Three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein image segmentation method and system

Abstract

The present invention relates to the field of image analysis technology, and in particular to a three-dimensional hepatobiliary, hepatic artery, portal vein and hepatic vein image segmentation method and system. The method comprises the following steps: acquiring multi-phase medical image data of the liver region; performing regional recognition on the image data to extract the hepatobiliary, hepatic artery, portal vein and hepatic vein regions; performing structural tracking on the hepatobiliary region to obtain hepatobiliary positioning results; performing improved contrast extraction on the hepatic artery region to obtain hepatic artery positioning data; performing directional segmentation analysis on the portal vein region to obtain portal vein positioning results; extracting branch angles of the hepatic vein region in combination with the portal vein direction to obtain hepatic vein positioning data; using the hepatic artery as a reference structure, elastically registering the above-mentioned vascular structures to generate registration data under unified coordinates; and completing three-dimensional visualization based on the registration results. The present invention realizes the joint segmentation and registration of multi-phase images, and has the technical advantage of strong structural recognition integrity.

Description

Technical Field

[0001] The present invention relates to the field of image analysis technology, and in particular to a method and system for segmenting three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein images. Background Art

[0002] With the rapid development of medical imaging equipment, multimodal imaging technologies such as computed tomography (CT) and magnetic resonance imaging (MRI) are increasingly being used in clinical diagnosis. This is particularly true in scenarios such as preoperative evaluation of liver disease, surgical planning for liver cancer / cholangiocarcinoma (especially hilar cholangiocarcinoma), and preoperative simulation for liver transplantation. These applications require higher precision and automation for the three-dimensional visualization and spatial positioning of key vascular structures within the liver, such as the hepatic artery, portal vein, hepatic vein, and bile duct. Traditional medical image segmentation methods rely heavily on image processing algorithms based on thresholding, region growing, and active contours. These methods struggle to adequately address the slender, branched, multi-directional, and grayscale-close characteristics of vascular structures in images. This is particularly true in microvessels, intersections, and areas with blurred or weak signals, often leading to segmentation breaks, positioning offsets, and structural misidentification. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention proposes a three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein image segmentation method and system to solve at least one of the above technical problems.

[0004] The present application provides a three-dimensional hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation method, comprising the following steps:

[0005] Acquiring multi-phase medical image data of a liver region;

[0006] Performing regional recognition on multi-phase medical image data to obtain hepatobiliary duct regional data, hepatic artery regional data, portal vein regional data, and hepatic vein regional data;

[0007] Perform structural tracking on the hepatobiliary region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data;

[0008] Elastic registration is performed based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; and three-dimensional visualization is performed based on the elastic registration data.

[0009] The present invention realizes multi-structure refined recognition based on multi-phase medical image data, and performs differentiated processing in combination with the relationship of liver anatomical structure. By performing structural tracking on the hepatobiliary regional data, the connectivity and integrity of the bile duct branches can be effectively enhanced; the boundary clarity and positioning accuracy of the hepatic artery can be improved through improved contrast extraction; the structural continuity of the portal vein branches can be enhanced through directional segmentation analysis; the branch angle extraction method is combined with the portal vein direction characteristics to effectively distinguish the hepatic vein from the portal vein, thereby improving the recognition accuracy. Through the elastic alignment operation between multiple structures, the unified spatial fusion of various vascular structures is achieved, which is convenient for three-dimensional visualization reconstruction and clinical auxiliary analysis. Compared with the traditional blood vessel segmentation method based on single-phase images or single enhancement processing, the spatial guidance strategy in the present invention is based on anatomical proximity, limits the search area of ​​the target structure, and effectively reduces the false detection rate; the direction modeling part enhances the model's perception of the continuity of the direction of slender blood vessels through main direction gradient calculation and direction field learning; the structure tracking module introduces a path cost function, using signal strength, direction consistency and proximity as constraints to ensure the coherent extraction of tiny branches; the elastic registration strategy uses the hepatic artery as a reference to construct a control point grid to flexibly align the positioning results of multiple structures, while maintaining the topological relationship and improving the accuracy of structural fusion. Overall, the present invention not only improves the extraction integrity and spatial expression clarity of vascular structures, but also ensures the visual coordination between multiple structures.

[0010] Preferably, the multi-phase medical image data includes arterial phase CT image data, portal venous phase CT image data, equilibrium phase CT image data, and T2-weighted MRI image data, and performing regional recognition on the multi-phase medical image data to obtain hepatobiliary duct regional data, hepatic artery regional data, portal vein regional data, and hepatic vein regional data, respectively, includes:

[0011] Performing coarse liver region segmentation on portal venous phase CT image data to obtain liver region data;

[0012] Determine the hepatic artery region according to the liver region data and the arterial phase CT image data to obtain the hepatic artery region data;

[0013] Hepatic vein region data is identified based on liver region data and equilibrium phase CT image data to obtain hepatic vein region data;

[0014] Directionally guided segmentation is performed based on the liver region data to obtain portal vein region data;

[0015] The hepatobiliary region is identified based on the liver region data and the T2-weighted MRI image data to obtain the hepatobiliary region data.

[0016] The present invention introduces arterial phase CT images, portal venous phase CT images, equilibrium phase CT images and T2-weighted MRI images, and combines coarse segmentation of the liver region with structure-specific processing to achieve targeted identification of the hepatic artery, portal vein, hepatic vein and bile duct regions. The hepatic artery region is identified by utilizing the vascular enhancement characteristics in the arterial phase image, and the liver parenchyma is segmented and portal vein directionality is extracted in combination with the portal venous phase image. The hepatic vein is separated based on the venous contrast in the equilibrium phase image, and the bile duct is identified by enhancing the bile duct signal using T2-weighted MRI images. Through multimodal and multi-phase fusion, the structural clarity, regional boundary accuracy and tissue separation ability of image segmentation are improved, effectively overcoming the problems of low recognition accuracy and organ confusion caused by the single phase and fuzzy structure in traditional methods, and achieving high-precision anatomical region extraction based on structure, image and strategy, which has significant medical application value and scalability.

[0017] Preferably, performing directionally guided segmentation based on the liver region data to obtain portal vein region data includes:

[0018] Perform voxel grayscale gradient extraction based on liver region data to obtain voxel grayscale gradient data;

[0019] Perform main direction clustering on the voxel gray gradient data to obtain main direction clustering data;

[0020] Perform tensor voting processing based on the main direction clustering data to obtain the vascular direction field data;

[0021] The preset direction-guided segmentation network model is used to segment the vascular direction field data to obtain the portal vein area data. The direction-guided segmentation network model is constructed by using the improved SD-UNet as the backbone network, and the attention map or guided flow is constructed as the output head.

[0022] The present invention extracts voxel grayscale gradients based on liver region data, and combines main direction clustering with tensor voting mechanism to construct vascular direction field, making full use of the directional consistency and anatomical coherence of the portal vein branch structure to achieve direction-aware driven segmentation modeling. The improved SD-UNet is used to construct a direction-guided segmentation network model, which integrates the direction field data and combines the attention map or guidance flow mechanism as the output guidance, which can effectively improve the responsiveness to thin branches and diversified directions during portal vein identification. Compared with traditional segmentation methods based on grayscale enhancement or static models, the present invention not only has stronger performance in segmentation accuracy and structural continuity, but also significantly enhances the model's efficiency in utilizing spatial directional features. It is suitable for identifying portal vein structures with slender branches and large bending variations. It has higher anatomical consistency and algorithm generalization capabilities, and is suitable for clinical applications and adaptive modeling needs of complex vascular systems.

[0023] Preferably, the step of constructing the preset direction-guided segmentation network model includes the following steps:

[0024] Obtain historical vascular direction field data and historical annotation data;

[0025] Perform multi-scale exaggerated convolution parallel processing based on historical vascular direction field data to obtain edge perception layer data;

[0026] The edge perception layer data is processed at the shallow edge layer and at the high-level feature layer to obtain shallow edge layer data and high-level feature layer data respectively. The shallow edge layer processing is performed through Sobel / LoG filter or edge channel convolution to enhance the boundary response.

[0027] Perform residual connection on shallow edge layer data and high-level feature layer data to obtain decoding layer data;

[0028] The vascular direction network is trained on the decoded layer data using historical annotated data to obtain direction-guided flow data. The vascular direction network training supervises the convolutional network by using the direction consistency loss combined with the Dice loss.

[0029] The direction-guided output of the direction-guided flow data is performed to obtain a direction-guided segmentation network model, wherein the direction-guided output is convolutionally fused with the decoding layer data and the direction confidence is calculated, or the direction-guided attention map is calculated through the direction-guided flow data, and the output result is regulated at the channel level.

[0030] The present invention constructs a direction-guided segmentation network model, integrates historical blood vessel direction field data and annotation data for network training, and has significant direction perception and structure recognition capabilities. Multi-scale exaggerated convolution is used to achieve structural perception at different scales, and edge channel convolution and Sobel / LoG filtering are used to enhance edge response, taking into account both details and semantic features. Then, through the residual connection between the shallow edge layer and the high-level feature layer, the clarity of the blood vessel boundary and the integrity of the branch structure are effectively maintained. The joint direction consistency loss and Dice loss are used for supervised training, so that the model can accurately fit the real blood vessel direction characteristics. Further, the channel-level regulation is performed by fusing the direction-guided flow with the decoding features, or generating a direction attention map, which significantly enhances the network's response to structural direction information. Compared with traditional image segmentation models, this solution has higher accuracy and robustness in the recognition of fine blood vessel branches, branch angle continuity modeling, and direction consistency maintenance, providing a more efficient and accurate model foundation for the intelligent recognition and three-dimensional modeling of detailed vascular systems.

[0031] Preferably, performing structural tracking on the hepatobiliary region data to obtain hepatobiliary positioning data includes:

[0032] Constructing a spatial guidance domain for the hepatobiliary region data to obtain the hepatobiliary spatial guidance domain data;

[0033] Perform local direction consistency enhancement based on the hepatobiliary spatial guidance domain data to obtain hepatobiliary enhancement data;

[0034] Performing path tracing based on the hepatobiliary enhancement data to obtain hepatobiliary duct tracking data, wherein the path tracing is to construct a cost function based on signal strength, direction consistency, and path proximity to perform minimum cost path tracing;

[0035] Connectivity optimization is performed on the hepatobiliary duct tracking data to obtain the hepatobiliary duct positioning data.

[0036] In the present invention, a spatial guidance domain based on the hepatobiliary region is constructed to limit the scope of structural recognition, and the local directional consistency enhancement method is combined to effectively enhance the structural coherence and directional characteristics in the bile duct image. By introducing a cost function of three factors: signal intensity, directional consistency, and path proximity, and performing a minimum cost path tracking operation, high-precision extraction of the bile duct trunk and fine branches can be achieved. Further combined with connectivity optimization processing, pseudo-branch structures and broken paths are effectively eliminated, thereby improving the continuity and accuracy of the overall tracking. Compared with the traditional method of extracting bile ducts based on thresholds or simple filtering, this method takes into account spatial guidance, directional modeling, and path optimization, significantly improving the structural recognition ability in cases of weak signals, complex directions, and small branches, and has stronger clinical adaptability and topological integrity, providing stable and reliable structural input for three-dimensional bile duct modeling and visualization applications.

[0037] Preferably, performing improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data includes:

[0038] Direction-selective enhancement kernel processing was performed on the hepatic artery region data to obtain the rough positioning data of the first hepatic artery;

[0039] The edge sensitivity of the hepatic artery region data is adjusted to obtain the rough positioning data of the second hepatic artery;

[0040] Attention perception enhancement is performed based on the coarse positioning data of the first hepatic artery and the coarse positioning data of the second hepatic artery to obtain weighted positioning data of the first hepatic artery and weighted positioning data of the second hepatic artery, respectively;

[0041] Confidence voting is performed based on the first hepatic artery weighted positioning data and the second hepatic artery weighted positioning data to obtain hepatic artery fusion positioning data, and centerline extraction is performed based on the hepatic artery fusion positioning data to obtain hepatic artery positioning data.

[0042] In the present invention, by introducing direction-selective enhancement kernel and edge-sensitivity regulation processing on the hepatic artery regional data, two structural enhancement pathways are constructed respectively, which can more accurately capture the direction and boundary features of the hepatic artery trunk and its small branches. The attention perception mechanism is used to perform weighted enhancement on the two types of coarse positioning results respectively, effectively improving the significance and noise suppression ability of the vascular response area. The results of the two enhancement pathways are further fused through the confidence voting mechanism, which strengthens the spatial continuity and positioning consistency of the structure while ensuring the recognition robustness. The centerline information is extracted based on the fusion results to achieve a structured expression of the hepatic artery path. Compared with traditional single-channel enhancement or fixed filtering extraction methods, this method has the advantages of multi-feature fusion, key enhancement, and strong adaptability to small blood vessels. It significantly improves the segmentation and positioning accuracy of the hepatic artery in complex backgrounds and low-contrast images, and is suitable for high-precision vascular modeling and intelligent analysis tasks.

[0043] Preferably, performing directional segmentation analysis on the portal vein region data to obtain portal vein positioning data includes:

[0044] Calculate the principal direction gradient map of the portal vein area data to obtain principal direction gradient map data;

[0045] Perform direction-guided segmentation according to the main direction gradient map data to obtain direction-guided segmentation data;

[0046] Branches are merged and breaks are repaired according to the direction-guided segmentation data to obtain portal vein positioning data.

[0047] In the present invention, the main directional gradient map of the portal vein area data is calculated to accurately capture the local directional characteristics of the blood vessel direction. Combined with the direction-guided segmentation operation, the model can fully perceive the spatial extensibility and direction consistency of the vascular structure during the segmentation process, effectively improving the recognition and connectivity maintenance capabilities of slender vascular branches. By performing branch merging and break repair on the direction-guided segmentation results, the topological integrity and structural continuity of the segmentation results are enhanced, and stronger structural recognition capabilities are possessed in the intersection area of ​​portal vein branches and weak signal areas. Compared with traditional segmentation methods that rely on grayscale or edge information, this method introduces a directional modeling mechanism, which has higher positioning accuracy and robustness when dealing with curved, slender, and multi-branched venous structures, and can more effectively support subsequent three-dimensional modeling and clinical auxiliary analysis tasks.

[0048] Preferably, extracting branch angles from the hepatic vein region data and the portal vein positioning data to obtain the hepatic vein positioning data includes:

[0049] Extract the centerline based on the portal vein positioning data to obtain portal vein direction map data;

[0050] Perform skeleton extraction based on the hepatic vein region data to obtain the hepatic vein branch data;

[0051] Perform linear fitting on the hepatic vein branch data to obtain hepatic vein branch fitting data, and calculate the angle of the nearest branch direction based on the hepatic vein branch fitting data to obtain branch angle characteristic map data;

[0052] Narrow neural network classification is performed based on the portal vein direction map data and branch angle feature map data to obtain the hepatic vein positioning data.

[0053] In the present invention, a branch angle discrimination mechanism for hepatic vein identification is constructed by extracting a centerline direction map based on the portal vein positioning result, and performing branch linear fitting and angle feature analysis in combination with the skeleton structure of the hepatic vein region. The angle relationship between the fitted branch direction and the portal vein direction is used to generate a branch angle feature map, and then the branch type is classified and determined through a narrow neural network to achieve accurate positioning of the hepatic vein. The present invention makes full use of the differences in branch direction and spatial orientation between the portal vein and the hepatic vein in the anatomical structure, improves the model's ability to identify and discriminate between the two types of venous structures, and has a higher classification accuracy in areas where structures are adjacent and signals overlap. Compared with traditional recognition methods based on single image segmentation or grayscale feature extraction, this method introduces a multi-feature intelligent analysis framework that combines structure fitting, angle modeling and lightweight neural networks. It has technical advantages such as strong structural understanding, low computational cost, and high topological recognition accuracy, providing more reliable technical support for the automatic anatomical recognition and three-dimensional modeling of detailed venous systems.

[0054] Preferably, performing elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain the elastic registration data includes:

[0055] Determine the hepatic artery positioning data as reference structure data;

[0056] Performing inter-structure pairing on the hepatobiliary duct positioning data, the portal vein positioning data, and the hepatic vein positioning data according to the reference structure data to obtain inter-structure pairing data;

[0057] The control point network is covered based on the paired data between structures to obtain the elastic registration data.

[0058] In the present invention, by using the hepatic artery positioning data as a reference structure, a unified spatial alignment benchmark between multiple vascular structures is established, and the positioning results of the bile duct, portal vein and hepatic vein are combined to perform structure pairing, which can fully utilize the anatomical proximity and spatial distribution patterns between the structures to construct an accurate structural correspondence. Based on the results of the structure pairing, an elastic registration operation is performed using a control point network coverage method, which can implement flexible spatial deformation on multiple structures while maintaining the continuity and topological relationship of the anatomical structure. Compared with traditional rigid registration or methods based on single structure guidance, the present invention effectively improves the spatial coupling degree and overall fusion quality between vascular structures through multi-structure collaborative drive and elastic deformation control point mechanism, especially in medical image scenes where multiple structures intersect, spatial dislocation or deformation is significant, it has higher structural alignment accuracy and three-dimensional modeling adaptability, providing a solid foundation for clinical precision visualization and personalized vascular reconstruction.

[0059] Preferably, a 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation system is characterized in that it is used to perform the 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation method described above, and the 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation system comprises:

[0060] A multi-phase image acquisition module, used for acquiring multi-phase medical image data of the liver area;

[0061] A region recognition module is used to perform region recognition on multi-phase medical image data to obtain bile duct region data, hepatic artery region data, portal vein region data, and hepatic vein region data;

[0062] The structural association positioning module is used to perform structural tracking on the hepatic artery region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; and perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data.

[0063] The three-dimensional structure reconstruction and display module is used to perform elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; and perform three-dimensional visualization operations based on the elastic registration data.

[0064] The beneficial effects of the present invention are: based on multi-phase medical image data, through modular processing methods with different structures and strategies, it is possible to achieve refined extraction and three-dimensional modeling of multiple key vascular systems in the liver. The bile duct is identified at the path level through structural tracking, the hepatic artery is enhanced with edge clarity and micro-branch positioning capabilities through improved contrast extraction, the portal vein introduces a directional segmentation analysis method to strengthen its structural continuity and branch integrity, and the hepatic vein is combined with the portal vein direction information to distinguish the branch angles and improve the accuracy of structural attribution judgment. Through multi-structure elastic registration and control point network construction based on the hepatic artery, unified spatial alignment and three-dimensional visualization expression of multiple vascular structures are achieved. Compared with the existing processing methods based on single structure or unified segmentation models, the present invention can make full use of the anatomical relationship characteristics between different image phases and structures, and has outstanding technical advantages such as accurate structural recognition, high registration robustness, and strong continuity of three-dimensional modeling, which significantly improves the anatomical restoration ability and intelligence level of clinical diagnostic assistance systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0066] Figure 1 A flowchart showing the steps of a method for segmenting three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein images according to one embodiment is shown;

[0067] Figure 2 A flowchart showing the steps of a multi-phase image region recognition method according to an embodiment is shown;

[0068] Figure 3 A flowchart of a multi-region structure association positioning method according to an embodiment is shown. DETAILED DESCRIPTION

[0069] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative work are within the scope of protection of the present invention.

[0070] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0071] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0072] See also Figures 1 to 3 The present application provides a three-dimensional hepatobiliary, hepatic artery, portal vein and hepatic vein image segmentation method, comprising the following steps:

[0073] S1. Acquire multi-phase medical image data of the liver region;

[0074] In one embodiment, the multi-phase medical image data described herein includes a multimodal image sequence from the same subject, specifically including arterial phase CT image data, portal venous phase CT image data, equilibrium phase (delayed phase) CT image data, and T2-weighted MRI image data. The CT image data is acquired using a spiral CT scanner with 64 or more slices, such as the GE Revolution CT (256-slice) clinical-grade spiral CT scanner. Each acquired image slice is no thicker than 1 mm. The scanning protocol is as follows: arterial phase scan: 20 seconds after intravenous injection of iodinated contrast agent; portal venous phase scan: 60 seconds after injection; equilibrium phase (delayed phase) scan: 180 seconds after injection. Scanning parameters include slice thickness of 1 mm, helical pitch of 1.375, tube voltage of 120 kVp, and automatic tube current regulation. T2-weighted MRI image data is acquired using a 3.0T high-field MRI scanner, such as the GE SignaPioneer 3.0T system or the Siemens MAGNETOM Skyra 3.0T system. A transverse fast spin echo (FSE) sequence was used with the following parameters: echo time (TE) of 80–100 ms, repetition time (TR) of 4000–6000 ms, slice thickness no greater than 3 mm, and matrix resolution no less than 256 × 256, to ensure sufficient soft tissue contrast for identification of hepatobiliary structures.

[0075] S2. performing region recognition on the multi-phase medical image data to obtain hepatobiliary duct region data, hepatic artery region data, portal vein region data, and hepatic vein region data;

[0076] In one embodiment, a three-dimensional deep convolutional neural network model based on the U-Net++ structure is used to construct independent semantic segmentation models for key anatomical structures of the liver. The structures involved include the bile duct, hepatic artery, portal vein, and hepatic vein. The U-Net++ model (also known as Nested U-Net) introduces a deep supervision mechanism and dense skip connections. The network input is a three-dimensional cubic image block (3D patch) with a size of Voxels are processed by grayscale normalization and voxel size resampling. The output is a structural probability map, which is the predicted probability of each voxel for its class. To improve segmentation accuracy and anatomical structure differentiation, different anatomical structures are modeled primarily based on their optimal visualization phase images: the hepatobiliary duct model is primarily based on delayed phase images; the hepatic artery model is primarily based on arterial phase images; the portal vein model is primarily based on portal venous phase images; and the hepatic vein model is primarily based on T2-weighted MRI image data. During model training, a weighted combination of the Dice coefficient loss and the focus loss function is used as the overall loss function to balance the overlap of foreground classes and focus on difficult-to-classify samples, significantly alleviating the class imbalance problem.

[0077] S3. Perform structural tracking on the hepatobiliary region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data;

[0078] In one embodiment, the operator manually selects the voxel with the highest grayscale value within the first 20 slices of the image center as the starting point of the hilar region. This point is typically located at the entrance of the common bile duct or common hepatic duct, but can also be automatically identified based on the anatomical model. Starting from this initial point, a forward propagation distance field based on a velocity function is constructed using the Fast Marching method to provide overall energy distribution guidance for path search. A minimum-cost path search algorithm (such as Dijkstra) is used within the distance field to extract the shortest energy path from the initial point to the end of each branch, reconstructing the tree-like centerline structure of the bile duct. The operator manually selects the voxel with the highest grayscale value within the first 20 slices of the image center as the starting point of the hilar region. This region corresponds to the entrance of the common bile duct or common hepatic duct, and can also be automatically identified based on the anatomical structure. Growth constraints are set: neighborhood directional consistency > 0.7. Structural consistency between adjacent voxels is calculated using the principal directions of the Hessian matrix. This ensures that the tracking path grows continuously along the blood vessels / bile ducts and suppresses lateral mis-propagation. The segmentation probability map value is > 0.4; this probability map is derived from the structural probability map output by the aforementioned U-Net++ model and indicates the confidence that the current voxel belongs to the hepatobiliary structure. The output is the centerline data of the 3D hepatobiliary tree structure, represented as a spatial curve composed of multiple key points.

[0079] The input images were normalized arterial phase contrast-enhanced CT images. The images were grayscale normalized and voxel size resampled to a uniform 1 × 1 × 1 mm³. A three-dimensional Frangi filter was applied to the arterial phase images to enhance the local contrast of slender, tubular structures; the filter parameters were set to a multiscale range. =0.5 to 2.5, with a step size of 0.5; Frangi filter analyzes the local structure morphology through the eigenvalue of the Hessian matrix and has a good response to linear structures. In the enhanced image, the grayscale mean within the local window is used and standard deviation Construct a dynamic threshold; if the gray value of a voxel , it is judged as a foreground voxel; this method can effectively deal with the problems of uneven image brightness and local contrast differences. The above segmentation results are logically ANDed with the liver region mask to retain only the high-response areas within the liver; the liver region mask is obtained by the above liver segmentation model. Connected domain analysis is performed on the cropped three-dimensional segmentation results; only the top three largest connected regions with a voxel number greater than 500 are retained to remove discrete noise and pseudo-response areas; the output three-dimensional mask represents the main trunk of the hepatic artery and its significant branch structures. The SimpleITK library is used for three-dimensional image filtering and segmentation, combined with NumPy for numerical calculations and connected domain analysis. The above tools are all open source scientific computing libraries suitable for medical image processing scenarios.

[0080] Directional filtering was performed on the portal venous phase images of enhanced CT images using an adjustable directional filter bank with a set number of directions, N = 12. This step enhances linearly distributed vascular structures in the image and obtains the response intensity of each pixel in each direction. For each pixel, the directional angle corresponding to the maximum value of the filtered response in 12 directions was recorded, thus forming a directional vector field for the entire image. The directional field was clustered using the K-Means clustering algorithm (K = 3), and the most frequently occurring direction was extracted as the dominant direction. Using the dominant direction as a guide, a 16×16 pixel, directionally aligned 2D slice window was constructed. A trained 2D convolutional neural network (2D CNN) was used to perform a binary classification of the central pixel of each window into portal vein / non-portal vein, thereby identifying and segmenting the portal vein path. The identification results were reconstructed into a 3D voxel structure based on spatial location to generate a portal vein segmentation mask. Simultaneously, path tracing was performed based on the dominant direction to obtain the directional centerline trajectory of the portal vein. The output consists of two parts: a three-dimensional portal vein voxel mask, which is used to represent the spatial distribution of the portal vein structure; and a central path extracted based on the dominant direction, which is used to represent the main direction of the blood vessel.

[0081] Arterial phase images and portal vein path data are input; the portal vein centerline is obtained by the aforementioned U-Net++ segmentation model and minimum path tracing method; the image used is the venous phase image (obtained after 60 to 90 seconds of enhancement delay). Based on the three-dimensional path of the portal vein centerline, all branch nodes with trifurcated or multi-forked structures are identified; the branch point is determined based on the degree of connection of the path node (degree) ≥ 3, or the direction change exceeds a specific angle threshold. A spherical neighborhood area is constructed at each branch point, with a radius set to 8 voxels, covering all branch structures; the remaining vascular paths and their main direction vectors are extracted within the area. The three-dimensional angle between the main branch direction of the portal vein and the directions of other paths in the neighborhood is calculated (based on the vector cosine); if the angle is between 40° and 80°, and the path length is not less than 20 voxels, it is marked as a candidate hepatic vein path. At each branch point Construct a spherical space neighborhood with a radius of 8 voxels around Extract all path segments within the neighborhood and calculate their main direction vectors; calculate the three-dimensional angle between the neighborhood path and the main trunk direction; if the following conditions are met, it is marked as a candidate hepatic vein path, and the angle range is: Path length ≥ 20 voxels; the end coordinate point is located at the top of the liver (such as the slice number Or near the outlet of the inferior vena cava (e.g., within 3 cm of the center coordinates of the superior vena cava segment of the hepatic vein). A 3D ridge extraction algorithm is used to refine the candidate path. A graph optimization method based on minimum cost graph cuts is used to extract the precise center path, eliminating excess noise or detours. The output is a 3D centerline representation of the main hepatic vein path.

[0082] S4. Perform elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; and perform three-dimensional visualization based on the elastic registration data.

[0083] In one embodiment, an elastic registration technique based on the B-Spline non-rigid deformation model is used to align multiple vascular structures (including the hepatic artery, portal vein, hepatic vein, and bile duct) in liver medical images to a unified spatial reference coordinate system. The registration process is implemented using the open-source image processing platform Elastix v4.9. Specifically, a fixed mask image of the portal vein structure is used as the reference coordinate system for spatial registration. Mask images of the hepatic artery, bile duct, and hepatic vein structures are used as registration targets. A non-rigid deformation model based on B-Spline is used to support local elastic alignment. A grid spacing of, for example, 32 pixels is set to control the fineness of the deformation grid. The primary metric function is the mutual information (MHI) used to compare the intensity distributions between grayscale images. The auxiliary metric, the structural similarity index (SSIM), is used to improve structural alignment accuracy. An adaptive stochastic gradient descent optimizer is used for metric calculation. A pyramid registration method is employed, with three resolution levels optimized from low resolution to the original resolution. This algorithm is based on Elastix v4.9 and supports extensions using open-source platforms such as Advanced Normalization Tools (ANTs). It outputs deformation masks for the hepatic artery, bile duct, and hepatic vein in a unified spatial coordinate system, while also generating the corresponding 3D non-rigid deformation field.

[0084] VTK is used to read segmentation mask data. The Marching Cubes isosurface reconstruction algorithm is applied to each structure type to generate a corresponding triangular mesh (PolyData). Each structure is assigned a unique color code, such as hepatic artery: red; portal vein: blue; hepatic vein: green; and bile duct: yellow. The reconstructed structures are exported to standard 3D model formats (such as STL or OBJ). The front-end uses the WebGL-based Three.js library for model loading and rendering. Interactive control functions include 3D rotation and scaling, cross-section browsing, model transparency adjustment, mouse dragging, positioning, and structure selection. Auxiliary functions include hovering structure name prompts, vascular path visualization, and distance, angle, and vessel diameter measurement tools. The back-end uses the Flask framework to read and interface model data. The front-end uses WebGL for model rendering, and the front-end and back-end communicate via HTTP. This allows for online browsing of segmentation results without client installation, allowing doctors and researchers to access structural models from multiple devices.

[0085] Preferably, the multi-phase medical image data includes arterial phase CT image data, portal venous phase CT image data, equilibrium phase CT image data, and T2-weighted MRI image data, and performing regional recognition on the multi-phase medical image data to obtain hepatobiliary duct regional data, hepatic artery regional data, portal vein regional data, and hepatic vein regional data, respectively, includes:

[0086] S21, performing coarse liver region segmentation on the portal venous phase CT image data to obtain liver region data;

[0087] In one embodiment, portal venous phase enhanced CT images are used as input data. This phase is acquired approximately 60 seconds after contrast agent injection, when the contrast between the liver parenchyma and surrounding tissues is clearest. The grayscale intensity of the image is limited to [-100, 300] Hounsfield units (HU) to remove abnormal low-density and high-density values ​​and enhance the grayscale contrast of the liver parenchyma tissue. The image is isotropically voxel-resampled to a uniform spatial resolution of To adapt the network input and enhance feature consistency, a 3D U-Net architecture was used for semantic segmentation of the liver region. The network input was a preprocessed, standardized CT image, and the output was a probability map. The network training data included the publicly available Liver Tumor Segmentation Challenge (LiTS) dataset and manually annotated clinical samples constructed for this project. Labels only included the entire liver region, excluding intrahepatic vascular and bile duct structures to ensure clear mask boundaries. A 3D connected domain analysis was performed on the segmentation results, retaining only the regions with the largest number of voxels and removing false positive artifacts or non-liver region responses. A morphological closing operation was performed on the mask, using a spherical structuring element (radius 2-3 voxels) to fill small edge holes and smooth boundary curves. The output was a 3D binary mask image (mask) representing the liver region. This mask served as a spatial cropping reference for the extraction of vascular, bile duct, and tumor structures, improving the accuracy and computational efficiency of structural localization.

[0088] S22, determining the hepatic artery region according to the liver region data and the arterial phase CT image data to obtain hepatic artery region data;

[0089] In one embodiment, the input is an arterial phase CT image, in which the hepatic artery structure shows a high contrast enhancement state. The obtained liver mask is called to perform spatial cropping on the CT image, retaining only the image voxels in the liver area and excluding the interfering structures outside the liver. The Frangi vessel enhancement filter is applied to the cropped image, and a multi-scale strategy is used to enhance the image quality. Enhances vascular structures with tubular features. Apply a preset threshold to the Frangi filter result. , set to the image mean plus 1.5 times the standard deviation , to segment the vascular regions with strong responses. The vascular segmentation results are refined using 3D morphology to extract a one-pixel-wide vascular centerline path. Using the centerline point closest to the center of the liver portal region as the starting point, centerline-based vascular structure tracking and growth are performed. The extracted vascular network is structurally screened to eliminate false-positive vascular segments that are not connected to the liver region. Only the main trunk and its vascular branches within two levels are retained, limiting the retention to the structural path from the trunk to the third-level node.

[0090] S23, identifying the hepatic vein region based on the liver region data and the equilibrium phase CT image data to obtain hepatic vein region data;

[0091] In one embodiment, an equilibrium-phase CT image, acquired approximately 120-180 seconds after contrast agent injection, is input. In this equilibrium-phase image, the density contrast between the venous vessels and the liver parenchyma is significant, making it suitable for hepatic vein identification. A three-dimensional mask of the liver region (liver region data) is simultaneously loaded as a basis for spatial cropping. The liver mask is used to crop the original image, retaining only image voxels within the liver region and excluding interference from non-liver regions. A three-dimensional steerable filter (3DSteerable Filter) or a multi-directional Gabor filter is applied to the cropped image. The filter enhances the response of structures with linear and tubular features, highlighting the direction of the hepatic veins. The enhanced image is then subjected to the Otsu adaptive threshold segmentation algorithm to produce a binary image. The Otsu method automatically selects the optimal segmentation threshold by analyzing the image histogram and is suitable for bimodal distribution scenarios. All connected regions were analyzed for features, and candidate hepatic vein regions that met the following criteria were retained: the principal axis was close to the Z-axis (vertical axis), meaning the angle between the principal structural direction and the Z-axis was less than 30°; the number of branches was greater than two, calculated using skeleton extraction and node statistics; and the region center was close to the liver apex, meaning the Z coordinate of the region center was within the top 10% of voxels of the liver mask. Starting from the selected apical vein region, a Fast Marching path tracing algorithm was employed. A cost map was constructed, using the inverse function of the vascular response value to trace the path to the underlying liver interior. A continuous path was extracted, and the centerlines of the main hepatic vein trunk and its branches were generated. The output was a hepatic vein region mask, a 3D binary image with the vein region marked, and a structural center path map, representing the 3D centerlines of the main hepatic vein trunk and its branches, represented as a set of continuous coordinate points.

[0092] S24, performing direction-guided segmentation based on the liver region data to obtain portal vein region data;

[0093] In one embodiment, the input is a portal venous phase CT image and a corresponding liver region mask. Within the area defined by the liver mask, the Hessian matrix of each voxel is calculated based on the image intensity values, and the principal curvature direction is obtained to construct a voxel-level directional gradient vector field. This direction field is used to represent the main orientation of the local vascular structure. A spherical region of interest (ROI) is defined in the portal region (the starting point of the portal vein trunk) as the starting point for path tracing. This region can be automatically generated based on anatomical location or physician annotation. With the starting point as the center, a path tracing operation is performed along the principal direction (e.g., the direction of maximum curvature) of each voxel in the direction field. The path tracing method uses a tube tracking algorithm or a flux voting method based on directional consistency. During the tracing process, a path map of the portal vein trunk is gradually established. Using the path map as a guiding skeleton, a local region growing operation is performed in combination with the grayscale features of surrounding voxels to obtain the overall segmentation result of the portal vein. The following enhancement strategies are used in the region growing process to improve accuracy and robustness: local maximum intensity projection is introduced in the path tracing region to enhance the continuity of small vessels and lateral tributaries; the cosine similarity threshold between the region growing direction and the local main direction is set. , restricting growth to the main axis of the vessel to prevent erroneous segmentation into non-vascular areas; constructing a preliminary path map to represent the main portal vein. Output includes a 3D voxel mask of the portal vein region; a path map of the portal vein center obtained by tracing; and the portal vein tributary structure, represented as a tree path or spatial point set.

[0094] S25. Perform hepatobiliary region identification based on the liver region data and the T2-weighted MRI image data to obtain hepatobiliary region data.

[0095] In one embodiment, the hepatobiliary region is identified and extracted based on T2-weighted magnetic resonance imaging (MRI) images and liver structural data. The input image is a T2-weighted MRI image, which exhibits good signal contrast in this sequence and is therefore suitable for bile duct extraction. The T2 MRI image is spatially aligned to the reference coordinate system of the portal venous phase CT image using rigid registration. The liver mask in the CT image is spatially transformed and mapped to the MRI image. Based on this, MRI voxels of the liver region are cropped to exclude extrahepatic interfering tissue. Local contrast enhancement and grayscale normalization, including histogram equalization and histogram matching, are applied to the cropped MRI image. A three-dimensional convolutional neural network (such as ResUNet3D) is used for automatic segmentation of bile duct structures. This network employs an encoder-decoder architecture, integrating multi-scale features with residual connections, making it suitable for the accurate extraction of elongated structures such as the hepatobiliary region. During the model training phase, independent labels are assigned to the bile duct and gallbladder regions to address the problem of similar signal quality and easy confusion between the bile duct and gallbladder in T2 images. This improves the model's discriminative ability and reduces missegmentation. Post-processing is performed on the segmentation results, including analyzing whether the bile duct structure is connected to the hepatic portal area to identify valid branches; removing isolated, broken, or artifact areas caused by noise interference to improve the anatomical rationality and traceability of the segmented structure. The output is a 3D voxel mask of the hepatobiliary duct, resulting in hepatobiliary region data.

[0096] Preferably, performing directionally guided segmentation based on the liver region data to obtain portal vein region data includes:

[0097] Perform voxel grayscale gradient extraction based on liver region data to obtain voxel grayscale gradient data;

[0098] In one embodiment, a portal venous phase enhanced CT image is used as input data. This image has undergone standard preprocessing steps, including grayscale cropping, normalization, and voxel resampling. The resulting three-dimensional mask of the liver region is used as the processing region, and voxel gradients are calculated only within the liver. A three-dimensional Sobel operator or an equivalent first-order differential convolution kernel is used to perform convolution operations in the x, y, and z spatial dimensions. The gradient component in each direction is calculated using the following formula: , , , for The rate of change of intensity in the direction, is the grayscale voxel value of the 3D CT image, To calculate The first-order differential convolution kernel of the directional gradient, for The rate of change of intensity in the direction, To calculate The first-order differential convolution kernel of the directional gradient, for The rate of change of intensity in the direction, To calculate First-order differential convolution kernel of directional gradient. Calculate the gradient vector of each voxel point and gradient magnitude , Obtain a gradient vector map, including the three-dimensional gradient value of each voxel in the liver area , expressed as a vector field, and a gradient magnitude map, i.e., a gradient modulus map , characterizing the amplitude of local intensity variation.

[0099] Perform main direction clustering on the voxel gray gradient data to obtain main direction clustering data;

[0100] In one embodiment, voxel grayscale gradient data is clustered by selecting voxel points whose gradient amplitude is greater than the 95th percentile of the entire image; all direction vectors participating in the clustering are normalized and converted into unit vectors. The unit vector calculation method is: ,in, For the The unit direction vector of the point, For the The direction vector of a point, For the The magnitude of the direction vector of the point, , For the The direction vector of the point is The direction of the component, For the The direction vector of the point is The direction of the component, For the The direction vector of the point is The spherical K-means algorithm is used to cluster unit direction vectors. This algorithm is applicable to unit sphere data and can efficiently extract multiple main directions. Its optimization goal is to maximize directional consistency, that is, to maximize the sum of the cosine similarities between all direction vectors and the direction of the cluster center to which they belong: , For the The unit direction vector of the voxel, is the cluster center direction vector, the cluster center direction vector to which the point belongs. The dot product represents the cosine value between the two directions. The larger the value, the more consistent the direction. For clustering index, specify Which direction the cluster belongs to. Number of clusters It can be set to 4 to 6 according to the actual vascular branch structure, which is enough to cover the main trunk and multi-level branch directions, thereby extracting The output includes a direction cluster label for each voxel participating in the cluster, indicating the category to which its dominant direction belongs; each direction class represents the main direction of a potential vascular structure.

[0101] Perform tensor voting processing based on the main direction clustering data to obtain the vascular direction field data;

[0102] In one embodiment, each main direction clustering data is initialized as a symmetric second-order tensor ,in is the gradient direction, is the transposed symbol; within the neighborhood window (such as 5×5×5), weighted propagation voting is performed, and each voxel obtains a weighted structure tensor, where the weighted propagation voting is: for point The weighted structure tensor of for point The neighborhood set of is a Gaussian weighting function, the farther the distance, the smaller the weight. for point and neighboring points The Euclidean distance between for point The original structure tensor at , , is the weight of tensor propagation, the closer the point , the greater the weight, is the natural exponential term, is the distance value, equal to , To control the scale parameter (window scale) of the Gaussian distribution width; perform eigenvalue decomposition on the tensor and extract the main direction: , is the structure tensor, For the eigenvalue decomposition operation, the tensor (or matrix) Perform eigenvalue decomposition, is the principal eigenvalue (largest), is the secondary eigenvalue, is another minor eigenvalue; by the ratio of the main eigenvalue to the total energy , and obtain the direction consistency index; the vascular direction field data includes the main direction vector field, and the main direction corresponding to each voxel constitutes the vascular direction vector field; the direction consistency map, the direction consistency index of each voxel, characterizes its prominence in the vascular structure.

[0103] The preset direction-guided segmentation network model is used to segment the vascular direction field data to obtain the portal vein area data. The direction-guided segmentation network model is constructed by using the improved SD-UNet as the backbone network, and the attention map or guided flow is constructed as the output head.

[0104] In one embodiment, the encoder part introduces a Dense Block module in each level of convolutional layer to enhance feature extraction capabilities; a spatial attention module is inserted into the jump connection to improve the spatial feature alignment effect. The input of the network consists of three channels: Channel 1: portal venous phase CT image; Channel 2: vascular direction field image (the direction vector can be encoded as a three-channel RGB or floating-point tensor); Channel 3: structural consistency score map, which is used to describe areas with prominent directional structures. The network output contains two branches. The main output branch generates a three-dimensional segmentation mask of the portal vein area through a Sigmoid function; the guided output branch generates a directional guidance map, which is used to enhance directional perception during training or to visualize subsequent results. During training, the loss function is in the following joint form: is the total loss, is the Dice coefficient loss, is the binary cross entropy loss, is the weighting factor of the directional loss, is the cosine similarity loss between direction vectors. The training parameters are set as follows, optimizer: ; Initial learning rate: Batch Size: 2; Input patch size: 64×64×64; Number of training epochs: 100-150. The output is a 3D portal vein mask image, i.e., portal vein region data.

[0105] Preferably, the step of constructing the preset direction-guided segmentation network model includes the following steps:

[0106] Obtain historical vascular direction field data and historical annotation data;

[0107] In one embodiment, historical vascular orientation field data is derived from the directional structure extraction results of existing annotated images; each sample includes the original CT image and the extracted vascular structure tensor orientation field (e.g., the main orientation map obtained through tensor voting and eigenvector calculation); the historical annotation data is a manually or semi-automatically annotated portal vein structure region mask (3D voxel label).

[0108] Perform multi-scale exaggerated convolution parallel processing based on historical vascular direction field data to obtain edge perception layer data;

[0109] In one embodiment, the input feature map undergoes three-dimensional convolution operations through three parallel paths, with convolution kernel sizes of 3×3×3, 5×5×5, and 7×7×7, respectively. Each path consists of a three-dimensional convolution layer, a batch normalization layer, and a nonlinear activation function (ReLU) to extract feature information under different receptive fields. The feature maps output by the three scale paths are spliced ​​in the channel dimension to form a fused multi-scale feature map. Subsequently, a 1×1×1 convolution layer is used for channel compression and feature fusion to reduce redundancy while maintaining structural representation capabilities.

[0110] In order to enhance the directional perception, a directional weighting mechanism is introduced in the above convolution process, such as obtaining the spatial direction vector of the current voxel point in the sliding window area of ​​each convolution kernel. , the direction comes from the aforementioned direction field calculation module; the response to the neighboring direction position that is consistent (or approximately consistent) with the direction vector is enhanced, for example, the main direction channel is multiplied by the weighting coefficient ; Maintain or suppress the responses in other directions, thereby simulating high-pass excitation behavior and strengthening the response of the image grayscale or structural mutation area. Let the current voxel position be , the position to be calculated in the convolution window is , and its directional consistency weight is defined as: , is the direction consistency weight factor, is the directional weighting coefficient (0.5~1.0), To maximize the function, is the main direction vector of the current voxel, is the current voxel (convolution center) position, is any neighborhood voxel position within the convolution window, For point to The unit direction vector of for The model, For point to The modulus of the unit direction vector; the direction-weighted convolution response is expressed as is the direction weighted convolution result, is any neighborhood voxel position within the convolution window, For The convolution window range centered on is the direction consistency weight factor, The convolution kernel is offset The value of For the input feature map The output is a set of edge-aware feature maps that fuse directional features, namely edge-aware layer data.

[0111] The edge perception layer data is processed at the shallow edge layer and at the high-level feature layer to obtain shallow edge layer data and high-level feature layer data respectively. The shallow edge layer processing is performed through Sobel / LoG filter or edge channel convolution to enhance the boundary response.

[0112] In one embodiment, the input voxel image In three directions Perform first-order gradient calculation on : ,in For The first-order gradient response in the direction, is the input edge perception layer data, for The first-order gradient convolution kernel in the direction, For The first-order gradient response in the direction, for The first-order gradient convolution kernel in the direction, For The first-order gradient response in the direction, for First-order gradient convolution kernel in the direction. Calculate edge response value , is the gradient magnitude map, the gradient magnitude map It will be used as the input of the shallow edge channel to significantly enhance the boundary area. The 3D LoG filter is used to perform a second-order differential operation on the image to respond to sharp structural changes, enhance the boundary signals of small blood vessels and weak response areas, and effectively improve the precision of structure detection, thereby constructing an edge guidance channel. (It can be a gradient map, LoG response, etc.); connect it in parallel with the backbone feature as the channel-level attention weight to guide the regional response of the deep feature; focus on the edge area of ​​the structure and obtain shallow edge layer data.

[0113] To obtain the global contextual features of vascular structures, a 3D version of the ResNet-18 encoder was used as the backbone extraction network: the encoding path contains four residual blocks, and each layer uses stride=2 convolution for downsampling to form a multi-level semantic abstraction; each block consists of two sets of convolution + normalization + ReLU + residual connection to improve the robustness of structure recognition; after encoding, the ASPP module is connected: hollow convolution kernels with expansion rates of 1, 6, 12, and 18 are used in parallel; the receptive field is expanded, the ability to capture large-scale structures is enhanced, and high-level feature layer data is obtained.

[0114] Perform residual connection on shallow edge layer data and high-level feature layer data to obtain decoding layer data;

[0115] In one embodiment, shallow edge layer data and high-level feature layer data are element-by-element added or concatenated in the channel dimension. To unify the fused feature dimensions, a 3×3×3 three-dimensional convolution operation is performed on the fused features. An attention mechanism module (such as an SE module or a spatial attention module) is introduced to enhance the response to significant edge features and improve the ability to integrate edges with global semantics. The fused feature map has the following properties: it preserves vascular edge details in the shallow input, such as the comb-like structure outline of capillaries; it integrates the global direction and structural semantics in the high-level output, which helps maintain the connectivity and complete identification of deep-layer vessels. The fused decoding layer feature map is output as the main input feature for the network's direction output branch and structural segmentation branch.

[0116] The vascular direction network is trained on the decoded layer data using historical annotated data to obtain direction-guided flow data. The vascular direction network training supervises the convolutional network by using the direction consistency loss combined with the Dice loss.

[0117] In one embodiment, a strategy for directional supervised training based on historically annotated data is used. A directional guidance stream output header is introduced at the end of the decoder to generate directional vector information corresponding to each voxel, which serves as a key input for structure recognition or guided path growth. Two types of supervisory information are used: a structure mask, such as using standard Dice Loss to optimize the main branch structure mask; and a directional consistency map, such as using historical expert annotated data to extract the main structural direction vector as a directional supervision label (directional vector map or tensor consistency map) to guide the directional branch output. The total loss function consists of the structural loss and the directional loss, and the weighted combination is as follows: , is the total loss, is the structural consistency loss, is the direction supervision weight, and the recommended value range is 0.3-0.5. is the direction consistency loss, where the structural consistency loss is: ,in represents the predicted structure mask, Represents the true structure mask, which represents the number of voxels. By maximizing the overlap area between the predicted result and the true label, It can effectively guide the network to learn the spatial distribution of vascular trunks and thin branches, and characterize the extraction integrity and accuracy of structural regions. The directional consistency loss is the cosine distance loss between directional vectors: ,in is the direction vector predicted by the network, is the label direction vector. To simultaneously generate the structure mask and directional flow, a dual output head is designed at the end of the network decoder. The main output head uses a sigmoid activation function to output a structure mask map for binary segmentation; the guided flow output head outputs a directional guided flow map, where each voxel position contains a unit 3D direction vector.

[0118] The direction-guided output of the direction-guided flow data is performed to obtain a direction-guided segmentation network model, wherein the direction-guided output is convolutionally fused with the decoding layer data and the direction confidence is calculated, or the direction-guided attention map is calculated through the direction-guided flow data, and the output result is regulated at the channel level.

[0119] In one embodiment, two methods are selected or combined: Method 1: 3D convolution fusion of the direction guidance flow data G and the decoding layer feature D: encoding the direction guidance flow data G into the direction convolution kernel parameter, which is used to perform local convolution weighting on the decoding layer feature D; the output is the direction confidence map , indicating the consistency of each voxel direction (for example: 0.85 = highly certain that the blood vessel runs in this direction). The direction confidence is used as an enhancement factor for the mask channel to improve the network focusing accuracy. The direction confidence map is used as a weight adjustment factor for the mask map generation module. ,in is the activation function, is a three-dimensional convolution kernel, which is a preset size such as , For the blood vessel direction field data, enhance the response weight of the region with consistent direction and output the segmentation result , thus obtaining the direction-guided segmentation network model.

[0120] Method 2: Direction-guided flow data is input into the directional attention module; the weight map of each channel is output , For the The weight coefficient of the channel indicates the degree to which the channel is consistent with the direction guidance. For directing flow from the direction The function module for calculating the weight of each channel, is the first channel index, is the total number of channels of the feature map, and performs channel-level guidance on the output mask map , It is a multi-channel spatial feature map, that is, direction-guided flow data; it is equivalent to letting the network select a channel with more consistent direction as the main basis for output, and obtaining a feature map after direction-guided attention regulation. . The feature map after adjustment The segmentation result with enhanced direction is output by 3D convolution to obtain the direction-guided segmentation network model.

[0121] Preferably, performing structural tracking on the hepatobiliary region data to obtain hepatobiliary positioning data includes:

[0122] Constructing a spatial guidance domain for the hepatobiliary region data to obtain the hepatobiliary spatial guidance domain data;

[0123] In one embodiment, a spatial coordinate tensor is constructed to express the normalized relationship of spatial positions. For each voxel, it is normalized according to the image size to: , the resulting three-channel tensor records the spatial coordinates of each position, and the range is normalized to Extract the central region of the spatial guide. According to the mask , calculate its voxel density centroid , morphological center point extraction method can also be used as a means to enhance stability. The construction radius is The three-dimensional spherical region of interest is used as the starting area for path tracing. Construct the spatial guidance domain tensor , which is defined as: is the spatial guide domain value, is the distance field weight coefficient, which represents the weight of the distance field from the current voxel to the center point, which is 1.0. is the distance field from the current voxel to the center point, is the edge gradient map weight, which represents the weight of the mask boundary gradient map. is the mask boundary gradient map (extracted using Sobel or Laplacian), which is 0.5. The spatial guidance domain tensor It will be used for directional consistency enhancement processing and path cost function modeling to provide spatial prior constraints for bile duct structure tracking.

[0124] Perform local direction consistency enhancement based on the hepatobiliary spatial guidance domain data to obtain hepatobiliary enhancement data;

[0125] In one embodiment, for any voxel point in the hepatobiliary spatial guidance domain data , calculate its gradient vector ,in is the original structure tensor, is the image gradient vector (three-dimensional), Transpose the image gradient vector (three-dimensional), and then construct the corresponding structure tensor: A Gaussian kernel is used for weighted smoothing in a 5×5×5 three-dimensional neighborhood. , is a smooth tensor, is the three-dimensional Gaussian weight, for Index in direction, for Index in direction, for Index in the direction to obtain the smoothed tensor . For smooth tensors Perform eigenvalue decomposition to obtain three sets of eigenvalues Among them, if , it means that the neighborhood of the voxel is a slender tubular structure; the corresponding eigenvector represents the local main direction vector. The local direction enhancement value is calculated, and the local direction enhancement function is as follows: ,in Enhanced data for the hepatobiliary duct, is the maximum tensor eigenvalue, Second only to The tensor eigenvalues ​​of , Second to and The tensor eigenvalues ​​of , Provides spatial guidance domain data for the hepatobiliary duct.

[0126] Performing path tracing based on the hepatobiliary enhancement data to obtain hepatobiliary duct tracking data, wherein the path tracing is to construct a cost function based on signal strength, direction consistency, and path proximity to perform minimum cost path tracing;

[0127] In one embodiment, the path , is a set of path points, representing a continuous path, consisting of several voxel points. is the first path point, is the second path point, For the path points, the overall path cost function is defined as, , is the path cost function, which represents the total cost of the “credibility” of the entire path. The smaller the value, the better. is the voxel point order term, is the number of voxel points, is the image intensity weight coefficient, , Enhanced data for the hepatobiliary duct, is the direction change weight coefficient, is the direction change penalty (cosine angle distance), The first Individual points, The first Individual points, is the distance penalty weight coefficient, is the spatial distance between the path and the guiding area (center of the liver portal), To guide the center of the region (such as the liver portal); extract the initial point set from the liver portal center region , serving as the starting point for path tracing. Path extension is performed in 3D image space based on the Fast Marching algorithm or the Dijkstra shortest path search algorithm. A maximum path length is set; the path direction change angle is set to less than 45°. If it is greater than 45°, a direction consistency filter is performed to prevent the path from crossing non-target tissue. The output is a set of central path points in the hepatobiliary duct, including a sequence of 3D path coordinates representing the trunk and its branches.

[0128] Connectivity optimization is performed on the hepatobiliary duct tracking data to obtain the hepatobiliary duct positioning data.

[0129] In one embodiment, the following attribute analysis is performed on all paths to calculate the length and number of branches of all tracking paths; sub-branches with a length less than 20px or not connected to the main trunk are deleted; a distance transformation map of the path point set is generated based on the closed completion of the distance map; a morphological closing operation is performed on path breakpoints with a distance less than 3px; Skeletonize 3D is used to extract the skeleton centerline; and BFS is used to merge all branches into a main structure tree to obtain the bile duct positioning data.

[0130] Preferably, performing improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data includes:

[0131] Direction-selective enhancement kernel processing was performed on the hepatic artery region data to obtain the rough positioning data of the first hepatic artery;

[0132] In one embodiment, the input is an enhanced CT image (arterial phase or equilibrium phase) that has been cropped to the liver region. A set of 3D convolution kernels with different spatial orientations is constructed. , whose direction set is defined as: ,in For Axially is the offset angle, , For direction The voxel response map below, For direction The corresponding convolution kernel, is the input 3D image volume, is a direction in the direction set, is a set of directions; each voxel takes the maximum response value in all directions as the enhancement value: , The maximum response map is the maximum response value of each voxel in all directions, representing the strongest direction enhancement intensity, which is used for hepatic artery enhancement positioning. It is a single direction response graph, the image is in the direction The convolution response value on reflects the intensity of the blood vessels or edge structures in that direction. is a direction in the direction set (for example, +30° offset along the Z axis), is the maximization function. At the same time, the direction number with the largest response is recorded: , is the optimal response direction index of the voxel, For the optimal direction index map function, find the direction number of each voxel with the largest response value in all directions. It is a single direction response graph, the image is in the direction The convolution response value on reflects the intensity of the blood vessels or edge structures in that direction; the obtained rough positioning data of the first hepatic artery include .

[0133] The edge sensitivity of the hepatic artery region data is adjusted to obtain the rough positioning data of the second hepatic artery;

[0134] In one embodiment, a Laplacian of Gaussian (LoG) filter or a Sobel operator combination is applied to the original image to extract edge responses: , It is the edge response map / gradient amplitude map, which reflects the intensity change degree of the image at each voxel point. The larger it is, the more obvious the edge is. For The image gradient component in the direction, For The image gradient component in the direction, For The image gradient component in the direction. Combined with local contrast analysis (local window standard deviation and local mean, the window size is or ) Generate a comparison chart: , It is a local contrast map, which indicates the texture contrast intensity of the local area. The larger the value, the more obvious the regional contrast. is the local standard deviation, and the local area centered on the current voxel (e.g. or ) reflects the grayscale standard deviation within the range of texture variation. is the local mean, the mean of the grayscale values ​​in the same area, used to normalize the contrast intensity, To prevent zero offset, a small constant is used to prevent the denominator from being 0. The edge sensitivity map obtained through the above is: , It is an edge sensitivity map that fuses the edge intensity and local contrast structure response map. It is often used to enhance small blood vessels and low contrast blood vessel areas. is the edge response weight coefficient, which controls the proportion of edge strength in fusion and is set to 0.6. is the edge response graph, is the contrast weight coefficient, which controls the weight of local contrast in fusion and is set to 0.4. For the local contrast map, set , .

[0135] Attention perception enhancement is performed based on the coarse positioning data of the first hepatic artery and the coarse positioning data of the second hepatic artery to obtain weighted positioning data of the first hepatic artery and weighted positioning data of the second hepatic artery, respectively;

[0136] In one embodiment, for the first hepatic artery rough positioning data and the second hepatic artery rough positioning data , respectively construct channel attention modules: , is the channel attention factor (vector), size is , used to adjust the response strength of each channel, is the Sigmoid activation function, the output range is (0,1), For the input feature map In the spatial dimension Do the averaging and output channel-level statistics with dimensions of , To compress the fully connected layer, compress the channel to C / r (such as r=8), is a nonlinear activation function, To expand the fully connected layer, expand back to the original number of channels ; Use this attention factor to regulate the response map itself: , , is the weighted response map of the first hepatic artery after adding channel attention, For The calculated channel attention factor, is the rough positioning data of the first hepatic artery, is the weighted response map of the second hepatic artery after adding channel attention, For The calculated channel attention factor, This is the rough positioning data of the second hepatic artery. Add spatial attention: , is a spatial attention map that outputs a weight factor (range [0,1]) for each voxel site. It is a channel fusion convolution layer that outputs a spatial weight map of 1 channel. To concatenate two pooled images in the channel dimension, generate Feature map, For the channel dimension Do average pooling, For the channel dimension Do the maximum pooling. The weighted response map that simultaneously adds channel and spatial attention is used as the weighted positioning data of the first hepatic artery. For The calculated channel attention factor, The spatial attention map generated for the first localization map, is the rough positioning data of the first hepatic artery, The weighted response map that simultaneously adds channel and spatial attention is used as the weighted positioning data of the second hepatic artery. For The calculated channel attention factor, The spatial attention map generated for the second localization map, This is the rough positioning data of the second hepatic artery.

[0137] Confidence voting is performed based on the first hepatic artery weighted positioning data and the second hepatic artery weighted positioning data to obtain hepatic artery fusion positioning data, and centerline extraction is performed based on the hepatic artery fusion positioning data to obtain hepatic artery positioning data.

[0138] In one embodiment, , is the fused positioning map (confidence fusion result), is the dynamic normalized weight of the first graph, is a weighted localization map of the first hepatic artery (such as an edge perception map or a direction consistency map), is the dynamic normalized weight of the second graph, A weighted localization map (such as a direction perception map or an attention output map) is created for the second hepatic artery; , , is the dynamic normalized weight of the first graph, is the dynamic normalized weight of the second graph, is a weighted localization map of the first hepatic artery (such as an edge perception map or a direction consistency map), A weighted localization map of the second hepatic artery (such as a direction perception map or an attention output map) is generated. To prevent small constants from dividing by zero, take values ​​such as ;

[0139] Fusion graph Set the confidence threshold (e.g. ), generate a binary mask map: , is a mask image, indicating the credible hepatic artery structure area, is the confidence threshold. Apply 3D skeletonization to extract the centerline skeleton, or use the medialaxis algorithm in SimpleITK or VTK. Further path optimization is performed, removing branches with lengths less than 10 px. Use a minimal path cost (such as Fast Marching) to complete the interruption points and obtain hepatic artery localization data.

[0140] Preferably, performing directional segmentation analysis on the portal vein region data to obtain portal vein positioning data includes:

[0141] Calculate the principal direction gradient map of the portal vein area data to obtain principal direction gradient map data;

[0142] In one embodiment, the portal vein region image Perform first-order differentiation in three dimensions (Sobel, Prewitt, or Scharr can be used): , is the gradient vector, the three-dimensional gradient vector of the current voxel, is the portal vein area data, is the voxel in the 3D image Location, is the voxel in the 3D image Location, is the voxel in the 3D image Position. A window around each voxel (e.g. ), calculate the structure tensor , is the local structure tensor, which represents the directional structure covariance of the current voxel in the neighborhood. is the index of the neighborhood index voxel, the index of each voxel coordinate in the local window involved in the calculation, is the neighborhood index voxel, and the coordinates of each voxel in the local window involved in the calculation, is the gradient vector, the three-dimensional gradient vector of the current voxel, is the transpose of the gradient vector, the transpose of the three-dimensional gradient vector of the current voxel, Find the eigendecomposition and select the eigenvector corresponding to the largest eigenvalue as the principal direction vector. This yields a principal direction gradient map, with one direction vector (unit vector) per voxel.

[0143] Perform direction-guided segmentation according to the main direction gradient map data to obtain direction-guided segmentation data;

[0144] In one embodiment, an initial seed point set is set near the portal vein trunk or the liver portal. ; The seed point can be set by anatomical rules, manually selected by experts, or automatically detected by the front module. , determine whether it meets the following three conditions within the 26-neighborhood: voxel point The main direction vector comes from the main direction gradient map, voxel point The main direction vector comes from the main direction gradient map, is the direction similarity threshold. The larger the value, the stricter the requirement for direction consistency. , Indicates the main direction vector of the point. The larger the threshold, the stricter the direction. ,in is the set grayscale threshold; , is the spatial distance function, is the Euclidean distance, The maximum expansion distance threshold is set to prevent jumping or wrong connection. The range is set to 1.5 to 2 px to avoid jumping in the diffusion path. If all the above conditions are met, Add to torrent collection , and continue to expand to its neighborhood until no new points can be added.

[0145] Branches are merged and breaks are repaired according to the direction-guided segmentation data to obtain portal vein positioning data.

[0146] In one embodiment, branch merging includes extracting all connected domains. For branches with shorter lengths (e.g., <20px), searching in space for: the Euclidean distance between the heads and tails of adjacent branches , is the Euclidean distance threshold (e.g. 5-10px), used to determine whether two branches are spatially adjacent; the main direction vector angle If satisfied, then use the shortest path to connect (using Fast Marching or shortest graph algorithm);

[0147] Fracture repair involves performing a 3D distance transformation D(x, y, z) on the output 3D mask; calculating the skeleton and performing a "bridging" operation on the skeleton breakpoints, finding connectable areas in the distance map, and connecting the breakpoints using the minimum cost path to obtain portal vein positioning data.

[0148] Preferably, extracting branch angles from the hepatic vein region data and the portal vein positioning data to obtain the hepatic vein positioning data includes:

[0149] Extract the centerline based on the portal vein positioning data to obtain portal vein direction map data;

[0150] In one embodiment, based on the portal vein 3D mask image, a 3D skeleton extraction algorithm (such as Medial AxisTransform or skeletonize_3d) is used to extract a skeleton point set, which is expressed as: , is the portal vein skeleton point set, is the first skeleton point, is the second skeleton point, For the Skeleton points, is the total number of skeleton points; for each skeleton point , take a symmetrical window in its neighborhood (such as the front and back points), define the neighborhood point set as: , for point The neighborhood set of is the starting point of the neighborhood boundary point, is the center store of the neighborhood point set, is the end point of the neighborhood boundary point; in this neighborhood, a principal direction line is fitted using the least squares fitting or principal component analysis (PCA) method. The principal direction vector of the fitting result is recorded as: , for point The main direction vector of The main direction unit vector obtained by fitting the skeleton points in the neighborhood reflects the direction of the skeleton segment where the point is located. The principal direction extraction function obtains the first principal component (maximum variance direction) based on least squares linear fitting or principal component analysis (PCA); after performing the above operation on all skeleton points, each principal direction vector is assigned to the corresponding voxel position to construct a three-dimensional sparse vector map ,in The portal vein direction map data is obtained, including the portal vein skeleton line point set and the direction vector of each skeleton point.

[0151] Perform skeleton extraction based on the hepatic vein region data to obtain the hepatic vein branch data;

[0152] In one embodiment, 3D skeleton extraction is performed on the hepatic vein region data (binary mask); all connected regions are retained; all branch segments are extracted (split by breakpoints); and the starting and ending coordinates of each branch are recorded as a point set sequence.

[0153] Perform linear fitting on the hepatic vein branch data to obtain hepatic vein branch fitting data, and calculate the angle of the nearest branch direction based on the hepatic vein branch fitting data to obtain branch angle characteristic map data;

[0154] In one embodiment, for each hepatic vein branch point set, linear least squares fitting or PCA is used to output the main direction vector; the direction vector closest to the branch is searched in the portal vein direction map, such as by using a nearest point search method, performing a KD tree or spatial index, and limiting the maximum radius, such as 10-15 px; and the cosine of the direction angle is calculated: , is the angle between the two vectors (degrees or radians), is the hepatic vein fitting direction vector, is the portal vein reference direction vector, is the modulus of the hepatic vein fitting direction vector, is the modulus of the portal vein reference direction vector; all angles are marked on the hepatic vein branch fitting data to obtain the branch angle feature map data.

[0155] Narrow neural network classification is performed based on the portal vein direction map data and branch angle feature map data to obtain the hepatic vein positioning data.

[0156] In one embodiment, narrow neural network classification refers to the use of a shallow, low-parameter, and low-dimensional neural network structure to complete the classification task. The width is no more than 16 and the depth is no more than 16 layers. Based on the feature classification of the narrow neural network, hepatic vein positioning data is generated. The classification model used is a lightweight narrow channel neural network (NarrowNet), which is used for voxel-level or branch-level structure recognition in a low-dimensional feature space. The network input features include the three-dimensional spatial position coordinates of each candidate point. ; Cosine value of the angle between the portal vein and the direction , or direction vector pair The neural network structure includes an input layer that accepts feature vectors of dimension 4; the first hidden layer contains 16 neurons with ReLU activation function; the second hidden layer (Dense2) contains 8 neurons with ReLU activation function; the output layer (Dense3) contains 1 neuron and uses Sigmoid activation function to output the probability value of the hepatic vein structure. The network training uses the binary cross entropy loss function (BinaryCrossEntropy), with manually labeled real hepatic vein voxels as positive samples and non-structural voxels or pseudo-vessels as negative samples. The network output is the classification confidence of each point , and the threshold The real hepatic vein area is screened out as the classification standard. The output result is the hepatic vein positioning mask map, that is, the hepatic vein positioning data.

[0157] Preferably, performing elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain the elastic registration data includes:

[0158] S31, determining the hepatic artery positioning data as reference structure data;

[0159] In one embodiment, the hepatic artery structure is stable, with minimal anatomical differences and a concentrated orientation; the contrast CT mid-arterial phase image is clear and has minimal error; the coordinate system of the hepatic artery centerline or 3D mask is set as the registration reference coordinate system; all target registration structures (hepatobiliary duct, portal vein, hepatic vein) need to be converted to this coordinate system.

[0160] S32, performing inter-structure pairing on the hepatobiliary duct positioning data, the portal vein positioning data, and the hepatic vein positioning data according to the reference structure data to obtain inter-structure pairing data;

[0161] In one embodiment, matching point pairs are extracted from each structure to provide control constraints for elastic registration, and the centerline point set of the reference structure (hepatic artery) is extracted. ; For each target structure (such as portal vein): extract its centerline point set , calculate the nearest neighbor pairs between point sets: , With reference point The nearest target point , To find the variable that minimizes a function (for nearest neighbors), is the centerline point set of any target structure, For the Hepatic artery centerline points, is the centerline point of any target structure; , is a point pair set, representing the set of all valid paired points, is the first point in the centerline point set of the hepatic artery reference structure (reference structure). Points, is the centerline point set of the portal vein (one of the target structures). Points, is the first point in the centerline point set of the hepatic artery reference structure (reference structure). Points, is the centerline point set of the hepatobiliary duct (one of the target structures). points; filter the point pairs to obtain the structure pairing data, such as setting a maximum distance threshold (such as <20px) and adding direction consistency (angle <30°).

[0162] In one embodiment, a tree-like mapping of the structures is performed; matching is performed from the trunk to the primary branches and then to the secondary branches (recursive structural constraints) to obtain inter-structure matching data. In this structural matching, trunk matching is prioritized, followed by recursive nearest neighbor and consistency matching of the primary and secondary branches to achieve anatomical consistency and topological rationality.

[0163] S33. Perform control point network coverage based on the inter-structure pairing data to obtain elastic registration data.

[0164] In one embodiment, the image / structure is covered with a grid of control points with a spacing of (or px number), the number of control points is automatically adjusted according to the structure density; the paired point set is used as the constraint driver to guide the elastic deformation of the control point grid, and the deformation field is defined as , represents a voxel The mapping position after deformation. Construct an optimization objective function to achieve deformation field formation: , is the total registration error loss, is a set of structure paired points, representing the key points in the source structure Corresponding points in the target structure A set of paired points formed, used as registration anchor points, is the position after deformation, point In the current deformation field The mapping position under The spline deformed mesh is interpolated to obtain, is the regularization coefficient, which controls the weight of the deformation smoothing term in the total loss function and is set to , the larger the value, the more likely it is to keep the deformation smooth. Regularization / deformation smoothness constrains the smoothness and continuity of the entire deformation field using bending energy or first / second-order derivative penalties. The optimization objective function consists of two parts: a paired point reprojection error term (for data fidelity) and a smooth regularization term (for controlling local continuity of registration). The optimization process is implemented using numerical optimization algorithms such as gradient descent and LBFGS. The output 3D non-rigid deformation field is used for visualization.

[0165] Preferably, a 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation system is characterized in that it is used to perform the 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation method described above, and the 3D hepatobiliary, hepatic artery, portal vein, and hepatic vein image segmentation system comprises:

[0166] A multi-phase image acquisition module, used for acquiring multi-phase medical image data of the liver area;

[0167] A region recognition module is used to perform region recognition on multi-phase medical image data to obtain bile duct region data, hepatic artery region data, portal vein region data, and hepatic vein region data;

[0168] The structural association positioning module is used to perform structural tracking on the hepatic artery region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; and perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data.

[0169] The three-dimensional structure reconstruction and display module is used to perform elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; and perform three-dimensional visualization operations based on the elastic registration data.

[0170] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes that fall within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0171] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein image segmentation method, characterized in that: The following steps are involved: Acquiring multi-phase medical image data of a liver region; Performing regional recognition on multi-phase medical image data to obtain hepatobiliary duct regional data, hepatic artery regional data, portal vein regional data, and hepatic vein regional data; Perform structural tracking on the hepatobiliary region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data; Perform elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; Perform 3D visualization based on elastic registration data; The structure tracking includes: Constructing a spatial guidance domain for the hepatobiliary region data to obtain the hepatobiliary spatial guidance domain data; Perform local direction consistency enhancement based on the hepatobiliary spatial guidance domain data to obtain hepatobiliary enhancement data; Performing path tracing based on the hepatobiliary enhancement data to obtain hepatobiliary duct tracking data, wherein the path tracing is to construct a cost function based on signal strength, direction consistency, and path proximity to perform minimum cost path tracing; Optimize the connectivity of the hepatobiliary duct tracking data to obtain the hepatobiliary duct positioning data; The improved contrast extraction includes: Direction-selective enhancement kernel processing was performed on the hepatic artery region data to obtain the rough positioning data of the first hepatic artery; The edge sensitivity of the hepatic artery region data is adjusted to obtain the rough positioning data of the second hepatic artery; Attention perception enhancement is performed based on the coarse positioning data of the first hepatic artery and the coarse positioning data of the second hepatic artery to obtain weighted positioning data of the first hepatic artery and weighted positioning data of the second hepatic artery, respectively; Confidence voting is performed based on the first hepatic artery weighted positioning data and the second hepatic artery weighted positioning data to obtain hepatic artery fusion positioning data, and centerline extraction is performed based on the hepatic artery fusion positioning data to obtain hepatic artery positioning data; The directional segmentation analysis includes: Calculate the principal direction gradient map of the portal vein area data to obtain principal direction gradient map data; Perform direction-guided segmentation according to the main direction gradient map data to obtain direction-guided segmentation data; Based on the direction-guided segmentation data, branches are merged and breaks are repaired to obtain portal vein positioning data; The multi-phase medical image data includes arterial phase CT image data, portal venous phase CT image data, equilibrium phase CT image data, and T2-weighted MRI image data. The multi-phase medical image data is subjected to region recognition to obtain hepatobiliary duct region data, hepatic artery region data, portal vein region data, and hepatic vein region data, respectively, including: Performing coarse liver region segmentation on portal venous phase CT image data to obtain liver region data; Determine the hepatic artery region according to the liver region data and the arterial phase CT image data to obtain the hepatic artery region data; Hepatic vein region data is identified based on liver region data and equilibrium phase CT image data to obtain hepatic vein region data; Directionally guided segmentation is performed based on the liver region data to obtain portal vein region data; Identify the hepatobiliary region based on the liver region data and T2-weighted MRI image data to obtain the hepatobiliary region data; The step of performing directionally guided segmentation based on the liver region data to obtain portal vein region data includes: Perform voxel grayscale gradient extraction based on liver region data to obtain voxel grayscale gradient data; Perform main direction clustering on the voxel gray gradient data to obtain main direction clustering data; Perform tensor voting processing based on the main direction clustering data to obtain the vascular direction field data; The vascular direction field data is segmented using a preset direction-guided segmentation network model to obtain portal vein region data. The direction-guided segmentation network model is constructed by using an improved SD-UNet as the backbone network and an attention map or guided flow as the output head. The steps for constructing the preset direction-guided segmentation network model include the following steps: Obtain historical vascular direction field data and historical annotation data; Perform multi-scale exaggerated convolution parallel processing based on historical vascular direction field data to obtain edge perception layer data; The edge perception layer data is processed at the shallow edge layer and at the high-level feature layer to obtain shallow edge layer data and high-level feature layer data respectively. The shallow edge layer processing is performed through Sobel / LoG filter or edge channel convolution to enhance the boundary response. Perform residual connection on shallow edge layer data and high-level feature layer data to obtain decoding layer data; The vascular direction network is trained on the decoded layer data using historical annotated data to obtain direction-guided flow data. The vascular direction network training supervises the convolutional network by using the direction consistency loss combined with the Dice loss. The direction-guided output of the direction-guided flow data is performed to obtain a direction-guided segmentation network model, wherein the direction-guided output is convolutionally fused with the decoding layer data and the direction confidence is calculated, or the direction-guided attention map is calculated through the direction-guided flow data, and the output result is regulated at the channel level.

2. The method according to claim 1, characterized in that The step of extracting branch angles from the hepatic vein region data and the portal vein positioning data to obtain the hepatic vein positioning data includes: Extract the centerline based on the portal vein positioning data to obtain portal vein direction map data; Perform skeleton extraction based on the hepatic vein region data to obtain the hepatic vein branch data; Perform linear fitting on the hepatic vein branch data to obtain hepatic vein branch fitting data, and calculate the angle of the nearest branch direction based on the hepatic vein branch fitting data to obtain branch angle characteristic map data; Narrow neural network classification is performed based on the portal vein direction map data and branch angle feature map data to obtain the hepatic vein positioning data.

3. The method according to claim 1, characterized in that The elastic registration is performed based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain the elastic registration data, including: Determine the hepatic artery positioning data as reference structure data; Performing inter-structure pairing on the hepatobiliary duct positioning data, the portal vein positioning data, and the hepatic vein positioning data according to the reference structure data to obtain inter-structure pairing data; The control point network is covered based on the paired data between structures to obtain the elastic registration data.

4. A three-dimensional hepatobiliary duct, hepatic artery, portal vein and hepatic vein image segmentation system, characterized in that: For executing the 3D hepatobiliary, hepatic artery, portal vein and hepatic vein image segmentation method according to claim 1, the 3D hepatobiliary, hepatic artery, portal vein and hepatic vein image segmentation system comprises: A multi-phase image acquisition module, used for acquiring multi-phase medical image data of the liver area; A region recognition module is used to perform region recognition on multi-phase medical image data to obtain bile duct region data, hepatic artery region data, portal vein region data, and hepatic vein region data; The structural association positioning module is used to perform structural tracking on the hepatic artery region data to obtain hepatobiliary duct positioning data; perform improved contrast extraction on the hepatic artery region data to obtain hepatic artery positioning data; perform directional segmentation analysis on the portal vein region data to obtain portal vein positioning data; and perform branch angle extraction on the hepatic vein region data and portal vein positioning data to obtain hepatic vein positioning data. The three-dimensional structure reconstruction and display module is used to perform elastic registration based on the hepatobiliary duct positioning data, the hepatic artery positioning data, the portal vein positioning data, and the hepatic vein positioning data to obtain elastic registration data; and perform three-dimensional visualization operations based on the elastic registration data.

Images