A pancreatic artery image segmentation method, system, storage medium and program product

Abstract

The application belongs to the technical field of medical imaging, and discloses a pancreatic artery image segmentation method, system, storage medium and program product, the method comprising: generating a first confidence blood vessel mask and a second confidence blood vessel mask based on a probability graph; performing connected component analysis and three-dimensional refinement processing on the second confidence blood vessel mask to generate a blood vessel center skeleton, determine a main skeleton and a candidate branch, and perform iterative connection and verification on the main skeleton and the candidate branch to obtain a final main skeleton; obtaining an initial fusion mask according to the final main skeleton and the first confidence blood vessel mask, and performing connected component screening and organ constraint filtering on the initial fusion mask to output a segmentation mask of the pancreatic artery. The application can automatically obtain a complete, continuous and complex three-dimensional structure of the pancreatic artery, and output a segmentation mask that can be directly used for three-dimensional reconstruction visualization, thereby providing a more reliable and accurate imaging basis for preoperative blood vessel evaluation and surgical planning of pancreatic cancer.

Description

Technical Field

[0001] This invention relates to the fields of medical imaging and computer vision technology, and in particular to a pancreatic artery image segmentation method, system, storage medium, and program product. Background Technology

[0002] The pancreas and its surrounding region are richly supplied with complex blood vessels, and there is a close spatial relationship between diseased tissue and these vessels. Accurately acquiring information on the morphology, course, and spatial distribution of important blood vessels around the pancreas is crucial for 3D reconstruction visualization of medical images, automatic identification of anatomical structures, and image-assisted analysis. However, the anatomical structure of the arteries surrounding the pancreas exhibits significant individual differences, with arterial variations being quite common. For example, the hepatic artery may originate from different arterial branches, and the morphology and branching patterns of the pancreaticoduodenal arterial arch also vary considerably; the course, number of branches, and spatial distribution of blood vessels can also show significant variations among individuals. This complex and diverse vascular anatomy places high technical demands on automated image segmentation of pancreatic arteries, increasing the difficulty of accurately extracting complete vascular structures.

[0003] In medical imaging vascular segmentation, existing methods are typically based on deep learning models trained on specific training datasets, thus making the models highly dependent on the distribution of the training data. However, the arteries surrounding the pancreas exhibit significant individual differences and complex anatomical variations, with substantial differences between individuals in vascular origin, course, and branching morphology. Applying existing methods to pancreatic arterial images with complex anatomical variations easily leads to problems such as incomplete segmentation, missed branches, or misidentification. This is especially true in regions with slender, small-diameter vascular branches, where it is difficult to accurately extract continuous vascular structures, resulting in incomplete vascular topology and insufficient stability and generalization ability of the overall segmentation results.

[0004] Furthermore, the aforementioned methods typically rely directly on the segmentation results output by deep learning models, exhibiting a strong dependence on specific training data distributions. When dealing with complex anatomical variations in the pancreatic artery, their segmentation stability is insufficient, making it difficult to obtain reliable and detailed vascular structures, thus affecting the robustness and adaptability of the methods. Simultaneously, existing methods are prone to branch omissions and vessel breaks in regions of slender vessel branches, making it difficult to achieve continuous and complete extraction of small vessel branches and obtain a complete vascular topology, thus limiting the integrity and visualization effects of 3D reconstruction of the pancreatic artery.

[0005] Therefore, how to provide a pancreatic artery image segmentation method, system, storage medium, and program product is an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a pancreatic artery image segmentation method, system, storage medium, and program product to solve the aforementioned technical problems existing in the prior art.

[0007] According to a first aspect of the present invention, a pancreatic artery image segmentation method is provided.

[0008] In one embodiment, a pancreatic artery image segmentation method includes: The acquired image is preprocessed and then input into a pre-trained deep learning segmentation network to output a probability map. A first confidence blood vessel mask and a second confidence blood vessel mask are generated based on the probability map. Connectivity component analysis and 3D refinement are performed on the second confidence vessel mask to generate the vessel center skeleton. Based on the vessel center skeleton, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton. Based on the final main skeleton and the first confidence vessel mask, an initial fusion mask is obtained. Connected component filtering and organ constraint filtering are then performed on the initial fusion mask to output the segmentation mask of the pancreatic artery, thereby achieving pancreatic artery image segmentation.

[0009] According to a second aspect of the present invention, a pancreatic artery segmentation system is provided.

[0010] In one embodiment, a pancreatic artery segmentation system includes: A multi-level confidence mask generation unit is used to preprocess the acquired image, input the preprocessed image into a pre-trained deep learning segmentation network, output a probability map, and generate a first confidence blood vessel mask and a second confidence blood vessel mask based on the probability map. The skeleton extraction and iterative connection unit is used to perform connected component analysis and three-dimensional refinement on the second confidence blood vessel mask to generate the central skeleton of the blood vessel. Based on the central skeleton of the blood vessel, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton. The fusion constraint and refinement output unit is used to obtain an initial fusion mask based on the final main skeleton and the first confidence vascular mask, and to perform connected component screening and organ constraint filtering on the initial fusion mask to output the segmentation mask of the pancreatic artery, so as to achieve pancreatic artery image segmentation.

[0011] According to a third aspect of the present invention, a computer-readable storage medium is provided.

[0012] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0013] According to a fourth aspect of the present invention, a computer program product is provided.

[0014] In one embodiment, the computer program product stores a computer program that, when executed by a processor, implements the steps of the method.

[0015] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: 1. This invention obtains high-confidence vascular structures and vascular structures containing all low-threshold responses by using a first confidence vascular mask and a second confidence vascular mask, respectively. While ensuring the complete preservation of high-confidence vascular structures, it effectively incorporates weak response signals from small vascular branches, making it easy to stably reproduce on different datasets and improving the robustness of the method to individual anatomical variations in pancreatic arteries.

[0016] 2. This invention performs three-dimensional thinning under endpoint protection, foreground and background local connectivity constraints, and Euler feature invariance constraints, extracting the central skeleton of the blood vessel that maintains the original blood vessel connectivity and bifurcation topology. This reduces the skeleton breakage or topological distortion problems caused by traditional thinning algorithms and provides a structurally reliable topological foundation for subsequent branch connections.

[0017] 3. This invention distinguishes between the main skeleton and candidate branches, and establishes pairing costs based on the distance between endpoints, local tangent directions, and spatial structural relationships. It constructs a graph search cost function with negative logarithmic probability and path length weighting, realizing probability-driven minimum cost path search. This can be used to recover blood vessel ruptures caused by incomplete deep learning segmentation. At the same time, through secondary confirmation of path rationality, it effectively suppresses erroneous connections and improves the continuity of slender, small-diameter branches and the overall topological integrity.

[0018] 4. This invention, through a collaborative processing flow of dual-threshold extraction, topology-preserving skeletonization, intelligent path connection, and secondary confirmation of path rationality, can automatically obtain complete and continuous three-dimensional structures of complex pancreatic arteries and output segmentation masks that can be directly used for three-dimensional reconstruction visualization, providing more reliable and accurate imaging evidence for preoperative vascular assessment and surgical planning in pancreatic cancer.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0021] Figure 1This is one of the flowcharts illustrating a pancreatic artery image segmentation method according to an exemplary embodiment; Figure 2 This is a schematic diagram of a pancreatic artery segmentation system according to an exemplary embodiment; Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a second flowchart illustrating a pancreatic artery image segmentation method according to an exemplary embodiment; Figure 5 This is the third flowchart illustrating a pancreatic artery image segmentation method according to an exemplary embodiment. Detailed Implementation

[0022] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0023] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0025] Figure 1 An embodiment of a pancreatic artery image segmentation method according to the present invention is shown.

[0026] In this optional embodiment, a pancreatic artery image segmentation method includes: Step S101: Preprocess the acquired image and input the preprocessed image into the pre-trained deep learning segmentation network to output a probability map. Generate a first confidence blood vessel mask and a second confidence blood vessel mask based on the probability map. Step S102: Perform connected component analysis and three-dimensional refinement on the second confidence vessel mask to generate the vessel center skeleton. Based on the vessel center skeleton, determine the main skeleton and candidate branches, and iteratively connect and verify the main skeleton and candidate branches to obtain the final main skeleton. Step S103: Based on the final main skeleton and the first confidence vessel mask, an initial fusion mask is obtained, and connected component filtering and organ constraint filtering are performed on the initial fusion mask to output the segmentation mask of the pancreatic artery, so as to achieve pancreatic artery image segmentation.

[0027] In this optional embodiment, the deep learning segmentation network is a convolutional neural network that includes an encoder, a decoder, and skip connections; wherein, the encoder is used to extract multi-scale features of the preprocessed image layer by layer, the decoder is used to restore the spatial resolution layer by layer, and the skip connections are used to concatenate and fuse the feature maps of the corresponding layers in the encoder and the corresponding layers in the decoder.

[0028] In this optional embodiment, generating the first and second confidence vessel masks based on the probability map includes: using a set of voxels in the probability map whose probability values ​​are greater than or equal to a first preset threshold as the first confidence vessel mask, and using a set of voxels in the probability map whose probability values ​​are greater than or equal to a second preset threshold as the second confidence vessel mask; wherein the second preset threshold is less than the first preset threshold.

[0029] In this optional embodiment, connected component analysis and 3D refinement are performed on the second confidence vessel mask to generate a vessel center skeleton. Based on the vessel center skeleton, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton, which includes: Connectivity component analysis is performed on the second confidence vessel mask to obtain the number of voxels in each connected component. The connected components are then filtered based on a preset minimum voxel threshold, and the second confidence candidate region is obtained based on the filtering results. The second confidence candidate region is subjected to topology-preserving 3D refinement to extract the central vascular skeleton; wherein the central vascular skeleton preserves the connectivity and bifurcation topology of the original vascular structure. Based on the central skeleton of the blood vessel, the largest connected component is determined as the main skeleton, and the remaining connected components are determined as candidate branches. By using endpoint pairing and graph search path connection strategies, the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton.

[0030] In this optional embodiment, the three-dimensional refinement process includes: protecting endpoints in the preset neighborhood whose number of foreground neighbor points is less than or equal to a target threshold from being deleted; dividing the voxel space into multiple sub-domains based on the parity of the three-dimensional coordinates, and sequentially deleting non-endpoints that meet the preset deletion conditions.

[0031] In this optional embodiment, the preset deletion conditions include: after deleting non-endpoints, the local connectivity between the foreground and the background remains unchanged, and the Euler eigenvalues ​​before and after deletion remain unchanged.

[0032] In this optional embodiment, an endpoint pairing and graph search path connection strategy is used to iteratively connect and verify the main skeleton and candidate branches, resulting in the final main skeleton including: Extract the endpoint coordinates of the main skeleton and candidate branches respectively, pair the endpoints of the main skeleton with the endpoints of the candidate branches based on the endpoint coordinates, and calculate the pairing cost of each endpoint pair. Endpoint pairs that satisfy the pairing constraints are determined as valid endpoint pairs; wherein, the pairing constraints include that the Euclidean distance between the two endpoints is less than a preset distance threshold, and the angle between the skeleton tangent directions at the two endpoints is less than a preset angle threshold. The local spatial region determined by each effective endpoint pair is used as the search region, and a three-dimensional weighted graph is constructed within the search region based on a preset neighborhood. Based on a 3D weighted graph, with one endpoint of an effective endpoint pair as the starting point and the other endpoint as the ending point, and using the weighted sum of the negative log probability value and the path length as the edge cost, we search for the minimum cost path from the starting point to the ending point. The minimum cost path is used as the candidate connection path between the main skeleton and the candidate branch. The candidate connection path is then confirmed a second time, and the candidate connection path and the corresponding candidate branch that are confirmed to be reasonable are merged into the main skeleton. The endpoint pairing, path search, and verification merging process is iteratively executed until no new candidate connection paths are accepted, resulting in the final main skeleton.

[0033] In this alternative embodiment, the pairing cost is determined based on at least one of the following factors: the Euclidean distance between the two endpoints, the angle between the tangent directions of the skeleton at the two endpoints, and the spatial relationship between the skeleton structures to which the two endpoints belong.

[0034] In this optional embodiment, the secondary confirmation includes: judging whether the connection path is reasonable based on the consistency between the end direction of the candidate connection path and the corresponding skeleton endpoint direction; if reasonable, performing fusion update; if unreasonable, re-performing endpoint pairing and path search.

[0035] In this optional embodiment, an initial fusion mask is obtained based on the final main skeleton and the first confidence vessel mask. Connectivity component filtering and organ constraint filtering are then performed on the initial fusion mask to output a segmentation mask for the pancreatic artery, thereby achieving pancreatic artery image segmentation. The final main skeleton is expanded using a preset spherical structural element, and then the expanded final main skeleton is fused with the first confidence blood vessel mask to obtain the initial fused mask. Connected component analysis is performed on the initial fusion mask to remove connected components with fewer voxels than the preset number of voxels. Combined with the preset auxiliary organ mask, connected components completely located within the auxiliary organ mask are deleted to obtain the segmentation mask of the pancreatic artery, thereby achieving pancreatic artery image segmentation.

[0036] Figure 2 An embodiment of a pancreatic artery segmentation system according to the present invention is shown.

[0037] In this alternative embodiment, a pancreatic artery segmentation system includes: The multi-level confidence mask generation unit 201 is used to preprocess the acquired image, input the preprocessed image into a pre-trained deep learning segmentation network, output a probability map, and generate a first confidence blood vessel mask and a second confidence blood vessel mask based on the probability map. The skeleton extraction and iterative connection unit 202 is used to perform connected component analysis and three-dimensional refinement processing on the second confidence blood vessel mask to generate the blood vessel center skeleton, determine the main skeleton and candidate branches based on the blood vessel center skeleton, and perform iterative connection and verification on the main skeleton and candidate branches to obtain the final main skeleton. The fusion constraint and refinement output unit 203 is used to obtain an initial fusion mask based on the final main skeleton and the first confidence blood vessel mask, and to perform connected component screening and organ constraint filtering on the initial fusion mask to output the segmentation mask of the pancreatic artery, so as to achieve pancreatic artery image segmentation.

[0038] In this optional embodiment, the skeleton extraction and iterative connection unit 202 includes: The connected component analysis module is used to perform connected component analysis on the second confidence blood vessel mask, obtain the number of voxels in each connected component, and filter the connected components based on a preset minimum voxel threshold, and obtain the second confidence candidate region based on the filtering results. The 3D refinement module is used to perform topology-preserving 3D refinement on the second confidence candidate region and extract the central skeleton of the blood vessel; wherein the central skeleton of the blood vessel maintains the connectivity and bifurcation topology of the original blood vessel structure; The connected component determination module is used to determine the largest connected component as the main skeleton and the remaining connected components as candidate branches based on the central skeleton of the blood vessel. The iterative connection and verification module is used to perform iterative connection and verification of the main skeleton and candidate branches using endpoint pairing and graph search path connection strategies to obtain the final main skeleton.

[0039] In this optional embodiment, the fusion constraint and refinement output unit 203 includes: The initial fusion mask generation module is used to expand the shape of the final main skeleton using preset spherical structural elements, and then fuse the expanded final main skeleton with the first confidence blood vessel mask to obtain the initial fusion mask; The segmentation mask acquisition module is used to perform connected component analysis on the initial fusion mask, remove connected components with fewer than the preset number of voxels, and combine with the preset auxiliary organ mask to delete connected components that are completely located within the auxiliary organ mask, so as to obtain the segmentation mask of the pancreatic artery and realize pancreatic artery image segmentation.

[0040] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows: 1. A U-shaped medical image segmentation network is used to segment preprocessed 3D CT images to obtain voxel-by-voxel pancreatic artery probability maps. The resampling interval, input block size, number of network layers, convolution kernel, and downsampling stride are determined based on the voxel spacing, image size, and memory constraints. An encoder-decoder and same-scale skip connections are used to output high-resolution vascular probability maps.

[0041] 2. Set fixed dual thresholds for the vascular probability map: use 0.5 as the high confidence threshold to obtain reliable vascular regions, i.e., the first confidence vascular mask; use 0.1 as the low confidence threshold to obtain vascular structures containing all low-threshold responses, i.e., the second confidence vascular mask. Perform topology-preserving 3D refinement on the second confidence vascular mask, deleting simple boundary points layer by layer while protecting endpoints and connectivity, to obtain the central skeleton of the vascular system.

[0042] 3. Take the largest connected component of the skeleton as the main skeleton and the remaining effective components as branches to be connected; detect the endpoints of each component and construct candidate endpoint pairs based on the distance between endpoints, local orientation consistency and branch structure relationship.

[0043] 4. In the local probability space corresponding to the candidate endpoint pair, construct the path cost with the negative logarithm of the blood vessel probability and search for the minimum cost connection path; perform a second verification of the rationality of the minimum cost connection path, and after verification, merge the path and the branch to be connected into the main skeleton and iteratively update.

[0044] 5. Finally, the updated main skeleton is expanded into a connected structure with a preset width and fused with a high-confidence vascular mask. Then, unreasonable regions are removed according to connectivity, structural scale and anatomical constraints, and continuous three-dimensional binary segmentation results of pancreatic artery are output.

[0045] Specifically, the input of this invention is an X-ray computed tomography (CT) image, and the complete blood vessel segmentation algorithm flow is as follows: Figure 4 As shown.

[0046] Step S1, Input CT Image and Image Preprocessing: The original CT image is spatially cropped using a pancreatic region localization strategy to obtain a raw 3D abdominal image containing the pancreas and surrounding vascular structures, reducing interference from irrelevant tissues in subsequent segmentation tasks. Resampling is performed according to the planned target spacing; high-order interpolation is used for continuous images, and low-order interpolation is used for discrete labels. Foreground intensity is limited to the low and high quantiles of the training set foreground intensity, and then standardized according to the mean and standard deviation of the training set foreground intensity. During the training phase, 3D input blocks are cropped from the resampled images, and samples are expanded through rotation, scaling, mirroring, noise, blurring, brightness, and contrast perturbations to achieve intensity normalization.

[0047] Semantic segmentation network and 3D probabilistic graph: The encoder employs a multi-level 3D convolutional structure. In one specific embodiment, six resolution levels are used, with an input block size of 192×192×192 voxels. The number of output channels for each level is 32, 64, 128, 256, 320, and 320, respectively. Each level consists of two concatenated 3D convolutional-instance normalization-linear activation units, without using random deactivation layers. To accommodate anisotropic data where the interlayer spacing is significantly larger than the in-plane pixel spacing, the encoder uses 3×3×3 convolutional kernels, simultaneously aggregating interlayer and in-plane information. The stride of the six levels is (1,1,1), (2,2,2), (2,2,2), (2,2,2), (2,2,2), and (1,2,2), where the first number corresponds to the layer direction, and the last two numbers correspond to the in-plane direction. This allows for joint downsampling of all directions in shallow layers and focused compression of in-plane dimensions in deeper layers. The decoder is configured with five upsampling levels that correspond one-to-one with the encoder, except for the bottleneck level. Each level first uses a 3D transposed convolution with the same stride as the corresponding encoder to upsample low-resolution features to the next higher scale, then concatenates them along the channel dimension with the encoder's output at the same scale to form a skip connection. The concatenated features are then fused using two 3D convolutions—instance normalization—linear activation units. Skip connections directly transmit information about superficial vessel edges and fine branches to the decoder, while deep features provide semantic distinction between vessels and surrounding tissues. A 1×1×1 3D convolutional segmentation head is set up at each decoding scale for deep supervised training. During training, lower-resolution labels are downsampled according to the corresponding scale, and the loss of each segmentation head is calculated and assigned decreasing weights according to resolution from high to low. During inference, only the highest resolution segmentation head is used. For binary classification of background and pancreatic artery, the highest resolution segmentation head outputs two class response values, which are then transformed by a normalized exponential function to obtain the pancreatic artery channel. The network training uses the sum of cross-entropy loss and Dice similarity coefficient loss, i.e., .in, Used to measure the error in voxel classification. Used to measure the difference in overlap between predicted vascular regions and actual vascular regions.

[0048] For a binary classification task involving background and pancreatic arteries, the cross-entropy loss is defined as: ; The Dessian similarity coefficient loss is defined as: ; in, This represents the total number of voxels involved in the loss calculation. Indicates the voxel number; Indicates the first The three-dimensional coordinates of an individual element; This indicates the probability that the network predicts the voxel belongs to the pancreatic artery; This indicates the true label of the voxel, with 1 representing blood vessels and 0 representing the background. To prevent the denominator from being zero and to avoid the instability of the logarithmic value, a positive number is set in this embodiment. .

[0049] During inference, an overlapping sliding window is used to traverse the 3D image. A high Gaussian weight is assigned to the center of the window, and a weighted average is applied to the overlapping areas to reduce the edge error of the input blocks. Mirror inference can be performed along the spatial axes allowed during training, and the results are averaged. Finally, the probability map is inversely resampled according to the preprocessed records and backfilled to the position before cropping to restore the spatial coordinates, voxel spacing, and orientation information of the original CT image for use in post-processing of blood vessels.

[0050] Step S2, post-vascular processing: such as Figure 5 As shown, the input is a three-dimensional probability map. Adaptive threshold calculation is performed to generate high-confidence and low-confidence maps, namely, a first-confidence vessel mask and a second-confidence vessel mask. In this embodiment, the high-confidence threshold... Low confidence threshold The corresponding mask is defined as: , ; in, Represents the three-dimensional coordinates of any voxel. These represent the coordinates in three spatial directions; This indicates the probability of predicting blood vessels for this voxel; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Graph nodes , Each represents the coordinates of a specific voxel, therefore The probability satisfies The voxels constitute the second confidence vascular mask, with a probability satisfying The voxels form the first-confidence vascular mask. Prior to skeleton extraction, [the following is observed / constructed / etc.]. Preprocessing is performed. To enhance connectivity, a 3×3×3 fully connected neighborhood (i.e., a 26-neighborhood) is used, and the connected components within it are labeled. The union of components with fewer than 15 voxels is denoted as the small component set. .set up Indicated by radius The spherical structuring elements of individual elements are continuously executed on set A. Submorphic expansion, then This represents a spherical structural element with a radius of 1 voxel. The expansion was repeated four times. The centerline extracted from this expanded region is denoted as... The enhanced low-threshold structure is uniformly denoted as ;in Represents the set difference. This represents the union of sets.

[0051] Step S3, for Preprocessing and skeleton extraction are performed. First, the non-zero bounding box is cropped and expanded outwards by 5 voxels. Iterative refinement based on Eulerian features is then performed in binary input mode. In each round, endpoints with no more than 1 foreground neighbor within a 26-neighbor range are protected. Then, eight subdomains are divided according to the parity of the 3D coordinates, and simple non-endpoint points are deleted sequentially. Simple points satisfy the following conditions: after deleting candidate points, the remaining foreground voxels form a connected component with 26 neighbors, the background voxels form a connected component with 6 neighbors, and the Eulerian features remain unchanged before and after deletion. ,and .in, , , and These represent the number of vertices, edges, faces, and 3D units in the local voxel complex, respectively. In this embodiment, the main iteration count is 10, and after completion, a central skeleton S is obtained that maintains connectivity, bifurcation relationships, and endpoints.

[0052] Step S4: Initialize the skeleton S and extract endpoint candidate points. Count the layer range where the skeleton has foreground features, assuming the first foreground layer is... The number of foreground layers is The effective master endpoint level range is defined as follows: , This is to reduce the impact of unstable endpoints at the top and bottom edges on the connection process.

[0053] S is decomposed into a set of connected components using a 26-neighborhood. ;in, It represents the set consisting of all skeleton connected components. Indicates the total number of connected components. Represents any connected component in the set. Indicates the number is The connected components.

[0054] The main skeleton is defined as: ,in Represents connected components Number of skeletal voxels included. The component with the largest number of voxels is represented; other connected components with a voxel count greater than 5 are considered candidate branches. Skeleton points are treated as nodes in the graph, and connections between adjacent skeleton points are considered edges. A voxel with an adjacency degree of 1 is found as the initial endpoint; from this endpoint, the process traces inward along its corresponding skeleton until another endpoint, a branch point, or a point without successors is reached. Only initial endpoints with a tracing length of at least 3 voxels are retained as valid endpoints. This minimum branch length parameter is used to exclude false endpoints formed by short spikes.

[0055] Step S5, search for connection paths; for candidate branches In its set of valid endpoints With the set of valid endpoints of the main skeleton Calculate the Euclidean distance between them: ;in, The first branch of the candidate branch One valid endpoint, The first part represents the main skeleton. One valid endpoint, , These are the indices in the two endpoint sets, This represents the Euclidean norm.

[0056] The endpoint pair with the smallest distance is denoted as... , representing the endpoints of the candidate branch and the main skeleton, respectively. The first iteration requires distance... Individual factors, and The plane coordinates are located at From respectively and Obtain at least two local skeleton points along the skeleton, fit the endpoint tangent through principal component analysis, and unify them to point from the inside of the skeleton to the outside of the endpoint.

[0057] set up The included angle between the two tangential directions. and These are the angles between the tangents of the candidate branch and the main skeleton end, and the vectors pointing to the opposite ends, respectively; the number of candidate branch voxels is uniformly denoted as... This embodiment requires... , and If multiple candidate branches correspond to the same endpoint of the main skeleton, only the endpoint pair with the smallest distance is retained.

[0058] Step S6: Determine if the connection path is reasonable. If so, perform a fusion update; otherwise, re-execute the extraction of endpoint candidate points. Establish a local search region Ω for the selected endpoint pairs. Take the smallest axis-aligned cuboid containing both endpoints and expand it outwards by 10 voxels in each of the three coordinate directions. Construct a 26-neighborhood graph from the voxels in Ω. ,in Let represent the set of all voxel nodes within Ω. This represents the set of edges between 26 neighboring nodes. and express Two adjacent voxel nodes in the node. The probabilistic cost is defined as: ;in, Represents a node Predicted probability of blood vessels at the location; Indicates will Restricted to closed intervals Inside; To prevent positive numbers with an undefined logarithm when the probability is 0.

[0059] Adjacent nodes , The edge cost is: ,and Dijkstra's algorithm can be used to search for the minimum cost path on an undirected graph; where This is the regularization weight for path length. From... arrive The optimal connection path is defined as: ; in, Let Ω represent the set of all feasible 26-neighborhood paths connecting the two endpoints. Indicates any one of these paths. This represents the path with the minimum cumulative edge cost. Representing a path Passing adjacent node pairs The cost of the edge.

[0060] When only node probability cost is used, let When path length needs to be considered, .

[0061] The optimal connection path is re-confirmed before the merge update. Let the candidate connection path be... The candidate branch endpoints and the main skeleton endpoints correspond to respectively and Select from both ends of the path A local path point, which is taken in this embodiment. and define the path end direction and .in, When the path length is less than 2 steps, take... .

[0062] Let the outward unit tangents of the candidate branch endpoint and the main skeleton endpoint be respectively... and Define the included angle for secondary confirmation. and .when and In this embodiment, the connection path is determined to be reasonable and a fusion update is performed. Otherwise, the connection path is deemed unreasonable, and the endpoint candidate point extraction step is returned to re-execute the endpoint pairing and path search; at the same time, the corresponding endpoint pair is marked as rejected and excluded in the current iteration; if no new valid endpoint pair exists, the iteration is terminated.

[0063] Merge Update: The optimal path and its corresponding candidate branches are merged into the main skeleton, and the update relationship is as follows: ; in, Indicates the iteration round; Indicates the first The main framework at the start of each iteration; This indicates the main framework for the next round after the merger; This indicates the optimal connection path accepted in this round; Indicates the first connected in this round There are several candidate branches. After updating, the endpoints are re-examined and the next round of candidate endpoint pairs is constructed; if no new path is accepted, the iteration stops.

[0064] Step S7, Segmentation Result: The final main skeleton Compared with first confidence vascular mask To perform a merge update, among which This represents the final main framework obtained at the end of the iteration. A spherical structural element with a radius of 1 voxel is used. One expansion is recorded as The fusion result is: ;in This represents the initial fusion mask before connected component filtering.

[0065] Using 26 neighborhood pairs Connected component labeling is performed, with components having at least 300 voxels considered as valid candidates. Discrete connected components are deleted, and components completely located within a preset excluded organ region are removed using an auxiliary organ mask. The reconstructed structure output is a three-dimensional binary pancreatic artery segmentation mask. The aforementioned voxel threshold can be proportionally adjusted based on the physical volume when the image resampling interval changes.

[0066] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0067] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0068] In addition, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0069] In addition, the present invention also provides a computer program product that, when executed by a processor, implements the steps of the method described.

[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0071] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for segmenting pancreatic artery images, characterized in that, The method includes: The acquired image is preprocessed and then input into a pre-trained deep learning segmentation network to output a probability map. A first confidence blood vessel mask and a second confidence blood vessel mask are generated based on the probability map. Connectivity component analysis and 3D refinement are performed on the second confidence vessel mask to generate the vessel center skeleton. Based on the vessel center skeleton, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton. Based on the final main skeleton and the first confidence vessel mask, an initial fusion mask is obtained. Connected component filtering and organ constraint filtering are then performed on the initial fusion mask to output the segmentation mask of the pancreatic artery, thereby achieving pancreatic artery image segmentation.

2. The pancreatic artery image segmentation method according to claim 1, characterized in that, The deep learning segmentation network is a convolutional neural network comprising an encoder, a decoder, and skip connections; wherein, the encoder is used to extract multi-scale features of the preprocessed image layer by layer, the decoder is used to restore spatial resolution layer by layer, and the skip connections are used to concatenate and fuse the feature maps of the corresponding layers in the encoder and the corresponding layers in the decoder.

3. The pancreatic artery image segmentation method according to claim 1, characterized in that, The generation of the first and second confidence vessel masks based on the probability graph includes: using a set of voxels in the probability graph whose probability values ​​are greater than or equal to a first preset threshold as the first confidence vessel mask, and using a set of voxels in the probability graph whose probability values ​​are greater than or equal to a second preset threshold as the second confidence vessel mask; wherein the second preset threshold is less than the first preset threshold.

4. The pancreatic artery image segmentation method according to claim 1, characterized in that, The second confidence vessel mask is subjected to connected component analysis and 3D refinement to generate a vessel center skeleton. Based on the vessel center skeleton, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton, which includes: Connectivity component analysis is performed on the second confidence vessel mask to obtain the number of voxels in each connected component. The connected components are then filtered based on a preset minimum voxel threshold, and the second confidence candidate region is obtained based on the filtering results. The second confidence candidate region is subjected to topology-preserving 3D refinement to extract the central vascular skeleton; wherein the central vascular skeleton preserves the connectivity and bifurcation topology of the original vascular structure; Based on the central skeleton of the blood vessel, the largest connected component is determined as the main skeleton, and the remaining connected components are determined as candidate branches. By using endpoint pairing and graph search path connection strategies, the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton.

5. The pancreatic artery image segmentation method according to claim 4, characterized in that, The three-dimensional refinement process includes: protecting endpoints with fewer than or equal to a target threshold number of foreground neighbor points within a preset neighborhood from being deleted; dividing the voxel space into multiple subdomains based on the parity of the three-dimensional coordinates, and sequentially deleting non-endpoints that meet the preset deletion conditions.

6. The pancreatic artery image segmentation method according to claim 5, characterized in that, The preset deletion conditions include: after deleting the non-endpoints, the local connectivity between the foreground and the background remains unchanged, and the Euler eigenvalues ​​before and after deletion remain unchanged.

7. The pancreatic artery image segmentation method according to claim 4, characterized in that, The method of using endpoint pairing and graph search path connection strategy to iteratively connect and verify the main skeleton and candidate branches to obtain the final main skeleton includes: Extract the endpoint coordinates of the main skeleton and candidate branches respectively, pair the endpoints of the main skeleton with the endpoints of the candidate branches based on the endpoint coordinates, and calculate the pairing cost of each endpoint pair. Endpoint pairs that satisfy the pairing constraints are determined as valid endpoint pairs; wherein, the pairing constraints include the Euclidean distance between the two endpoints being less than a preset distance threshold, and the angle between the skeleton tangent directions at the two endpoints being less than a preset angle threshold. The local spatial region determined by each effective endpoint pair is used as the search region, and a three-dimensional weighted graph is constructed within the search region based on a preset neighborhood. Based on a 3D weighted graph, with one endpoint of an effective endpoint pair as the starting point and the other endpoint as the ending point, and using the weighted sum of the negative log probability value and the path length as the edge cost, we search for the minimum cost path from the starting point to the ending point. The minimum cost path is used as the candidate connection path between the main skeleton and the candidate branch. The candidate connection path is then confirmed a second time, and the candidate connection path and the corresponding candidate branch that are confirmed to be reasonable are merged into the main skeleton. The endpoint pairing, path search, and verification merging process is iteratively executed until no new candidate connection paths are accepted, resulting in the final main skeleton.

8. The pancreatic artery image segmentation method according to claim 7, characterized in that, The pairing cost is determined based on at least one of the following factors: the Euclidean distance between the two endpoints, the angle between the tangent directions of the skeleton at the two endpoints, and the spatial relationship between the skeleton structures to which the two endpoints belong.

9. The pancreatic artery image segmentation method according to claim 7, characterized in that, The secondary confirmation includes: judging whether the connection path is reasonable based on the consistency between the end direction of the candidate connection path and the corresponding skeleton endpoint direction; if reasonable, then perform fusion update; if unreasonable, then re-perform endpoint pairing and path search.

10. The pancreatic artery image segmentation method according to claim 1, characterized in that, The process of obtaining an initial fusion mask based on the final main skeleton and the first confidence vessel mask, and performing connected component filtering and organ constraint filtering on the initial fusion mask to output the segmentation mask of the pancreatic artery, thereby achieving pancreatic artery image segmentation, includes: The final main skeleton is expanded using a preset spherical structural element, and then the expanded final main skeleton is fused with the first confidence blood vessel mask to obtain the initial fused mask. Connected component analysis is performed on the initial fusion mask to remove connected components with fewer voxels than the preset number of voxels. Combined with the preset auxiliary organ mask, connected components completely located within the auxiliary organ mask are deleted to obtain the segmentation mask of the pancreatic artery, thereby achieving pancreatic artery image segmentation.

11. A pancreatic artery segmentation system, characterized in that, The system includes: A multi-level confidence mask generation unit is used to preprocess the acquired image, input the preprocessed image into a pre-trained deep learning segmentation network, output a probability map, and generate a first confidence blood vessel mask and a second confidence blood vessel mask based on the probability map. The skeleton extraction and iterative connection unit is used to perform connected component analysis and three-dimensional refinement on the second confidence blood vessel mask to generate the central skeleton of the blood vessel. Based on the central skeleton of the blood vessel, the main skeleton and candidate branches are determined, and the main skeleton and candidate branches are iteratively connected and verified to obtain the final main skeleton. The fusion constraint and refinement output unit is used to obtain an initial fusion mask based on the final main skeleton and the first confidence vascular mask, and to perform connected component screening and organ constraint filtering on the initial fusion mask to output the segmentation mask of the pancreatic artery, so as to achieve pancreatic artery image segmentation.

12. The pancreatic artery segmentation system according to claim 11, characterized in that, The skeleton extraction and iterative connection unit includes: The connected component analysis module is used to perform connected component analysis on the second confidence blood vessel mask, obtain the number of voxels in each connected component, and filter the connected components based on a preset minimum voxel threshold, and obtain the second confidence candidate region based on the filtering results. A three-dimensional refinement module is used to perform topology-preserving three-dimensional refinement on the second confidence candidate region to extract the central skeleton of the blood vessels; wherein, the central skeleton of the blood vessels maintains the connectivity and bifurcation topology of the original blood vessel structure; The connected component determination module is used to determine the largest connected component as the main skeleton and the remaining connected components as candidate branches based on the central skeleton of the blood vessel. The iterative connection and verification module is used to perform iterative connection and verification of the main skeleton and candidate branches using endpoint pairing and graph search path connection strategies to obtain the final main skeleton.

13. The pancreatic artery segmentation system according to claim 11, characterized in that, The fusion constraint and refinement output unit includes: The initial fusion mask generation module is used to expand the shape of the final main skeleton using preset spherical structural elements, and then fuse the expanded final main skeleton with the first confidence blood vessel mask to obtain the initial fusion mask; The segmentation mask acquisition module is used to perform connected component analysis on the initial fusion mask, remove connected components with fewer than the preset number of voxels, and combine with the preset auxiliary organ mask to delete connected components that are completely located within the auxiliary organ mask, so as to obtain the segmentation mask of the pancreatic artery and realize pancreatic artery image segmentation.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images