branch image feature information.

[0056] Furthermore, after obtaining the fused representation vector, it can be concatenated with the original spatial feature information and image feature information of the vascular branches to form the final fused feature. This fused feature contains, on the one hand, the joint representation obtained after deep interaction between spatial and image features, and on the other hand, retains the direct discriminative information from the original spatial and image features, thus giving the final fused feature a more complete expressive capability. For subsequent classification networks, this fused feature can provide both cross-modal coupled features and independent features for each modality, thus facilitating accurate differentiation between left-circular and non-left-circular branches.

[0057] It should be noted that the fusion process is not limited to a fixed combination of network layers. As long as it can jointly model the spatial feature information of blood vessel branches and the image feature information of blood vessel branches, and output a fusion feature that can simultaneously represent the coupling relationship between the two types of information, it is applicable to this embodiment. For example, in other embodiments, the number of convolutional enhancement layers, weighting method, compression method, or splicing order can be adjusted according to specific application requirements, and this invention does not limit this.

[0058] S104. Classify and determine each vascular branch according to the fusion features, obtain the determination result of whether each vascular branch is a left circumflex branch, and output the position of the left circumflex branch according to the determination result.

[0059] In practice, the fusion features corresponding to each vascular branch can be input into a classification network. The classification network then performs category discrimination on the fusion features and outputs the classification result for the corresponding vascular branch. The classification result is used to characterize whether the vascular branch belongs to a left circumflex branch or a non-left circumflex branch.

[0060] Preferably, the classification network may include a multi-layer fully connected processing structure, which maps and discriminates the fused features layer by layer to output the category information corresponding to each vascular branch. Since the fused features already contain the spatial structure representation of the vascular branch, the image content representation, and the coupling relationship between the two, the classification network can identify each vascular branch based on more sufficient discriminative criteria, thereby improving the accuracy of the left circumflex branch identification.

[0061] Furthermore, during the classification process, a category identification process can be performed separately for each vascular branch on the left anterior descending artery segment. That is, the fusion features corresponding to each vascular branch are sequentially input into the classification network to obtain the classification result for that vascular branch; or, the fusion features corresponding to multiple vascular branches are input into the classification network in batches to output the classification results for multiple vascular branches in batches.

[0062] In some implementations, the classification network outputs a binary classification result, directly indicating whether the target vascular branch belongs to the left circumflex branch or not; alternatively, it outputs judgment information characterizing the confidence level of different categories, and then determines the final category result according to preset judgment rules. For example, when the judgment value corresponding to the left circumflex branch meets preset conditions, the vascular branch is determined to be a left circumflex branch; otherwise, the vascular branch is determined to be a non-left circumflex branch.

[0063] After classifying and determining the branches of each blood vessel, a corresponding classification label sequence can be generated based on the determination results. This classification label sequence characterizes the category distribution of each branch on the left anterior descending artery segment. Branches identified as branches of the left circumflex artery correspond to target labels, while other branches not identified as branches of the left circumflex artery correspond to non-target labels. The classification label sequence can correspond one-to-one with multiple previously established image sequences of blood vessel branches, or it can correspond one-to-one with the spatial structure information of each blood vessel branch, thereby establishing a correlation between the classification results and specific blood vessel branches.

[0064] Furthermore, outputting the location of the left circumflex branch based on the determination result may include: filtering out the target vessel branch identified as a left circumflex branch from the classification results of all vessel branches, and reading the branch position information in the spatial structure information corresponding to the target vessel branch to determine the location of the left circumflex branch in the left anterior descending artery segment. The location can be expressed as the relative position of the left circumflex branch in the target vessel segment, or as the start frame position, end frame position, or center position of the left circumflex branch in a continuous cross-sectional image sequence. In this way, the physician can quickly confirm the corresponding region of the left circumflex branch in the continuous pull-back images by combining the output location result.

[0065] In specific applications, the output of the position of the left circumflex branch may also include mapping the position result of the left circumflex branch to a continuous cross-sectional image sequence, and marking the image frame interval, branch opening region or corresponding branch number where the left circumflex branch is located; or the position result may be output to a display terminal for doctors to observe, analyze and perform subsequent operations.

[0066] It should be noted that in some embodiments, if the classification result shows that only one vascular branch on the left anterior descending artery segment meets the criteria for determining a left circumflex branch, the position of that vascular branch can be directly output as the position of the left circumflex branch. If multiple candidate vascular branches meet the preset criteria, the target left circumflex branch can be further determined from the multiple candidate vascular branches by combining the judgment value, prior spatial location information, or other auxiliary rules. This invention does not limit this aspect.

[0067] The above scheme will now be described in conjunction with specific implementation methods.

[0068] See Figure 2 As shown, Figure 2 A schematic flowchart illustrating a spatial feature network (SFN) provided in an embodiment of this disclosure is shown.

[0069] Step 11: Acquire vascular images and establish branches.

[0070] The patient's LAD vascular segment is scanned using intravascular imaging OCT or IVUS equipment to acquire a continuous cross-sectional image sequence. The cross-sectional sampling interval can be 0.05mm, 0.1mm, or 0.2mm, etc., to ensure that all branches of the entire vascular segment are covered. The vascular segmentation network is used to segment the lumen and branch cross-sections of each frame in the continuous cross-sectional image sequence to obtain the branch contour of each frame. The contours of the same branch are combined to form a branch.

[0071] Step 12: Obtain spatial structure information and construct training labels.

[0072] Calculate the branch location, cross-sectional area, and length of each branch to form the spatial structure information of each branch. Label all branches on the LAD vascular segment as either LCX branches or non-LCX branches, such as... Figure 2 As shown on the left, this serves as a supervision label.

[0073] Step 13: Train the spatial feature network and train the 3D target image feature extraction network, and perform feature extraction simultaneously.

[0074] The training spatial feature network (SFN) takes the spatial structure information of branches and the spatial feature information (SFI) of blood vessel branches as input, with a dimension of 256. Specifically, firstly, a 1D CNN is used to extract hierarchical features from the spatial structure information such as branch position, branch cross-sectional area, and branch length, obtaining a dense intermediate feature representation. Subsequently, the features of the training set are subjected to dimensionality reduction processing using PCA to obtain low-dimensional principal component feature vectors. Then, a residual fusion strategy is used to linearly map and add the dense SFN features and PCA features to form the feature vector SFI. During training, an orthogonality constraint loss is introduced to reduce the redundancy between the dense SFN features and PCA features. At the same time, supervised learning is used to reduce the cosine distance of features of branches of the same category and increase the cosine distance of features of branches of different categories, enabling the network to effectively learn discriminative and complementary spatial feature information (SFI).

[0075] The training of the 3D target image feature extraction network (IFN) involves inputting a multi-frame branch image sequence, progressively extracting 3D and 2D information, and compressing it into a one-dimensional convolution to obtain image feature information (IFI) with a dimension of 256. Specifically, it jointly extracts image features of a blood vessel branch from multiple frames, effectively extracting 3D target features into the target's feature vector. The network's Conv3d-Conv2d-Conv1d structure first extracts 3D image features from the 3D target image, then further extracts 2D image features, and finally compresses them into a one-dimensional convolution to extract the feature vector.

[0076] Step 14, Feature Fusion and Classification (SIFN).

[0077] See Figure 3 As shown, Figure 3 A schematic diagram of the structure of a spatial feature network provided in an embodiment of this disclosure is shown.

[0078] like Figure 3The spatial feature network shown takes as input spatial structure information including branch location, branch cross-sectional area, and branch length. The network outputs spatial feature information (SFI) of blood vessel branches. Intermediate features of the spatial structure of blood vessel branches are extracted through one-dimensional convolution, batch normalization, ReLU activation, max pooling, and a residual network. Global average pooling is then used to obtain a globally dense feature vector, and PCA is combined to reduce the dimensionality of the features. Based on this, a residual fusion strategy is used to fuse the dimensionality-reduced features with the dense features, while orthogonality constraints are introduced to reduce feature redundancy. The final output is 256-dimensional spatial feature information (SFI) of blood vessel branches. This feature is dense, interpretable, and can effectively measure the spatial structural similarity of different branches. It can be used for applications such as classification, clustering, and anomaly detection, including: Step 131, Network Input.

[0079] The spatial structural information of vascular branches is as follows: ,

[0080] Where bs is the batch size. Let R be the branch location, branch cross-sectional area, and branch length. R is the real number field, indicating that each element in the tensor is a real number. Meaning: For each blood vessel branch in a batch, there are 3 input channels, and each channel contains a sequence of length 1.

[0081] Step 132, SFN backbone network.

[0082] The SFN backbone network adopts a 1D convolution + BN + ReLU + MaxPool + ResNet1D structure, and outputs an intermediate feature representation H:

[0083]

[0084] in, This represents the number of intermediate channels, with each channel having a feature length of 1.

[0085] Global features are obtained through global average pooling:

[0086] Step 133, SFN projection layer (SFL).

[0087] To align with the feature scale of PCA and facilitate residual fusion, a linear mapping layer (SFL) can be learned:

[0088] in, , These are the parameters for the fully connected layer. These are the global features output by the SFN backbone network.

[0089] Step 134: Dynamic PCA dimensionality reduction processing, such as... Figure 4 As shown, Figure 4 A schematic diagram of a structure for calculating a PCA projection matrix provided by an embodiment of this disclosure is shown: Global feature matrix extracted from training set data via FSN backbone network Construct the covariance matrix and perform eigenvalue decomposition. Select the first d=256 principal components to obtain the projection matrix. .

[0090] Dynamic characteristics explanation: Because the SFN backbone network is continuously updated during training, its extracted global features... This also changes accordingly, so the current features can be re-extracted and updated after each training epoch. This ensures that the PCA projection matrix and the SFN feature distribution are updated synchronously.

[0091] The low-dimensional principal components of each sample are represented as follows:

[0092] in, These are the global features output by the SFN backbone network.

[0093] Step 135, Residual Fusion Strategy.

[0094] Fusing SFN dense features with PCA feature residuals:

[0095] in, This is a learnable linear mapping matrix used to transform PCA features to match dense features. The final output feature vector is obtained as follows: .

[0096] Here, SFN dense features Playing a leading role, maintaining the integrity of the original spatial structure information; PCA features It provides low-dimensional supplementary information and fuses them through residuals, so that the final fused features retain the expressive power of dense features and enhance feature complementarity.

[0097] Unlike sparse coding vectors, most of the dimensions are non-zero and carry information, hence the name dense features. Dense features can make full use of the information in each dimension in cosine similarity calculation, so that the spatial structural differences of different blood vessel branches can be distinguished in detail.

[0098] Step 136, Orthogonality constraint loss.

[0099] To reduce feature redundancy and enhance complementarity, an intra-batch covariance orthogonal constraint is introduced:

[0100] in, , This represents the L2 normalized vector.

[0101] Orthogonality constraint loss function:

[0102] in, The Frobenius norm is used. This constraint ensures that dense features and PCA projected features are as geometrically orthogonal as possible, reducing redundancy and improving discriminability. During training, [the following is used:] The loss from the main task is jointly optimized with the loss from the main task based on weights.

[0103] Step 137: Comparative learning.

[0104] A contrastive learning network based on cosine similarity is used to train the network. Contrastive learning can directly optimize the geometric similarity of spatial features, and SFI is more in line with the discrimination requirements of the spatial structure of blood vessel branches.

[0105] Contrastive learning: All branch spatial feature information (SFI) is initially a 0 vector. The SFI of all branches is updated and calculated once in each epoch and stored in the feature set of all branches. During training of each batch size, the branch information is processed by the spatial structure network to output SFI. The feature cosine distance is calculated once with the SFI of each branch in the feature set. The smaller the feature cosine distance between branches of the same category (e.g., all are LCX branches or all are non-LCX branches), the better. The larger the feature cosine distance between branches of different categories, the better.

[0106] The loss function for contrastive learning is as follows:

[0107] Where: N is the total number of branches in the feature set. It is an anchor point feature. These are features of similar samples. These are different class of sample features, and m is the margin used to control the minimum interval between different classes of samples (in this example, m=0.2).

[0108] The training process for spatial feature networks includes: Step S1: Input the branch spatial information (location, cross-sectional area, length) and obtain the global features through the SFN backbone network. ; Step S2: The SFL network layer will transfer global features. Mapping to SFN dense features ; Step S3: Extract features from the current training set And obtain dynamic PCA Through global features and calculate ; Step S4: Residual fusion yields the final features. , It is a learnable linear mapping matrix; Step S5: Calculate the task loss for contrastive learning. + Orthogonality constraint loss function Update all learnable parameters (SFN backbone, SFL network layers, ... ).

[0109] The network structure of a 3D target image feature extraction network is as follows: Figure 5 As shown, the 3D target image feature extraction network IFN extracts image features of a blood vessel branch from multiple frames of images. The process includes: the first layer takes N frames of images for each branch as input, with input dimensions [bs, N, C, H, W], where C=1 is the number of image channels, H=256 is the image height, and W=256 is the image width. This network first extracts features (size [bs, N, C1, H, W]) from the 3D target (N frames corresponding to the blood vessel branch) using a 5-layer 3D convolutional network, where C1=64 is the number of feature channels. Then, a 3D-2D module compresses the feature dimensions (size [bs, N×C1]). The network uses a 2D convolutional network (1, H, W) to adapt to a 2D convolutional network. Features are then further extracted using a 2D convolutional network ResNet (size [bs, 256, 16, 16]). The image feature output layer (IFL) uses Avgpool (size [bs, 256, 1, 1]), squeezes to [bs, 256, 1], passes through Conv1d (size [bs, 256, 1]), and then squeezes again to [bs, 256] as the image feature information (IFI) for blood vessel branches. The network's last output layer determines whether a branch is an LCX.

[0110] During training, such as Figure 6As shown, the network is trained with supervised labels to classify LCX / non-LCX, and then the image feature information (IFI) of the branches can be obtained from the image feature output layer (IFL). The training objective of this stage is to extract the features of the vascular branch images. The role of the network's classification output layer is to output the classification results of LCX / non-LCX. By using supervised labels for training, the network as a whole becomes more distinguishable between the two types of vascular branch features, thus making the extracted features of the two types of vascular branches more obvious and discriminative.

[0111] The method SIFN, which fuses spatial feature information (SFI) of blood vessel branches and image feature information (IFI), is structured as follows: Figure 7 The SFIu shown is [bs,1,256,1]. The IFIu [bs,1,1,256] layer can fully fuse two types of features to form a fusion matrix SIFI [bs,1,256,256]. SIFI, after passing through a convolutional layer, can utilize neighborhood information to form even more effective features, which are then further processed by SIFI [bs,1,256,256]. The repeat(IFIu) layer [bs,1,256,256] enhances the attention mechanism of image features. Then, four convolutional layers with stride=2 are used to shrink the feature map (feature map size[bs,1,16,16]). The feature map can be reshaped to size[bs,256]. At this time, SFI, IFI, and SIFI are concatenated to form a new feature size[bs,768]. The purpose of this is to strengthen the fusion of SFI and IFI features. Then, two fully connected layers are used to output the classification result. Thus, the fusion of spatial structure features and image features to determine whether the branch is an LCX branch has been achieved.

[0112] The process of feature fusion and classification specifically includes: 1) The fusion of two types of features forms a two-dimensional matrix: SFI and IFI are upgraded to a second-order interaction matrix (SFI × IFI → 256×256) to capture high-order correlations and achieve high-dimensional modeling of the relationship between features; Spatial features of vascular branches Image feature information To construct a two-dimensional fusion matrix through the outer product, the original features are first increased in dimension: Spatial feature reshape:

[0113] Image feature reshape:

[0114]

[0115] in, The outer product is used to fuse spatial features and image features into a two-dimensional matrix; 2) Local convolution enhances neighborhood information: Convolutional layers are used to process the interaction matrix, modeling combination patterns between multiple pairs of features.

[0116] 3) Adaptive weighting of fusion results using image feature-based conditional attention: An IFI-based conditional attention mechanism is introduced to adaptively weight different elements of the fusion matrix, enhancing important features and suppressing irrelevant features. First, expand IFI to the same dimension:

[0117] Then perform element-wise multiplication:

[0118] This step essentially uses the patterns of image features to determine which parts of the fusion matrix are more important. All elements have "equal" weights, but through , As an inner product, dimensions with strong image features amplify the weights of corresponding regions, while weak dimensions suppress them. This becomes adaptive, because the IFI (Integral Function Index) differs for different input images, and the weights will change accordingly. 4) Progressive downsampling (4 convolutions with stride=2, equivalent to downsampling 256→16): The processed fusion matrix is ​​downsampled and flattened.

[0119] Reshape into a vector:

[0120] 5) Multi-source feature concatenation: Concatenate with the original SFI and IFI vectors to form the final fused feature vector (768 dimensions), concatenating the three types of features:

[0121] 6) Classification output:

[0122] in , , , These are the parameters for the fully connected layer.

[0123] In this embodiment, the feature fusion process is specifically as follows: SFI u [bs,1,256,1] IFI u The [bs,1,1,256] layer can fully fuse two types of features to form a fusion matrix SIFI [bs,1,256,256]. SIFI, after passing through a convolutional layer, can utilize neighborhood information to form even more effective features, which are then further processed by SIFI [bs,1,256,256]. repeat(IFI u The [bs,1,256,256] layer enhances the attention mechanism for image features. Then, four convolutional layers with stride=2 are used to shrink the feature map (feature map size[bs,1,16,16]), which can be reshaped to size[bs,256]. At this point, SFI, IFI, and SIFI are concatenated to form a new feature size[bs,768]. This aims to strengthen the fusion of SFI and IFI features. Finally, two fully connected layers output the classification result, thus achieving the fusion of spatial structure features and image features. Ultimately, it can determine whether a branch is an LCX branch.

[0124] This disclosure provides a method for locating the left circumflex branch of the left anterior descending artery (LAD) in a cardiovascular system. By acquiring a continuous sequence of cross-sectional images, separating the vascular branches along the LAD segment, and combining the spatial structural information and image feature information of the vascular branches for fusion determination, the method achieves automatic location of the left circumflex branch. Compared to existing methods that primarily rely on single image information or manual experience, this method effectively improves branch recognition accuracy, enhances adaptability to noise, occlusion, and image quality fluctuations, reduces ambiguity in branch attribute judgment, and thus improves the accuracy, continuity, and auxiliary analysis efficiency of clinical procedures.

[0125] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0126] Based on the same inventive concept, this disclosure also provides a positioning device for the left anterior descending artery and the left circumflex artery, corresponding to the positioning method for the left anterior descending artery and the left circumflex artery. Since the principle of the device in this disclosure for solving the problem is similar to the positioning method for the left anterior descending artery and the left circumflex artery described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0127] Please see Figure 8 , Figure 8 This is a schematic diagram of a positioning device for the superior left circumflex branch of the left anterior descending artery, provided as an embodiment of this disclosure. Figure 8As shown in the illustration, the positioning device 800 for the left anterior descending artery and the superior left circumflex artery provided in this embodiment of the present disclosure includes: The spatial information acquisition module 810 is used to acquire a continuous cross-sectional image sequence of the target vascular segment, separate multiple vascular branch image sequences on the left anterior descending artery segment based on the continuous cross-sectional image sequence, and determine the spatial structure information corresponding to each vascular branch.

[0128] The image feature acquisition module 820 is used to extract spatial feature information of blood vessel branches based on the corresponding spatial structure information for each blood vessel branch, and to extract image feature information of blood vessel branches based on the corresponding branch image sequence.

[0129] The feature fusion module 830 is used to fuse the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused features of the corresponding blood vessel branch.

[0130] The location discrimination module 840 is used to classify and determine each vascular branch according to the fusion features, obtain the determination result of whether each vascular branch is a left circumflex branch, and output the location of the left circumflex branch according to the determination result.

[0131] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0132] This disclosure provides a device for locating the left circumflex branch of the left anterior descending artery. By acquiring a continuous sequence of cross-sectional images, it separates the vascular branches on the left anterior descending artery segment and combines the spatial structural information and image feature information of the vascular branches for fusion determination, thus achieving automatic location of the left circumflex branch. Compared to existing methods that mainly rely on single image information or manual experience, this device effectively improves branch recognition accuracy, enhances adaptability to noise, occlusion, and image quality fluctuations, reduces ambiguity in branch attribute judgment, and thereby improves the accuracy, continuity, and auxiliary analysis efficiency of clinical operations.

[0133] Corresponding to Figure 1 The present disclosure also provides an electronic device 900 for locating the left anterior descending artery and superior circumflex artery in cardiovascular arteries. Figure 9 The diagram shown is a structural schematic of an electronic device 600 provided in an embodiment of this disclosure, including: The system includes a processor 91, a memory 92, and a bus 93. The memory 92 stores execution instructions and includes main memory 921 and external memory 922. Main memory 921, also called internal memory, temporarily stores computational data in the processor 91, as well as data exchanged with external memory such as a hard disk. The processor 91 exchanges data with external memory 922 through main memory 921. When the electronic device 900 is running, the processor 91 communicates with the memory 92 via the bus 93, enabling the processor 91 to execute... Figure 1 The steps of the method for locating the left anterior descending artery and the superior left circumflex artery in the cardiovascular system.

[0134] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of the method for locating the left circumflex branch of the left anterior descending artery as described in the above-described method embodiments. The storage medium can be volatile or non-volatile computer-readable storage.

[0135] This disclosure also provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, they can perform the steps of the method for locating the left anterior descending artery and the left circumflex artery as described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0136] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0137] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed device and method can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0138] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0139] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0140] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, 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 this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for locating the superior left circumflex branch of the left anterior descending artery of a cardiovascular system, characterized in that, include: A continuous cross-sectional image sequence of the target vascular segment is obtained, and multiple vascular branch image sequences on the left anterior descending artery segment are separated based on the continuous cross-sectional image sequence, while the spatial structure information corresponding to each vascular branch is determined. For each vascular branch, spatial feature information of the vascular branch is extracted based on the corresponding spatial structure information, and image feature information of the vascular branch is extracted based on the corresponding branch image sequence; The spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are fused to obtain the fused features of the corresponding blood vessel branches; Based on the fusion features, each vascular branch is classified and determined to determine whether each vascular branch is a left circumflex branch, and the position of the left circumflex branch is output according to the determination result. The spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are fused to obtain the fused features of the corresponding blood vessel branches, specifically including: The spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are subjected to dimensionality-upgrading processing to construct a fusion matrix representing the interaction relationship between the two types of features; The fusion matrix is ​​convolved to enhance the expression of the combination relationship between adjacent features; The fusion matrix after convolution processing is adaptively weighted based on the vascular branch image feature information to enhance key fusion features related to the determination of the left circumflex branch; The weighted fusion matrix is ​​downsampled and flattened to obtain the fusion representation vector. The fused representation vector is concatenated with the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused feature.

2. The method according to claim 1, characterized in that, Obtain a continuous cross-sectional image sequence of the target vessel segment, and separate multiple vessel branch image sequences on the left anterior descending artery segment based on the continuous cross-sectional image sequence. Simultaneously, determine the spatial structure information corresponding to each vessel branch, specifically including: Intravascular imaging was performed on the left anterior descending artery segment to obtain cross-sectional images continuously distributed along the vessel axis. The vascular lumen region and branch region are identified in each frame of cross-sectional image to obtain the branch contour information in each frame; The branch contour information in different frames is correlated, and multiple branch contours belonging to the same blood vessel branch are combined to form a corresponding blood vessel branch image sequence. The spatial structure information of the corresponding vascular branch is determined based on the vascular branch image sequence. The spatial structure information includes at least one of the following: branch location, branch cross-sectional area, and branch length.

3. The method according to claim 1, characterized in that, Based on the corresponding spatial structure information, spatial feature information of blood vessel branches is extracted, specifically including: The spatial structural information representing the branch location, branch cross-sectional area, and branch length is input into the spatial feature extraction network; During training, the spatial feature extraction network performs similarity-bringing constraints on spatial features corresponding to the same type of vascular branches, performs similarity-difference constraints on spatial features corresponding to different types of vascular branches, and applies redundancy removal constraints to spatial features from different sources before and after fusion. The spatial feature extraction network is used to perform convolutional feature extraction and aggregation on the spatial structure information to obtain the first spatial feature. The first spatial feature is subjected to dimensionality reduction representation processing to obtain the second spatial feature; The first spatial feature and the second spatial feature are fused together to obtain the spatial feature information of blood vessel branches.

4. The method according to claim 3, characterized in that, The first spatial feature is subjected to dimensionality reduction representation processing to obtain the second spatial feature, specifically including: During training, a feature distribution is constructed based on the spatial features extracted at the current training stage; Based on the feature distribution, a low-dimensional projection relationship matching the current training phase is determined; The first spatial feature is reduced in dimension using the low-dimensional projection relationship to obtain the second spatial feature.

5. The method according to claim 3, characterized in that, The first spatial feature and the second spatial feature are fused to obtain the spatial feature information of blood vessel branches, specifically including: Perform a mapping transformation on the second spatial feature to make the second spatial feature correspond to the first spatial feature in the feature representation dimension; The transformed second spatial features are used as supplementary information and residually fused with the first spatial features to obtain the spatial feature information of the blood vessel branches.

6. The method according to claim 1, characterized in that, Extract vascular branch image feature information from the corresponding branch image sequence, specifically including: Input multiple frames of images of the corresponding blood vessel branches into a 3D image feature extraction network; During the training process, the three-dimensional image feature extraction network labels each vascular branch on the left anterior descending artery to obtain labels for the left circumflex branch and non-left circumflex branch. The labels are used to supervise the training of the three-dimensional image feature extraction network, and the vascular branch image feature information is extracted from the feature output layer of the three-dimensional image feature extraction network. Temporal correlation features and spatial texture features are extracted from the multi-frame images using 3D convolution processing; The features processed by 3D convolution are compressed in dimension and extracted by 2D convolution to obtain deep image representations. The deep image representation is converged and vectorized to obtain the vascular branch image feature information.

7. The method according to claim 1, characterized in that, Based on the fusion features, each vascular branch is classified and determined to determine whether each vascular branch is a left circumflex branch. The location of the left circumflex branch is then output based on the determination result, specifically including: The fused features are input into the classification network; The output vascular branch category results are processed through at least two fully connected layers, and it is determined whether the corresponding vascular branch is a left circumflex branch based on the category results; Based on the determination results of whether each vascular branch is a left circumflex branch, a corresponding classification label sequence is generated; The classification label sequence is mapped to the branch positions in the left anterior descending artery segment to output the position of the left circumflex artery branch.

8. A positioning device for the superior left circumflex branch of the left anterior descending artery of a cardiovascular system, characterized in that, The method for locating the left superior circumflex branch of the left anterior descending artery as described in any one of claims 1 to 7, the device comprising: The spatial information acquisition module is used to acquire a continuous cross-sectional image sequence of the target vascular segment, separate multiple vascular branch image sequences on the left anterior descending artery segment based on the continuous cross-sectional image sequence, and determine the spatial structure information corresponding to each vascular branch. The image feature acquisition module is used to extract spatial feature information of blood vessel branches based on the corresponding spatial structure information for each blood vessel branch, and to extract image feature information of blood vessel branches based on the corresponding branch image sequence. The feature fusion module is used to fuse the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused features of the corresponding blood vessel branch. Specifically, the spatial feature information of the blood vessel branches and the image feature information of the blood vessel branches are subjected to dimensionality-upgrading processing to construct a fusion matrix representing the interaction relationship between the two types of features; The fusion matrix is ​​convolved to enhance the expression of the combination relationship between adjacent features; The fusion matrix after convolution processing is adaptively weighted based on the vascular branch image feature information to enhance key fusion features related to the determination of the left circumflex branch; The weighted fusion matrix is ​​downsampled and flattened to obtain the fusion representation vector. The fused representation vector is concatenated with the spatial feature information of the blood vessel branch and the image feature information of the blood vessel branch to obtain the fused feature; The location discrimination module is used to classify and determine each vascular branch according to the fusion features, obtain the determination result of whether each vascular branch is a left circumflex branch, and output the location of the left circumflex branch according to the determination result.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is in operation, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the method for locating the left circumflex branch of the left anterior descending artery as described in any one of claims 1 to 7.

Images