An Automated Liver Segmentation Modeling Method and System Based on Hybrid Neural Networks

Abstract

This application provides an automated liver segmentation modeling method and system based on a hybrid neural network, relating to the field of medical imaging technology. The method includes: preprocessing acquired raw liver images to obtain a three-dimensional liver image; wherein the three-dimensional liver image includes liver parenchyma, portal vein, and hepatic vein; inputting the three-dimensional liver image into a trained segmentation model to obtain liver parenchyma masks, portal vein masks, and hepatic vein masks; dividing the liver parenchyma into multiple watershed partitions based on the liver parenchyma masks, portal vein masks, and hepatic vein masks; determining the standard liver segment corresponding to each watershed partition based on preset liver segmentation rules; and elastically deforming the watershed partitions based on the standard liver segments corresponding to the watershed partitions to obtain deformed watershed partitions. This application can reflect the individualized vascular variation characteristics of patients.

Description

Technical Field

[0001] This application relates to the field of medical imaging technology, and in particular to an automated liver segmentation modeling method and system based on hybrid neural networks. Background Technology

[0002] Liver segmentation is crucial in liver surgery, liver transplantation, and interventional procedures. Accurate segmentation not only helps define the extent of liver resection but also provides a reliable basis for assessing residual liver function and preoperative planning. The currently accepted Couinaud segmentation system, based on the boundary between portal vein supply and hepatic veins, divides the liver into eight anatomical functional units and has become an internationally recognized standard. However, this method, based on an idealized anatomical model, fails to fully reflect the individualized vascular variations of patients, resulting in insufficient clinical applicability. Summary of the Invention

[0003] This application is made in view of at least one of the above-mentioned technical problems existing in the prior art, and is able to reflect the individualized vascular variation characteristics of patients.

[0004] In a first aspect, embodiments of this application provide an automated liver segmentation modeling method based on a hybrid neural network, including:

[0005] The acquired raw liver images are preprocessed to obtain three-dimensional liver images; wherein, the three-dimensional liver images include liver parenchyma, portal vein and hepatic vein;

[0006] The three-dimensional liver image is input into the trained segmentation model to obtain liver parenchyma mask, portal vein mask, and hepatic vein mask;

[0007] Based on the liver parenchyma mask, the portal vein mask, and the hepatic vein mask, the liver parenchyma is divided into multiple watershed partitions;

[0008] Based on the preset liver segmentation rules, the standard liver segments corresponding to each watershed partition are determined;

[0009] Based on the standard liver segment corresponding to the watershed partition, the watershed partition is elastically deformed to obtain the deformed watershed partition.

[0010] Secondly, embodiments of this application provide an automated liver segmentation modeling system based on a hybrid neural network, comprising:

[0011] The preprocessing module is configured to preprocess the acquired raw liver images to obtain a three-dimensional liver image; wherein the three-dimensional liver image includes liver parenchyma, portal vein and hepatic vein;

[0012] The segmentation module is configured to input the three-dimensional liver image into a trained segmentation model to obtain liver parenchyma masks, portal vein masks, and hepatic vein masks;

[0013] The partitioning module is configured to divide the liver parenchyma into multiple watershed partitions based on the liver parenchyma mask, the portal vein mask, and the hepatic vein mask;

[0014] The segmentation module is configured to determine the standard liver segment corresponding to each watershed partition based on a preset liver segmentation rule; and to perform elastic deformation on the watershed partition based on the standard liver segment corresponding to the watershed partition to obtain the deformed watershed partition.

[0015] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory stores an executable program, and the processor executes the executable program to perform the steps of the method as described in any of the preceding claims.

[0016] Fourthly, embodiments of this application provide a computer program product, which, when executed by a processor, implements any of the methods described above.

[0017] This application provides an automated liver segmentation modeling method and system based on hybrid neural networks. Through preprocessing steps, it effectively eliminates non-standardized interference from the original images, ensuring clear structure and accurate location of the liver parenchyma, portal vein, and hepatic veins in 3D liver images. The segmentation model automatically segments key structures, avoiding the enormous workload of manually or semi-manually delineating liver parenchyma and blood vessels in traditional methods. Based on a hierarchical division logic of watershed partitioning and pre-defined liver segment rules, it first achieves natural anatomical partitioning based on the characteristics of the vascular watershed, and then anchors to standard liver segment rules, which not only conforms to the physiological characteristics of liver hemodynamics but also meets clinically accepted liver segmentation standards. Elastic deformation processing can adapt to individual liver morphological variations, ensuring a high degree of match between the partitioning results and the anatomical boundaries of standard liver segments, significantly reducing segmentation deviation. The resulting deformed watershed partitioning simultaneously considers individual anatomical realism and the standardization of liver segments, meeting the need for precise location of lesions in surgical planning and adapting to the standardized lesion localization description in imaging diagnosis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart of an automated liver segmentation modeling method based on a hybrid neural network, provided in one embodiment of this application;

[0020] Figure 2 This is a schematic diagram of a watershed partitioning process provided in one embodiment of this application.

[0021] Figure 3 This is a schematic diagram of an automated liver segmentation modeling system based on a hybrid neural network, provided in one embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the embodiments of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] like Figure 1 As shown in the figure, this application provides an automated liver segmentation modeling method based on a hybrid neural network, including:

[0024] Step 101: Preprocess the acquired raw liver image to obtain a three-dimensional liver image; wherein the three-dimensional liver image includes the liver parenchyma, portal vein and hepatic vein.

[0025] The raw liver images acquired clinically, such as DICOM sequence images from CT and MRI, are standardized and processed to output a three-dimensional liver image containing the complete liver parenchyma, portal vein, and hepatic vein structures.

[0026] The preprocessing process may specifically include: converting the 2D DICOM sequence into standard medical image formats that support 3D analysis, such as NIFTI / MHD; simultaneously unifying voxel spacing through spatial resampling to eliminate spatial scale differences caused by different devices and scanning parameters, ensuring accurate restoration of the 3D positional relationship of the liver and vascular structures; performing grayscale normalization based on image modal characteristics, and removing scanning noise by combining algorithms such as nonlocal mean denoising and wavelet denoising; eliminating interference through methods such as metal artifact compensation and respiratory motion artifact correction to avoid noise and artifacts affecting the subsequent differentiation of blood vessels and parenchyma; and extracting candidate liver regions through coarse segmentation methods such as threshold segmentation and region growing, excluding non-target organs such as the spleen, kidneys, and gastrointestinal tract in the abdominal cavity, reducing the interference of irrelevant data on subsequent model calculations, and improving processing efficiency.

[0027] Step 102: Input the 3D liver image into the trained segmentation model to obtain the liver parenchyma mask, portal vein mask, and hepatic vein mask.

[0028] The preprocessed 3D liver image is input into a trained deep learning segmentation model. Through the model's feature extraction and classification capabilities, three types of binary masks are output: liver parenchyma mask (marking the liver parenchyma region, with non-parenchyma regions as the background), portal vein mask (marking the main portal vein trunk and its branches), and hepatic vein mask (marking the main hepatic vein trunk and its branches).

[0029] Three-dimensional segmentation networks such as 3D-CNN are typically used. Compared with two-dimensional networks, they can completely capture the three-dimensional topological structure of the liver and blood vessels, avoiding problems such as blood vessel breakage and structural discontinuity caused by two-dimensional slice segmentation.

[0030] The training data needs to cover liver images with different types of vascular variations and different disease states. Multi-scale feature fusion and attention mechanism are used to improve the model's ability to identify small blood vessels. A composite loss function of Dice loss and boundary loss is used to solve the problem of low segmentation accuracy caused by the small proportion of blood vessel area and blurred boundaries.

[0031] All three types of masks are binary matrices with the same size as the input 3D image (1 represents the target region and 0 represents the background). The spatial coordinates of the portal vein and hepatic vein masks are matched with those of the liver parenchyma mask to ensure accurate association in subsequent watershed partitioning calculations based on masks.

[0032] Step 103: Based on the liver parenchyma mask, portal vein mask, and hepatic vein mask, divide the liver parenchyma into multiple watershed partitions.

[0033] The liver parenchyma is divided into multiple individualized watershed partitions corresponding to the portal vein blood supply range by using a liver parenchyma mask to limit the partition range, a portal vein mask to determine the perfusion source point, and a hepatic vein mask as the boundary constraint.

[0034] Step 104: Based on the preset liver segmentation rules, determine the standard liver segments corresponding to each watershed partition.

[0035] Using clinically accepted liver segmentation rules (such as the internationally recognized Couinaud eight-segment rule) as the standard, a corresponding standard liver segment label (such as segment I, segment II... segment VIII) is assigned to each individualized watershed partition, and a mapping relationship between individualized partitions and standardized liver segments is established.

[0036] Step 105: Based on the standard liver segment corresponding to the watershed partition, perform elastic deformation on the watershed partition to obtain the deformed watershed partition.

[0037] Based on the standard liver segments corresponding to the watershed partitions, elastic deformation optimization is performed on the individualized watershed partitions. By adjusting the voxel positions of the partition boundaries, the deformed watershed partitions retain the individualized vascular perfusion characteristics of the patients and better conform to the anatomical contours of the standard liver segments in terms of overall morphology and spatial range, ultimately obtaining deformed watershed partitions that balance standardization and individualization.

[0038] This application's embodiments effectively eliminate non-standardized interference from the original images through preprocessing steps, ensuring clear structures and accurate locations of the liver parenchyma, portal vein, and hepatic veins in the 3D liver images. The segmentation model automatically segments key structures, avoiding the enormous workload of manually or semi-manually delineating the liver parenchyma and blood vessels in traditional methods. Based on a hierarchical division logic of watershed partitioning and preset liver segment rules, natural anatomical partitioning is first achieved based on the characteristics of the vascular watershed, and then standard liver segment rules are anchored, which not only conforms to the physiological characteristics of liver hemodynamics but also meets clinically accepted liver segmentation standards. Elastic deformation processing can adapt to individual liver morphological variations, ensuring a high degree of match between the partitioning results and the anatomical boundaries of standard liver segments, significantly reducing segmentation deviation. The resulting deformed watershed partitioning simultaneously considers individual anatomical realism and the standardization of standard liver segments, meeting the need for precise location of lesions in surgical planning and adapting to lesion localization descriptions under unified standards in imaging diagnosis.

[0039] In one embodiment of this application, the segmentation model includes: an input module, an encoder, a feature fusion module, a decoder, and an output module;

[0040] The input module is used to segment the 3D liver image into multiple image blocks;

[0041] An encoder is used to extract features of different sizes from multiple image patches;

[0042] The feature fusion module is used to fuse features of different sizes based on the channel attention mechanism;

[0043] The decoder is used to generate a feature map of the same size as the image patch based on the fused features output by the feature fusion module.

[0044] The output module is used to output liver parenchyma masks, portal vein masks, and hepatic vein masks based on the feature maps output by the decoder.

[0045] The input module segments the complete 3D liver image into multiple fixed-size small image blocks, which serve as the input to the encoder. These are typically segmented into 64×64×32 or 32×32×16 voxels. If the size is too small, the 3D topological features of the blood vessels are easily lost; if the size is too large, it increases memory usage and computation time.

[0046] The encoder employs a cascaded structure of multiple convolutional layers and downsampling layers to perform layer-by-layer feature extraction and size compression on each image patch, outputting multi-scale feature maps ranging from low-dimensional detail features to high-dimensional semantic features, capturing different levels of features from the liver parenchyma, portal vein, and hepatic veins. For example, it can consist of four layers of 3D convolutional units, each containing: a 3D convolutional layer (Conv3D, kernel=3×3×3): extracting local spatial features; a batch normalization layer: accelerating convergence and stabilizing training; a ReLU activation layer: introducing non-linear features; and a downsampling layer (MaxPooling3D, stride=2): progressively compressing spatial dimensions to extract high-level semantic information.

[0047] The feature fusion module receives multi-scale feature maps output by the encoder, dynamically adjusts the feature weights of different scales and channels through a channel attention mechanism, and then achieves multi-scale feature fusion through upsampling and stitching to highlight the key feature channels for blood vessel segmentation.

[0048] A compression-excitation structure is adopted, which performs the following operations on the feature map for each scale: Compression: The 3D feature map (C×H×W×D, where C is the number of channels) is compressed into a 1×1×1×C channel descriptor through global average pooling, aggregating the global information of each channel; Excitation: The channel descriptor is mapped to C weight values ​​through two fully connected layers. The higher the weight value, the more important the channel is for blood vessel segmentation; Weight assignment: The calculated channel weights are multiplied with the original feature map channel by channel to enhance key feature channels and suppress redundant channels.

[0049] A top-down fusion approach is adopted, in which deep high-dimensional features are upsampled to the same size as shallow features through 3D transposed convolution, and then concatenated with shallow features that have been weighted by channel attention to form a feature map that integrates global semantics and local details.

[0050] The decoder employs a multi-transposed convolutional layer and a cascaded convolutional layer structure symmetrical to the encoder. It performs layer-by-layer upsampling and feature refinement on the fused features, gradually restoring the feature map size. Finally, it outputs a high-resolution feature map that is completely consistent with the size of the input image block, preserving the spatial location and boundary details of the liver parenchyma and blood vessels.

[0051] The decoder and encoder layers are symmetrical in structure, and the resolution is restored by upsampling layer by layer; each decoder layer contains: a deconvolution layer or an upsampling interpolation layer; a feature concatenation layer; and 3D convolution, batch normalization and ReLU activation layers.

[0052] The output module transforms the high-resolution feature map output by the decoder into three types of binary masks through convolutional mapping and activation functions, and then stitches together the segmented image patch masks to obtain the complete mask result of the liver region of interest.

[0053] The encoder's layered design, combining shallow and deep layers, allows for the simultaneous extraction of low-dimensional detail features and high-dimensional semantic features. The channel attention mechanism in the feature fusion module dynamically strengthens key feature channels and suppresses redundant features, avoiding segmentation bias caused by imbalances in multi-scale feature weights. For example, in images of patients with cirrhosis, the liver parenchyma exhibits uneven grayscale. Shallow detail features help locate blurred blood vessel edges, while deep semantic features help distinguish between blood vessels and parenchymal lesions. The fusion of these two features significantly improves the completeness of blood vessel segmentation.

[0054] In one embodiment of this application, the liver parenchyma is divided into multiple watershed partitions based on liver parenchyma masks, portal vein masks, and hepatic vein masks, including:

[0055] Based on the portal vein mask, multiple source points are identified; these source points are used to characterize the origin of blood perfusion.

[0056] Based on multiple source points, liver parenchyma masks, and hepatic vein masks, the liver parenchyma is divided into multiple watershed partitions.

[0057] The algorithm automatically locates the endpoints or segment points of the portal venous tree using distance transformation and skeletonization algorithms. These points are the final starting points for blood perfusion into the liver parenchyma via the portal venous system. Each identified point is defined as an independent source point, representing a microscopic, independent blood perfusion source that supplies blood to the surrounding liver parenchyma. When regions diffusing from different source points meet, a partition boundary is formed. This means that the liver parenchyma at that location is being supplied by two or more perfusion source points competing for blood. The algorithm assigns the region to the unique source point according to preset rules.

[0058] The liver parenchyma is divided into multiple non-overlapping, continuous three-dimensional watersheds. Each watershed theoretically corresponds to the blood supply area of ​​a terminal portal vein branch and the subsegment of the hepatic vein that drains from it.

[0059] This application's embodiments utilize the patient's own imaging data to automatically generate watershed partitions, reflecting the patient's unique vascular tree structure and blood perfusion domain, providing a surgical navigation map for precision liver surgery that transcends traditional segmentation.

[0060] In one embodiment of this application, multiple source points are determined based on a portal vein mask, including:

[0061] Identify connected regions in the portal vein mask;

[0062] The connected region containing the largest number of voxels is identified as the portal vein trunk tree;

[0063] Generate a portal vein tree based on the portal vein trunk tree;

[0064] Based on the portal vein tree, multiple source points are identified.

[0065] Using the portal vein mask output by the segmentation model as input, the three-dimensional connected component analysis algorithm traverses all voxels with a value of 1 in the mask, divides spatially connected voxels into the same connected region, and outputs a set of all independent connected regions in the mask (such as region A, region B, region C, etc.).

[0066] The 26-connectivity rule, commonly used in clinical image analysis, is applied. If two voxels with a value of 1 are adjacent in three-dimensional space in six orthogonal directions (up / down, left / right, front / back), or in twelve diagonal directions, or in eight spatial diagonal directions, they are considered connected and classified as belonging to the same region. This rule can completely capture the spatial continuity of portal vein branches and avoids vascular tree breakage caused by overly strict connectivity determination.

[0067] The connected regions with the largest number of voxels are selected and identified as the main portal vein tree. Simultaneously, all small connected regions with a voxel count less than the noise threshold are removed. The noise threshold needs to balance noise removal and preservation of small branches. If set too low, segmentation noise may not be completely removed; if set too high, small terminal branches of the portal vein may be mistakenly deleted. In practice, a noise threshold of 1% of the main portal vein tree volume is typically used. Small regions with a volume less than this noise threshold are removed, ensuring that irrelevant noise is removed while preserving clinically significant small branches.

[0068] A 3D vascular skeletonization and topological analysis were performed on the portal vein trunk tree, transforming solid portal vein regions into a tree-like topological structure reflecting branching and bifurcation relationships, i.e., the portal vein tree. An iterative refinement algorithm was employed to remove edge voxels from the portal vein trunk tree through multiple iterations, retaining only a single voxel column along the vessel centerline. Based on the skeletonization results, the portal vein tree topology was constructed through node and edge modeling.

[0069] Nodes: Branch endpoints on the skeleton (26-1 skeleton voxel in the neighborhood), bifurcation points (26-3 or more skeleton voxels in the neighborhood);

[0070] Edges: A column of skeleton voxels connecting adjacent nodes, representing the branch segments of the portal vein (such as the main trunk, left branch, and right posterior branch). The final output is a portal vein tree topology graph containing node positions, edge connections, and branch lengths, clearly reflecting the branch hierarchy of the portal vein.

[0071] Based on the portal vein tree, key nodes that can characterize the initiation of blood perfusion are selected as the source points for watershed segmentation.

[0072] Traditional methods directly locate the source point based on the portal vein mask, which is susceptible to segmentation noise, leading to mislocation of the source point. The embodiments of this application, through connected component identification and trunk tree screening, can completely eliminate segmentation noise, ensuring that subsequent vascular tree generation and source point localization are based solely on the actual portal vein trunk and branches. Simultaneously, through skeletonized topology analysis, the source point localization error can be controlled within 1-2 voxels, an improvement of over 60% compared to traditional methods, fundamentally guaranteeing the anatomical accuracy of the source point.

[0073] This application's embodiments generate source points based on the actual structure of the patient's own portal vein mask. For example, if the patient has an early-onset right anterior branch of the portal vein, the trunk tree screening will include this early-onset branch, and topological analysis will set its bifurcation point as the source point. The corresponding watershed partition will also be adjusted synchronously to accurately reflect the blood supply range of this variant branch. If the patient has multiple portal veins converging, through volume screening and anatomical location verification, the main converging branch can be accurately located as the trunk tree, avoiding interference from secondary branches.

[0074] In one embodiment of this application, multiple source points are identified based on the portal vein tree, including:

[0075] Based on the portal vein tree, multiple endpoints and multiple bifurcation points are identified;

[0076] Based on a preset radius ratio threshold, key bifurcation points are determined from multiple bifurcation points; where multiple endpoints and multiple bifurcation points constitute the source point.

[0077] Each node on the portal vein tree skeleton is traversed, and 26-neighborhood connectivity is used for determination. If there is only one portal vein tree skeleton voxel connected to a node in its 26-neighborhood, it is determined to be an endpoint. The endpoint corresponds to the terminal end of the portal vein branch and is the starting point of the smallest functional unit of blood perfusion. Its spatial location directly determines the extent of the terminal perfusion area.

[0078] For each node, a 26-neighborhood connectivity determination and branching analysis are performed. If a node has ≥3 portal vein tree skeleton voxels connected to it within its 26-neighborhood, and these connected voxels extend in different spatial directions, it is identified as a bifurcation point. The bifurcation point corresponds to the shunting location of the portal vein branch and is the regional shunting starting point for blood perfusion.

[0079] For all branches connected to each bifurcation point, the effective radius of each branch near the bifurcation point is calculated using the distance transformation method. Taking the bifurcation point as the center, a 5mm long branch segment is taken along the branch extension direction. The maximum distance from all skeleton voxels in this branch segment to the branch centerline (i.e., the half diameter of the branch) is calculated, and the average value is taken as the radius of the branch.

[0080] A preset threshold for the ratio of the radius of the main branch to the secondary branch is established. For each branching point, the following steps are performed: From the branches connected to the branching point, the branch with the largest radius is selected as the main branch; the radius ratio of all secondary branches to the main branch is calculated; if there are ≥2 secondary branches whose radius ratio is ≥ the preset threshold for the ratio of the radius of the main branch to the secondary branch, then the branching point is a critical branching point; if the radius ratio of all secondary branches is < the preset threshold for the ratio of the radius of the main branch to the secondary branch, then the branching point is a non-critical branching point and is removed.

[0081] Traditional methods use only a single type of source point, such as the main bifurcation point or the terminal endpoint, which cannot cover the needs of different surgical scenarios. This application's embodiment achieves layered and precise positioning by combining endpoints and key bifurcation points as source points. Key bifurcation points correspond to perfusion zones covering 10%-50% of the total liver volume, meeting the segmentation requirements of hemihepatectomy and segmentectomy.

[0082] In one embodiment of this application, such as Figure 2 As shown, based on multiple source points, liver parenchyma masks, and hepatic vein masks, the liver parenchyma is divided into multiple watershed regions, including:

[0083] Step 201: Construct a graph structure based on the liver parenchyma mask; where liver voxels in the liver parenchyma mask are nodes, and the connections between liver voxels and their adjacent liver voxels are edges.

[0084] Step 202: Determine the weights of the edges based on the liver parenchyma mask and the hepatic vein mask.

[0085] Step 203: Based on the graph structure and edge weights, calculate the weighted path distance from the liver voxel to each source point.

[0086] Step 204: Based on the weighted path distance from liver voxels to each source point, divide the liver parenchyma into multiple watershed partitions.

[0087] All voxels with a value of 1 in the liver parenchyma mask are directly defined as nodes of the graph structure. Each node corresponds to a unique three-dimensional coordinate. The nodes only contain liver parenchyma voxels, ensuring that the graph structure only covers the target area that needs to be partitioned and eliminating background interference.

[0088] The 26-neighborhood connectivity rule is used to define the edges between nodes. For any two nodes, if they are adjacent in three-dimensional space in six orthogonal directions (up / down, left / right, front / back), or in twelve diagonal directions, or in eight spatial diagonal directions (i.e., the absolute value of the coordinate difference is ≤1), then an undirected edge is established between the two nodes.

[0089] Based on the three-dimensional coordinates of the two nodes, the Euclidean distance is calculated as the basic weight of the edge.

[0090] Introducing a hepatic vein mask as an anatomical constraint factor, edges passing through the hepatic vein region are weighted and penalized, ultimately affecting the edge weight. As shown in equation (1),

[0091] (1)

[0092] Where ||u - v||² is the Euclidean distance, α is the penalty coefficient (α>0), and M hv (u) refers to the value of voxel u belonging to the hepatic vein mask (1 if it is, 0 otherwise), M hv (v) refers to the value of voxel v belonging to the hepatic vein mask (1 if it is, 0 otherwise), and Equation (1) makes the cost higher when the path passes through the hepatic vein region.

[0093] A multi-source Dijkstra's algorithm is employed, with all source nodes serving as starting nodes simultaneously. Initially, the path distances of all source nodes are set to 0, while the path distances of other nodes are set to infinity. The algorithm calculates the weighted path distance from each liver voxel to all source nodes. The weighted path distance from source node s to liver voxel v is the sum of the weights on all paths from source node s to voxel v. For each voxel v, the weighted path distances to all source nodes are iterated, and v is assigned to the source node that minimizes the weighted path distance.

[0094] This application's embodiments accurately simulate the path of blood diffusion from the source point to the liver parenchyma. Through weighted path distance calculation, the distance value of each voxel directly reflects the probability that the voxel receives blood supply from a certain source point, providing a quantitative basis for subsequent voxel allocation. Simultaneously, because the edge weights are incorporated into the hepatic vein constraint, the distance field naturally forms a partition feature with the hepatic vein as the boundary, laying the boundary foundation for subsequent watershed partitioning. Traditional watershed segmentation does not consider the physical barrier effect of the hepatic vein, easily leading to partitions crossing the hepatic vein.

[0095] This application's embodiment uses weight penalties to make the weights of edges crossing hepatic veins in the graph structure much higher than normal edges. Subsequent path calculations automatically avoid these high-weight edges, ensuring that the watershed partition boundaries are consistent with the anatomical boundaries of the hepatic veins, conforming to physiological perfusion patterns. Traditional watershed segmentation is directly based on grayscale or Euclidean distance calculations, which cannot accurately simulate the flow path of blood within the liver parenchyma. This application's embodiment, through graph structure modeling, transforms the perfusion zone division into a shortest path problem on the graph, laying the foundation for subsequent calculation of weighted path distances in conjunction with anatomical constraints.

[0096] In one embodiment of this application, the liver parenchyma is divided into multiple watershed partitions based on the weighted path distance from liver voxels to each source point, including:

[0097] Based on the weighted path distance from liver voxels to each source point, liver voxels are assigned to the source point with the smallest weighted path distance; among them, liver voxels assigned to the same source point constitute the initial region.

[0098] Multiple initial regions are merged based on a preset volume threshold to obtain multiple watershed partitions.

[0099] Based on the voxel spacing of the liver parenchyma mask, the number of voxels contained in each initial region is calculated. The region volume is then the product of the volume of a single voxel and the number of voxels. The volume of each initial region is calculated. If the volume is less than 0.5% of the total liver volume, the initial region is merged into the adjacent initial region that shares the longest boundary with it.

[0100] In practical applications, three-dimensional morphological closing operations can be performed on the boundary voxels of each initial region to eliminate jagged noise and peninsula-shaped protruding voxels at the boundaries, making the partition boundaries more in line with the natural anatomical shape of the liver parenchyma while maintaining the connectivity of the partitions.

[0101] The initial region may contain small areas due to excessively dense source points or distance calculation errors. These areas have no independent value for liver function assessment and may interfere with subsequent standard liver segment mapping. In this embodiment, meaningless small areas are merged into adjacent large regions by using volume threshold merging, ensuring that the final number of watershed partitions is controlled at 8-12, and that each partition has independent clinical functional significance.

[0102] In one embodiment of this application, the standard liver segment corresponding to each watershed partition is determined based on a preset liver segmentation rule, including:

[0103] Based on the preset liver segmentation rules, the parameters of multiple standard liver segments are determined;

[0104] For each watershed partition: based on the parameters of each standard liver segment, calculate the similarity between the watershed partition and each standard liver segment; based on the similarity, determine the target standard liver segment corresponding to the watershed partition.

[0105] Based on clinically accepted liver segmentation rules (such as the internationally recognized Couinaud eight-segment rule, Healey-Schroy four-segment rule, etc., this application uses the Couinaud eight-segment rule as an example for illustration), the abstract anatomical definition is transformed into calculable and comparable standard liver segment parameters.

[0106] The Couinaud eight-segment rule divides the liver into eight independent functional segments based on the portal vein supply area and using the hepatic veins (middle hepatic vein, left hepatic vein, right hepatic vein), portal vein branches (left sagittal branch, right main branch), and falciform ligament as boundaries. These segments are: (Segment I: caudate lobe; Segment II: upper segment of left lobe; Segment III: lower segment of left lobe; Segment IV: medial segment of left lobe, including subsegments IVa and IVb; Segment V: anteroinferior segment of right lobe; Segment VI: posteroinferior segment of right lobe; Segment VII: posterosuperior segment of right lobe; Segment VIII: anterosuperior segment of right lobe).

[0107] The parameters of a standard liver segment can include any one or more of the following: volume parameters, boundary features, and vascular parameters.

[0108] Volume parameters: The percentage range of each standard liver segment relative to the total liver volume (Segment I: 1–3%; Segment II: 10–15%; Segment III: 8–12%; Segment IV: 15–20%; Segment V: 12–18%; Segment VI: 10–15%; Segment VII: 11–16%; Segment VIII: 15–20%)

[0109] Boundary features: The main anatomical boundaries of each standard liver segment (e.g., for segment II, the upper boundary is the diaphragmatic surface, the lower boundary is the left branch of the portal vein, the medial boundary is the falciform ligament of the liver, and the lateral boundary is the left hepatic vein).

[0110] Vascular parameters: the angle range of portal vein branches (e.g., the angle of segment II branches is 30–45°) and the diameter range (1.5–3 mm for tertiary branches).

[0111] Calculate the similarity between the watershed partition and each standard liver segment, which can specifically include any one or more of the following: boundary similarity, topological similarity, and spatial location similarity.

[0112] Boundary similarity is calculated using the Jaccard similarity coefficient. Based on the aforementioned boundary features, the boundary between the watershed partition and the standard liver segment is represented as a three-dimensional polygonal mesh. After alignment, the ratio of the intersection area to the union area is calculated. This index quantifies the degree of matching between the watershed partition and the standard liver segment on the boundary.

[0113] Topological similarity is calculated using graph theory methods, specifically for watershed partitions, calculating hierarchical depth, branch angle, and vessel diameter. The topological similarity, TopologySim, is obtained by calculating the matching degree between these features and the corresponding features of the tree constructed from the standard liver (the aforementioned vessel parameters), as shown in Equation (2).

[0114] (2)

[0115] Wherein, α=0.4, β=0.3, γ=0.3; LevelSim is the branch level matching degree (by marking the levels of the vascular tree layer by layer, statistically analyzing the level differences of all matching vascular segments, and finally calculating the matching degree with the smaller the difference, the higher the score); AngleSim is the branch angle matching degree (by measuring the angle of the bifurcation point of each pair of matching vessels, calculating the relative deviation of the angle difference, and finally generating the matching degree with the smaller the deviation, the higher the score); DiameterSim is the vessel diameter matching degree (by estimating the diameter of each pair of matching vascular segments, calculating the relative proportion of the diameter difference, and finally generating the matching degree with the smaller the proportion, the higher the score).

[0116] Spatial location similarity is calculated by measuring the relative spatial position of the watershed partition and the standard liver segment, including the center point distance and volume ratio. The smaller the center point distance and the closer the volume ratio is to 1, the higher the spatial location similarity. The calculation method is shown in Equation (3).

[0117] (3)

[0118] Where: distance is the Euclidean distance between the center of the watershed partition and the center of the standard liver segment, volumeRatio is the ratio of the volume of the watershed partition to the volume of the standard liver segment, and SpatialSim is the spatial similarity.

[0119] The similarity between the watershed partition and the standard liver segment can be the sum of the three similarities mentioned above, or it can be as shown in equation (4).

[0120] Similarity = w1×BS + w2×TS + w3×SS (4)

[0121] Wherein, BS is the boundary similarity (the degree of matching between the liver segment boundary and the watershed partition boundary), TS is the topological similarity (the degree of matching between the vascular branch hierarchy and the spatial relationship), SS is the spatial location similarity (the degree of matching with the relative position of the standard liver segment in space), and w1, w2 and w3 are the corresponding weights, for example, w1=0.4, w2=0.3, w3=0.3.

[0122] Ultimately, the standard liver segment with the highest similarity was selected as the target standard liver segment.

[0123] Traditional methods require doctors to manually compare liver segments against established rules, identifying each segment on a 3D image. This process takes approximately 20-30 minutes per case, and the results are highly dependent on the doctor's experience. This application's implementation automates the entire process, significantly reducing the mapping time per case. Furthermore, based on a unified parameter library and calculation rules, the mapping results for different cases and operators achieve over 90% consistency, resolving the issues of high subjectivity and poor consistency associated with manual mapping.

[0124] Traditional mapping methods rely solely on a single dimension of spatial location or volume, resulting in matching error rates as high as 30%-40% in cases of vascular variations. This claim, through multi-dimensional similarity calculations of space, anatomy, and volume, combined with vascular association consistency determination, can accurately adapt to complex variations.

[0125] In one embodiment of this application, based on the standard liver segment corresponding to the watershed partition, the watershed partition is elastically deformed to obtain a deformed watershed partition, including:

[0126] Based on the target standard liver segment and the preset three-dimensional boundary model of the target standard liver segment, the watershed partition is elastically deformed to obtain the deformed watershed partition.

[0127] Object to be deformed: watershed partition.

[0128] Deformation target constraint: a pre-defined target standard liver segment three-dimensional boundary model (a standardized liver segment three-dimensional contour constructed based on clinical anatomical rules, such as the idealized three-dimensional boundary model of Couinaud II segment).

[0129] The three-dimensional boundary model is the result of statistical optimization based on a large amount of liver anatomical data from healthy individuals.

[0130] The specific deformation process includes:

[0131] The thin-plate spline elastic deformation algorithm or the B-spline deformation algorithm are preferred. Both algorithms support precise local adjustment and global shape fitting. The core logic is regional weighted deformation.

[0132] Landmark matching: On the individualized watershed partition and the standard liver segment model, key anatomical landmarks (such as the intersection of the partition and the hepatic vein, the geometric center of the partition, and the boundary point corresponding to the portal vein source point) are selected respectively. The correspondence between the source landmark (on the partition) and the target landmark (on the standard model) is established. Usually, 10-15 pairs of landmarks are selected to ensure that the main outline of the partition is covered.

[0133] Deformation field calculation: Based on the correspondence of marker points, the deformation vector of all voxels in the partition is calculated by TPS / B spline interpolation. Voxels closer to the marker point have deformation directions and amplitudes that are closer to the target marker point; voxels farther from the marker point have attenuated deformation amplitudes to ensure that the perfusion continuity within the partition is preserved.

[0134] Anatomical constraints are incorporated: Three types of mask constraints are introduced into the deformation field calculation: If the voxel is close to the boundary of the liver parenchyma mask, the deformation amplitude is limited to ≤5mm (to avoid the region from exceeding the liver area); if the voxel is close to the hepatic vein mask, the deformation direction must be away from the hepatic vein (to avoid the region crossing the hepatic vein after deformation); if the voxel belongs to the area around the portal vein mask (±3mm), the deformation amplitude is limited to ≤2mm (to avoid disrupting the blood supply connection between the portal vein and the region).

[0135] Voxel-by-voxel deformation execution: Based on the calculated deformation vector, adjust the three-dimensional coordinates of each voxel to generate deformed watershed partitions, and ensure the connectivity and boundary smoothness of the deformed watershed partitions.

[0136] This application's embodiments utilize elastic deformation to ensure that the segmentation morphologically conforms to standard liver segments while preserving individualized perfusion function. For example, if a patient's individualized hepatic tract segment (corresponding to segment II) is geometrically offset by approximately 6mm from the standard segment II due to deviation of the left branch of the portal vein, the segment boundary is adjusted to the outline of the standard segment II after deformation. Simultaneously, the relative positions of the internal voxels and the origin points of the left branch of the portal vein remain unchanged, conforming to the physician's anatomical understanding of segment II while reflecting the patient's actual blood supply range. This effect allows the segmentation results to be directly used for clinical communication and surgical navigation, addressing the core pain point of poor clinical adaptability in traditional methods.

[0137] In practical applications, deformed watersheds belonging to the same standard liver segment are assigned the same color label to facilitate identification by clinicians.

[0138] Three-dimensional reconstruction was performed based on the deformed watershed partitions to generate a three-dimensional model of the eight Couinaud segments of the liver. Key indicators such as volume and blood supply ratio of each segment were calculated based on the three-dimensional model for preoperative planning and functional assessment.

[0139] In practical applications, the 3D model can also be rendered in real time. Doctors can view the 3D structure of the liver on desktop or mobile devices, supporting rotation, zoom, and cross-sectional browsing to facilitate observation from different angles and levels. At the same time, blood vessels, tumors, segment boundaries, and surgical planning references can be overlaid to achieve structural-level visualization and comparison.

[0140] like Figure 3 As shown, this application provides an automated liver segmentation modeling system based on a hybrid neural network, comprising:

[0141] The preprocessing module 301 is configured to preprocess the acquired raw liver image to obtain a three-dimensional liver image; wherein the three-dimensional liver image includes the liver parenchyma, portal vein and hepatic vein;

[0142] The segmentation module 302 is configured to input a three-dimensional liver image into a trained segmentation model to obtain a liver parenchyma mask, a portal vein mask, and a hepatic vein mask.

[0143] The partitioning module 303 is configured to divide the liver parenchyma into multiple watershed partitions based on the liver parenchyma mask, portal vein mask, and hepatic vein mask;

[0144] The segmentation module 304 is configured to determine the standard liver segment corresponding to each watershed partition based on the preset liver segmentation rules; and to perform elastic deformation on the watershed partition based on the standard liver segment corresponding to the watershed partition to obtain the deformed watershed partition.

[0145] In one embodiment of this application, the partitioning module 303 is configured to determine multiple source points based on the portal vein mask; wherein the source points are used to characterize the source of blood perfusion; and the liver parenchyma is divided into multiple watershed partitions based on the multiple source points, the liver parenchyma mask, and the hepatic vein mask.

[0146] In one embodiment of this application, the partitioning module 303 is configured to identify connected regions in the portal vein mask; determine the connected region containing the largest number of voxels as the portal vein trunk tree; generate a portal vein tree based on the portal vein trunk tree; and determine multiple source points based on the portal vein tree.

[0147] In one embodiment of this application, the partitioning module 303 is configured to determine multiple endpoints and multiple bifurcation points based on the portal vein tree; and to determine key bifurcation points from the multiple bifurcation points based on a preset radius ratio threshold; wherein the multiple endpoints and multiple bifurcation points constitute the source point.

[0148] In one embodiment of this application, the partitioning module 303 is configured to construct a graph structure based on a liver parenchyma mask; wherein, liver voxels in the liver parenchyma mask are nodes, and the connections between a liver voxel and its adjacent liver voxels are edges; the weights of the edges are determined based on the liver parenchyma mask and the hepatic vein mask; the weighted path distance from the liver voxel to each source node is calculated based on the graph structure and the edge weights; and the liver parenchyma is divided into multiple watershed partitions based on the weighted path distance from the liver voxel to each source node.

[0149] In one embodiment of this application, the partitioning module 303 is configured to assign liver voxels to the source point with the smallest weighted path distance based on the weighted path distance from the liver voxel to each source point; wherein, liver voxels assigned to the same source point constitute an initial region; and multiple initial regions are merged based on a preset volume threshold to obtain multiple watershed partitions.

[0150] In one embodiment of this application, the segmentation module 304 is configured to determine the parameters of multiple standard liver segments based on preset liver segmentation rules; for each watershed partition: based on the parameters of each standard liver segment, calculate the similarity between the watershed partition and each standard liver segment; and based on the similarity, determine the target standard liver segment corresponding to the watershed partition.

[0151] In one embodiment of this application, the segmentation module 304 is configured to elastically deform the watershed partition based on the target standard liver segment and a preset three-dimensional boundary model of the target standard liver segment to obtain a deformed watershed partition.

[0152] This application provides an electronic device, including a memory and a processor. The memory stores an executable program, and the processor executes the executable program to perform the steps of the methods described in any of the above embodiments.

[0153] This application provides a computer program product that, when executed by a processor, implements the method as described in any of the above embodiments.

[0154] It should be understood that in the embodiments of this application, the processor may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0155] It should also be understood that the memory mentioned in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (Read-Only Memory). Only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0156] It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA, or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, the memory (storage module) is integrated into the processor.

[0157] It should be noted that the memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0158] In addition to the data bus, this bus may also include a power bus, a control bus, and a status signal bus. However, for clarity, all buses are labeled "bus" in the diagram.

[0159] It should also be understood that the first, second, third, fourth and various numerical designations used herein are merely for descriptive convenience and are not intended to limit the scope of this application.

[0160] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0161] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0162] In the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0163] Those skilled in the art will recognize that the various illustrative logical blocks (ILBs) and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0164] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0165] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0167] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0168] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automated liver segmentation modeling based on hybrid neural networks, characterized in that, include: The acquired raw liver images are preprocessed to obtain three-dimensional liver images; wherein, the three-dimensional liver images include liver parenchyma, portal vein and hepatic vein; The three-dimensional liver image is input into the trained segmentation model to obtain liver parenchyma mask, portal vein mask, and hepatic vein mask; Identify the connected regions in the portal vein mask; The connected region containing the largest number of voxels is identified as the portal vein trunk tree; Based on the portal vein trunk tree, a portal vein tree is generated; Based on the portal vein tree, multiple endpoints and multiple bifurcation points are identified; Based on a preset radius ratio threshold, key bifurcation points are determined from multiple bifurcation points; wherein, the multiple endpoints and the multiple bifurcation points constitute source points; wherein, the source points are used to characterize the source of blood perfusion; Based on the liver parenchyma mask, a graph structure is constructed; wherein, the liver voxels in the liver parenchyma mask are nodes, and the connections between the liver voxels and their adjacent liver voxels are edges; The weights of the edges are determined based on the liver parenchyma mask and the hepatic vein mask; Based on the graph structure and the edge weights, calculate the weighted path distance from the liver voxel to each source point; Based on the weighted path distance from the liver voxels to each source point, the liver parenchyma is divided into multiple watershed partitions; Based on the preset liver segmentation rules, the standard liver segments corresponding to each watershed partition are determined; Based on the standard liver segment corresponding to the watershed partition, the watershed partition is elastically deformed to obtain the deformed watershed partition.

2. The method as described in claim 1, Its features are, in, The segmentation model includes: an input module, an encoder, a feature fusion module, a decoder, and an output module; The input module is used to segment the three-dimensional liver image into multiple image blocks; The encoder is used to extract features of different sizes from the plurality of image blocks; The feature fusion module is used to fuse features of different sizes based on a channel attention mechanism; The decoder is used to generate a feature map with the same size as the image patch based on the fused features output by the feature fusion module. The output module is used to output the liver parenchyma mask, the portal vein mask, and the hepatic vein mask based on the feature map output by the decoder.

3. The method as described in claim 1, characterized in that, Based on the weighted path distance from the liver voxels to each source point, the liver parenchyma is divided into multiple watershed partitions, including: Based on the weighted path distance from the liver voxels to each source point, the liver voxels are assigned to the source point with the smallest weighted path distance; wherein, the liver voxels assigned to the same source point constitute the initial region; Multiple initial regions are merged based on a preset volume threshold to obtain multiple watershed partitions.

4. The method as described in claim 1, characterized in that, Based on preset liver segmentation rules, the standard liver segments corresponding to each of the aforementioned watershed partitions are determined, including: Based on the preset liver segmentation rules, the parameters of multiple standard liver segments are determined; For each watershed partition: based on the parameters of each standard liver segment, calculate the similarity between the watershed partition and each standard liver segment; based on the similarity, determine the target standard liver segment corresponding to the watershed partition.

5. The method as described in claim 4, characterized in that, Based on the standard liver segment corresponding to the watershed partition, the watershed partition is elastically deformed to obtain deformed watershed partitions, including: Based on the target standard liver segment and the preset three-dimensional boundary model of the target standard liver segment, the watershed partition is elastically deformed to obtain the deformed watershed partition.

6. A liver segmentation modeling system based on hybrid neural networks, characterized in that, include: The preprocessing module is configured to preprocess the acquired raw liver images to obtain a three-dimensional liver image; wherein the three-dimensional liver image includes liver parenchyma, portal vein and hepatic vein; The segmentation module is configured to input the three-dimensional liver image into a trained segmentation model to obtain liver parenchyma masks, portal vein masks, and hepatic vein masks; A partitioning module is configured to: identify connected regions in the portal vein mask; determine the connected region containing the largest number of voxels as the portal vein trunk tree; generate a portal vein tree based on the portal vein trunk tree; determine multiple endpoints and multiple bifurcation points based on the portal vein tree; determine key bifurcation points from the multiple bifurcation points based on a preset radius ratio threshold; wherein the multiple endpoints and multiple bifurcation points constitute source points; wherein the source points are used to characterize the source of blood perfusion; construct a graph structure based on the liver parenchyma mask; wherein liver voxels in the liver parenchyma mask are nodes, and the connections between a liver voxel and its adjacent liver voxels are edges; determine the edge weights based on the liver parenchyma mask and the hepatic vein mask; calculate the weighted path distance from the liver voxel to each source point based on the graph structure and the edge weights; and divide the liver parenchyma into multiple watershed partitions based on the weighted path distance from the liver voxel to each source point. The segmentation module is configured to determine the standard liver segment corresponding to each watershed partition based on a preset liver segmentation rule; and to perform elastic deformation on the watershed partition based on the standard liver segment corresponding to the watershed partition to obtain the deformed watershed partition.

Images