Liver and gall image segmentation method based on 2.5 D model and drainage basin analysis

Abstract

The invention provides a liver and gall image segmentation method based on a 2.5 D model and drainage basin analysis, and the method comprises the steps: constructing a voxel-level liver segment label graph through Couinaud liver segment structure analysis based on a portal vein center line; secondly, a 2.5 D model guided by structure perception is adopted for liver and gall structure segmentation, three-dimensional context information and a structure attention mechanism are fused in the model, and the discrimination ability at a complex junction is enhanced; then, in the intraoperative navigation stage, a stable registration relation is established by utilizing the structure number diagram and a visual image obtained in real time, dynamic posture adjustment is achieved, and the navigation stability is improved; and finally, the postoperative focus quantitative analysis module performs attribution analysis, volume quantification and credibility calculation on the focus area in combination with the structure numbering diagram and the model thermodynamic diagram. The whole process integrates dissection consistency, cross-stage adaptability and interpretability, and the requirements of the hepatobiliary surgery department for precise operation planning, real-time navigation and postoperative follow-up visit are met.

Description

A hepatobiliary image segmentation method based on 2.5D model and watershed analysis Technical Field

[0001] This invention belongs to the field of medical image processing, and particularly relates to a method for segmenting hepatobiliary images based on 2.5D models and watershed analysis. Background Technology

[0002] Surgical treatment of hepatobiliary diseases, especially malignant tumors such as liver cancer and cholangiocarcinoma, heavily relies on the accuracy of preoperative image analysis and the precision of intraoperative anatomical localization. In clinical practice, physicians typically need to identify the spatial location of liver parenchyma, blood vessels, bile ducts, and lesions based on preoperative CT images, while simultaneously determining the anatomical affiliation of the disease using Couinaud liver segment partitioning. However, most existing automatic segmentation technologies focus on pixel-level prediction based on image texture and contrast, lacking an explicit understanding of the blood flow topology of liver segments and failing to achieve structurally consistent segmentation at the anatomical blood supply level. While two-dimensional models have low computational overhead, they struggle to capture the complex cross-layer structural relationships of the liver; three-dimensional models, while possessing stronger structural representation capabilities, are significantly dependent on memory and computational resources, making them difficult to deploy clinically. The proposed two-dimensional-plus-three-dimensional model, while utilizing limited context, still learns based on pure image cues, failing to incorporate liver segment blood supply areas and portal vein structures as explicitly coded structural information into the model. This leads to classification instability or structural jumps at liver segment boundaries and lesion boundaries with blurred borders.

[0003] Meanwhile, while traditional watershed-based liver segmentation methods can construct Couinaud segments based on the portal vein centerline, they remain disconnected from existing segmentation models, lacking a linkage mechanism between structural priors and the segmentation network. Therefore, they cannot directly improve segmentation performance and are instead limited to an independent post-processing stage. With the increasing maturity of mixed reality technology in hepatobiliary surgery, preoperatively constructed 3D virtual models are used for intraoperative organ localization and surgical path planning. However, current registration methods rely on fixed calibration objects and initial position calibration, lacking correspondence with liver segment structures and failing to achieve dynamic correction when tissue displacement, organ surface deformation, or changes in camera perspective occur. Traditional image feature registration lacks robustness in soft tissue environments, making it difficult to guarantee the safety of intraoperative navigation. Furthermore, the interpretability of deep learning models is also a crucial factor affecting clinical application. Currently used heatmap methods mostly focus on visualization, lacking integration with structural localization and lesion quantification analysis, and thus cannot provide reliable credibility assessment for postoperative follow-up. Overall, existing technologies have significant shortcomings in terms of anatomical consistency, cross-stage information linkage, intraoperative registration stability, and the interpretability of lesion quantification. There is an urgent need for a method that can integrate portal vein basin structure, segmentation model, intraoperative registration, and lesion assessment into a unified framework to achieve intelligent support for the entire process from preoperative structural modeling and real-time intraoperative navigation to postoperative lesion quantification. Summary of the Invention

[0004] The purpose of this invention is to design a hepatobiliary image segmentation method based on 2.5D model and watershed analysis, which can be applied throughout the entire process of preoperative, intraoperative and postoperative procedures, and improve the anatomical consistency of the segmentation results.

[0005] To achieve the above objectives, this invention provides a hepatobiliary image segmentation method based on a 2.5D model and watershed analysis. The method includes: acquiring preoperative abdominal tomographic image data of the patient; extracting the portal vein vascular structure based on the abdominal tomographic image data and generating a portal vein centerline; identifying blood supply seed points for each liver segment based on the portal vein centerline, performing region growth starting from each liver segment blood supply seed point, and constructing a voxel-level liver segment label map by combining liver boundary constraints; fusing the liver segment label map with the abdominal tomographic image data as structural guidance information input to the 2.5D segmentation model; the 2.5D segmentation model uses a three-channel image composed of a central slice and its adjacent slices as input, and in the encoding... The decoder-device architecture incorporates a structural attention mechanism and a structural consistency regularization term. It outputs semantic segmentation results of the hepatobiliary organs and their lesions, spatially aligned with the input image data, along with a heatmap generated by the 2.5D segmentation model, and simultaneously generates a structural numbering map. The structural numbering map is used to extract the boundary points of liver segments, and structural matching relationships are established by combining intraoperative visual images. The rigid body transformation matrix is ​​optimized through weighted registration error and structural order consistency constraints to achieve pose alignment between the preoperative 3D segmentation model and the intraoperative scene. Based on the semantic segmentation results, the structural numbering map, and the heatmap, the liver segment attribution, volume ratio calculation, and credibility scoring of the lesion region are performed, generating structured postoperative quantitative analysis results.

[0006] Furthermore, the extraction of the portal vein vascular structure employs a grayscale threshold filtering and three-dimensional connected region growth algorithm, with the central voxel of the hepatic hilum region vascular region used as the seed point.

[0007] Furthermore, during the construction of the liver segment label map, region growth uses Euclidean distance as the expansion criterion, and when a voxel is simultaneously competed for by multiple liver segment regions, it is assigned to the liver segment corresponding to the nearest liver segment blood supply seed point.

[0008] Furthermore, the input of the 2.5D segmentation model consists of a three-channel image composed of a central slice and its adjacent slices above and below it, and the liver segment label image corresponding to the central slice is converted into a one-hot encoded image and then concatenated with the image channels.

[0009] Furthermore, the structural attention mechanism generates spatial weights based on the liver segment label map after each level of the encoder outputs a feature map, and dynamically reweights the feature channels.

[0010] Furthermore, the structural consistency regularization term suppresses classification jumps at the liver segment boundaries by imposing a smoothness constraint on the prediction probability of the liver segment boundary region.

[0011] Furthermore, the boundary points of the liver segments are obtained by detecting the positions of changes in liver segment numbers within the eight neighboring regions of the structural numbering diagram.

[0012] Furthermore, in the weighted registration error, the weight of each liver segment boundary point is set according to the variation range of the liver segment number in its neighborhood; the greater the variation range of the liver segment number, the higher the weight.

[0013] Furthermore, the structural sequence consistency constraint prevents mismatch of non-anatomical sequences by restricting the coordinate arrangement direction of the boundary points of adjacent liver segments on the intraoperative visual image projection plane to be consistent with the natural order in the anatomical structure.

[0014] Furthermore, the credibility score is jointly determined by the mean of the model heatmap response within the lesion area and the variance of the model heatmap response. A higher credibility score is achieved when the model heatmap response is concentrated and consistent.

[0015] The beneficial technical effects of this invention are at least as follows: Addressing the aforementioned problems, this invention provides a hepatobiliary image segmentation method based on a 2.5D model and watershed analysis. By explicitly constructing the Couinaud liver segment structure based on the portal vein central line as a voxel-level structural label map, the segmentation model possesses the ability to perceive the spatial region of the liver segment during the learning phase, fundamentally improving the anatomical consistency of the segmentation results. This invention introduces a structural attention mechanism and structural consistency regularization into the segmentation model, enabling the model to have more robust discrimination capabilities at liver segment boundaries and complex organ junctions. In the intraoperative navigation stage, this invention extracts stable anatomical boundary points through structural numbering maps and establishes structural matching relationships by combining intraoperative images, constructing a registration optimization model based on structural order constraints. This allows the 3D virtual model to achieve dynamic posture adjustment even with slight organ displacement or viewpoint changes, significantly improving the stability of intraoperative navigation. In the postoperative evaluation stage, this invention proposes a lesion quantification method combining structural numbering maps and model heatmaps for liver segment-level attribution analysis, volume quantification, and reliability calculation of lesion regions, achieving a unified expression between lesion detection results and anatomical structure and model interpretability. By systematically integrating liver segment labeling, structure-aware segmentation, structure-driven registration, and structure-dependent lesion quantification, this invention constructs an intelligent hepatobiliary imaging analysis framework that combines anatomical consistency, cross-stage adaptability, and interpretability, thereby meeting the key needs of hepatobiliary surgery for preoperative planning, intraoperative navigation, and postoperative follow-up. Attached Figure Description

[0016] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1 is a flowchart of the hepatobiliary image segmentation method based on 2.5D model and watershed analysis of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] In one or more embodiments, as shown in Figure 1, a hepatobiliary image segmentation method based on a 2.5D model and flow domain analysis is disclosed. The method includes the following: S1: Acquiring preoperative abdominal tomographic image data of the patient; extracting the portal vein vascular structure based on the abdominal tomographic image data and generating the portal vein centerline; identifying the blood supply seed points of each liver segment according to the portal vein centerline, performing region growth starting from each liver segment blood supply seed point, and constructing a voxel-level liver segment label map in combination with liver boundary constraints; Specifically, this step aims to construct an anatomically consistent liver segment label map based on the preoperative abdominal medical tomographic image data, which is used for structural region guidance of the subsequent segmentation model. This label map analyzes the blood flow path of the portal vein, combines the anatomical zoning rules of the liver segments, and assigns a voxel number to each liver segment, thereby forming a spatially consistent and topologically correct three-dimensional structural layer, providing a basis for structural perception learning and intraoperative spatial localization of the segmentation results.

[0020] The input is the abdominal transverse image sequence obtained before the patient's surgery, denoted as... ,in , , These represent the row, column, and slice number of the image, respectively. This image data was acquired using a spiral computed tomography (CT) scanner, typically with a slice thickness of 0.75 mm to 1.0 mm and a spatial resolution of [missing information]. Pixels, with grayscale values ​​in standard radiometric attenuation intensity. The data has already undergone isotropic resampling and image normalization in the preprocessing stage; therefore, this step directly analyzes the normalized 3D volumetric data.

[0021] This step mainly includes three processing stages: portal vein structure extraction, blood flow path analysis, and liver segment expansion. First, in the image sequence... The portal vein vascular structure was extracted. This process was implemented based on voxel value range filtering and a 3D connected region growing algorithm. Typically, the voxel grayscale values ​​of the portal vein are distributed within... Within the interval (e.g.) , By selecting the central voxel of the blood vessels in the porta hepatis region as the seed point, combined with three-dimensional... The neighborhood growth rule extracts the maximum connected region, denoted as . This area represents the initial distribution range of the portal vein.

[0022] Subsequently, the extracted portal vein region Perform a centerline extraction operation to obtain the blood flow path diagram of the portal vein. The centerline extraction employs a 3D skeletonization algorithm to obtain the central axis structure. Then, based on the anatomical rules of liver segments, it is identified... Construct a seed point set based on the endpoints of typical portal vein branches. Each point This corresponds to a blood supply center for a specific liver segment.

[0023] With each Starting from the liver segment, a region growing algorithm is executed to construct the corresponding liver segment region. The region expansion rules are as follows: ;in Indicates the first During the nth iteration A set of voxels for each liver segment region Represents the three dimensions of the set Neighborhood voxel set, For the nearest voxels to be determined, Indicates the first The blood supply seed points of each liver segment for and Euclidean distance It is the maximum growth radius of the region, controlling the maximum extent of expansion of a single liver segment, and is usually set to... to Individual element unit.

[0024] If candidate voxels Other liver segments If there is an overlap, assign it to the nearest one. Corresponding region. The conditions for growth termination are: first, the voxel exceeds the liver boundary (judged by the mask provided by the liver parenchyma segmentation results); second, the voxel has been absorbed by other liver segments.

[0025] Finally, a complete liver segment labeling map is formed. Its definition is as follows: ;in Voxel representation Liver segment numbering, For indicator functions, when Time to take Otherwise take The number range is... ,in Generally This is consistent with the anatomy of the liver segments.

[0026] The output of this step is: It is a three-dimensional voxel-level integer label map whose spatial coordinates are related to the input image. Fully aligned.

[0027] S2: The liver segment label map is fused with the abdominal tomographic image data and used as structural guidance information input to the 2.5D segmentation model. The 2.5D segmentation model uses a three-channel image composed of a central slice and its adjacent slices as input, and introduces a structural attention mechanism and a structural consistency regularization term into the encoder-decoder structure. It outputs semantic segmentation results of the hepatobiliary organs and their lesions that are spatially aligned with the input image data, as well as a heatmap generated by the 2.5D segmentation model, and simultaneously generates a structural number map. Specifically, this step aims to realize the construction of a two-and-a-half three-dimensional hepatobiliary structural segmentation model under structure-aware guidance, focusing on the liver segment structural label map generated in the previous step. By incorporating a segmentation model, the model gains the ability to perceive the anatomical regions of liver segments. In this approach, the model uses the liver segment labels generated in the previous step based on portal vein flow domain analysis as a guide. By enhancing the independent modeling of tissue features within different liver segment structures, it achieves semantic segmentation of the liver and its accessory structures with consistent structures and clear boundaries, and provides traceable structural information for subsequent intraoperative registration, lesion quantification, and other tasks.

[0028] The model's input includes the liver segment label map output from the previous stage. and normalized abdominal tomographic image sequence Considering the significant continuity of cross-layer structures but limited number of channels in hepatobiliary CT images, this model adopts a two-and-a-half-three-dimensional input method, with each input starting from the central layer. The two adjacent floors above and below it and The image consists of three layers, stitched together into a three-channel image input tensor. Simultaneously extract The layer, representing the liver segment structure map corresponding to the current segmentation target image, is encoded using integers to generate a one-hot encoded form. , This represents the total number of liver segments. This structural information is then concatenated with the image to form a fused input tensor. ;in It is a complete input tensor containing three layers of image grayscale and spatial location encoding for each liver segment, ensuring that the network can identify the anatomical location of each structural region during the learning process.

[0029] The model employs an encoder-decoder structure, with the encoder consisting of four levels of two-dimensional convolutional blocks, each level consisting of two... It consists of convolutional layers, batch normalization layers, and activation functions, with the number of convolutional kernels starting from the first layer. Increased to the fourth level Structural diagram information It not only participates in the input but also incorporates a structural channel attention mechanism in the intermediate feature fusion module. After each encoder output, the structural attention module spatially reweights the features. This structural attention mechanism uses a structural graph as the basis for generating channel weights, enabling dynamic adjustment of channels in different liver segments.

[0030] To enhance the model's ability to identify liver segment boundary transition zones, a structural consistency regularization term is further introduced, which modifies the class map predicted by the model. With liver segment label map The structural boundary correspondence between them is used as the regularization target. A boundary region mask is constructed. , representing the location in the eight-neighborhood where the liver segment number changes, this mask is used to construct the weighted boundary penalty term. The final loss function consists of the basic cross-entropy loss. With structural regularization terms constitute: ;in The structural regularization coefficient controls the importance of the consistency of liver segment boundaries; its empirical value is set to [value missing]. to The structural regularization term is defined as: This item imposes a smoothness constraint on the predicted probability map of the boundary location to prevent the model from generating excessive fluctuations in the structural boundary region. For position The set of eight neighboring domains, This is the predicted probability vector for the current pixel. This indicates whether the location is a structural boundary point (if it is 1, then proceed to regularization calculation).

[0031] The model's final output consists of two parts: the first part is a 3D segmentation map. Each voxel value corresponds to an organ category number, such as liver parenchyma, portal vein, bile duct, lesion, etc.; the second part is the structural numbering diagram. directly by It was copied and is used for subsequent intraoperative navigation structure alignment and lesion attribution determination.

[0032] S3: Extract the boundary points of liver segments using the structural numbering map, establish structural matching relationships by combining intraoperative visual images, and optimize the rigid body transformation matrix through weighted registration error and structural order consistency constraints to achieve pose alignment between the preoperative 3D segmentation model and the intraoperative scene. Specifically, this step is used in hepatobiliary surgery to spatially align the preoperative segmentation model results with the patient's actual liver anatomy, achieving accurate positioning and real-time navigation capabilities of the virtual hepatobiliary model in mixed reality devices. This step relies on the semantic segmentation map output in step two. and structural numbering diagram Combined with visual images obtained during the operation By constructing a structural consistency mapping relationship, the transformation matrix is ​​calculated. It also updates the pose of the segmentation model and outputs the final visualized 3D model for intraoperative mixed reality navigation. .

[0033] Input in progress, This is the semantic segmentation map of the liver and gallbladder structure output from the previous step. Each voxel represents a category number (such as liver parenchyma, portal vein, bile duct, lesion, etc.). A liver segment numbering diagram consistent with its coordinates is used to represent the Couinaud liver segment to which each voxel belongs, and the two are spatially consistent; These are two-dimensional color image frames acquired by intraoperative mixed reality terminals (such as binocular camera devices). The images have been calibrated to the world coordinate system and possess depth or spatial reconstruction capabilities. These images are used to construct the structural correspondence between the images and the three-dimensional segmentation model.

[0034] This step begins with the structural numbering diagram. Extracting the set of boundary points at the junction of liver segments ,in This represents the coordinates of three-dimensional points on the boundary of liver segment structures in model space. Each boundary point is obtained by detecting changes in structure numbering within its eight neighborhoods, ensuring that the point set is distributed at the anatomical interface. This point set is then projected onto the camera projection matrix. Downmapped to the intraoperative image plane to form the desired alignment point Meanwhile, in intraoperative images By using edge detection and anatomical region segmentation, significant structural boundary points are extracted from actual images. .

[0035] Because intraoperative images may contain partial occlusion, soft tissue deformation, or registration errors, this step designs a weighted matching model based on the structural stability of liver segments to ensure structural matching stability. A structural boundary order regularization term is introduced to suppress non-anatomical order mismatches. First, the basic registration error is defined as follows: ;in Represents the rigid body transformation matrix from the 3D model to the intraoperative image, including rotation and translation; It is a structural point The confidence weight is defined as: This weight depends on the rate of change of the structure number. This is the boundary sensitivity coefficient. Larger variations in structure numbering indicate that the point is located at an important anatomical boundary and should be assigned a higher matching weight.

[0036] To further reduce spatial jump problems caused by mismatched liver segment order, a structural order consistency regularization term is introduced. This constraint ensures that the orientation of the liver segment boundary points on the projection plane is consistent with their natural orientation in the anatomical structure. The definition is as follows: ;where the set This represents all spatially adjacent pairs of liver segment boundary points. Represents a symbolic function. Point In the image The coordinate components of the axis, This indicates that the value is 1 for items with inconsistent signs, and 0 otherwise. This regularization term is used to restrict the matching point pairs to maintain consistency between the structural numbering order and the spatial location order in the image.

[0037] The ultimate optimization objective is to simultaneously minimize the matching error and the structural order mismatch term: ;in The regularization weights control the degree to which the structural order constrains the registration result. A gradient descent-based rigid body pose optimization algorithm is used to solve the problem. In the initial preoperative calibration matrix Fine-tuning and updates are made based on this. The pose adjustment results are used to drive the generation of the 3D structure voxel model from the segmentation model. Updated to The output includes two variables, one of which is the final transformation matrix. The first is used to control the spatial relationship between the 3D model and the intraoperative scene in the mixed reality system; the second is the 3D visual model after posture adjustment. It is used for visualization, path planning, and lesion identification in navigation systems.

[0038] S4: Based on the semantic segmentation results, the structural numbering map, and the heatmap, the liver segment attribution, volume ratio calculation, and credibility score are performed on the lesion region to generate structured postoperative quantitative analysis results. Specifically, this step is used to perform postoperative quantitative analysis of the lesion region generated by the preoperative segmentation model at the liver segment level. Combined with the model's interpretive heatmap, structured analysis results with anatomical attribution and credibility assessment capabilities are output. The main goal of this step is to solve the problems of "ambiguous lesion location," "difficulty in quantifying credibility," and "weak anatomical correlation" existing in current deep segmentation models in clinical postoperative assessment. By jointly processing the segmentation output with the liver segment structural labeling map and the attention heatmap generated by the model, each suspicious lesion is assigned a clear anatomical number, spatial location, volume estimate, and credibility score, thereby assisting doctors in completing lesion residual assessment, postoperative follow-up strategy formulation, and risk visualization.

[0039] The input for this step consists of three parts, all derived from the output of the previous step and aligned to the intraoperative coordinate space after pose updates. The first is the hepatobiliary semantic segmentation map. The output is from the structure-aware segmentation model, where each voxel represents its category number. The lesion category number is uniformly set in the system as follows: (e.g., set to 5). The second is the structure numbering diagram. This is used to characterize the Couinaud liver segment number to which each voxel belongs; the numbering range is typically [range missing]. to The third is the three-dimensional heat map generated during the model interpretation phase. Each voxel value reflects the model's level of attention to the location as a lesion, ranging from [value missing]. . It is generated by the interpretation module of the structure-aware model in step two, and is generated using a channel gradient weighted attention mechanism, such as the response of the Grad-CAM mechanism at the sensitive location of the structural boundary.

[0040] To make this step feasible in clinical practice, the structural numbering diagram was first segmented into liver segments to obtain a voxel set of eight liver segments. Each set represents the first set. All voxels within each liver segment. Extracting the lesion area The volume ratio index is obtained by calculating the ratio of the number of lesion bodies to the total number of body bodies in the liver segment. This indicator reflects the proportion of lesions in each liver segment and is one of the core indicators in lesion quantification analysis. To address issues such as false positives and ambiguous boundaries, it is necessary to further incorporate interpretability information from the heatmap. Within each liver segment, the mean heatmap response intensity of the lesion area is calculated, and a heatmap reliability factor is designed based on the spatial dispersion of heatmap values. The first term in this formula is the average thermal response value of the lesion area, and the second term is the variance of the thermal response within the lesion. The discrete penalty coefficient (usually set to) to Its physical meaning is: if the model's attention to the lesion area is focused and consistent, then... A larger value indicates higher reliability; if the attention distribution is discrete, or high responses occur in a few locations, then... A decrease reflects forecast instability.

[0041] The final comprehensive scoring index is defined as follows: ,when Exceeding the set threshold (like This indicates the presence of a structurally distinct and highly relevant lesion in the liver segment, suggesting postoperative clinical monitoring or further examination. For example, a lesion was found in the fourth liver segment (middle liver segment) postoperatively. , ,get higher than If so, the lesion segment is marked as a "high-confidence residual lesion" and highlighted in red in the visualized 3D model.

[0042] This invention also provides a hepatobiliary image segmentation device based on a 2.5D model and watershed analysis, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps as described in the above embodiments of the hepatobiliary image segmentation method based on a 2.5D model and watershed analysis, such as steps S1 to S4 in Figure 1; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0043] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the hepatobiliary image segmentation device based on 2.5D model and watershed analysis.

[0044] The hepatobiliary image segmentation device based on 2.5D model and watershed analysis can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the device may also include input / output devices, network access devices, buses, etc.

[0045] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASACs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the hepatobiliary image segmentation device based on 2.5D model and watershed analysis, connecting various parts of the device via various interfaces and lines.

[0046] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the hepatobiliary image segmentation device based on 2.5D model and watershed analysis by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the operation of the air conditioning controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0047] The integrated module of the hepatobiliary image segmentation device based on 2.5D model and watershed analysis, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0048] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0049] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A hepatobiliary image segmentation method based on 2.5D model and watershed analysis, characterized in that, The method includes: acquiring preoperative abdominal tomographic images of the patient; extracting the portal vein vascular structure based on the abdominal tomographic images and generating a portal vein centerline; identifying blood supply seed points for each liver segment based on the portal vein centerline, performing region growth starting from each blood supply seed point, and constructing a voxel-level liver segment label map by combining liver boundary constraints; fusing the liver segment label map with the abdominal tomographic images and inputting it as structural guidance information into a 2.5D segmentation model; the 2.5D segmentation model uses a three-channel image composed of a central slice and its adjacent slices as input, and introduces a structural attention mechanism and structure into the encoder-decoder structure. The consistency regularization term outputs semantic segmentation results of the hepatobiliary organs and their lesions that are spatially aligned with the input image data, as well as a heatmap generated by the 2.5D segmentation model, and simultaneously generates a structural numbering map. The structural numbering map is used to extract the boundary points of liver segments, and structural matching relationships are established by combining intraoperative visual images. The rigid body transformation matrix is ​​optimized through weighted registration error and structural order consistency constraints to achieve pose alignment between the preoperative 3D segmentation model and the intraoperative scene. Based on the semantic segmentation results, the structural numbering map, and the heatmap, the liver segment attribution, volume ratio calculation, and credibility score of the lesion region are performed, generating structured postoperative quantitative analysis results.

2. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The extraction of the portal vein vascular structure was performed using a grayscale threshold filtering and three-dimensional connected region growth algorithm, with the central voxel of the hepatic hilum vascular region as the seed point.

3. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, During the construction of the liver segment label map, region growth is based on Euclidean distance as the expansion criterion, and when a voxel is simultaneously competed for by multiple liver segment regions, it is assigned to the liver segment corresponding to the nearest liver segment blood supply seed point.

4. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The input to the 2.5D segmentation model consists of a three-channel image composed of a central slice and its adjacent slices above and below it. The liver segment label image corresponding to the central slice is converted into a one-hot encoded image and then concatenated with the image channels.

5. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The structural attention mechanism generates spatial weights based on the liver segment label map after each level of the encoder outputs a feature map, and then dynamically reweights the feature channels.

6. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The structural consistency regularization term suppresses classification jumps at the junctions of liver segments by imposing a smoothness constraint on the predicted probability of the liver segment boundary region.

7. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The boundary points of the liver segments are obtained by detecting the positions of changes in liver segment numbers within the eight neighboring regions of the structural numbering diagram.

8. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, In the weighted registration error, the weight of each liver segment boundary point is set according to the variation range of the liver segment number in its neighborhood. The greater the variation range of the liver segment number, the higher the weight.

9. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The structural sequence consistency constraint restricts the coordinate arrangement direction of the boundary points of adjacent liver segments on the intraoperative visual image projection plane to be consistent with the natural order in the anatomical structure, thus preventing mismatch of non-anatomical order.

10. The hepatobiliary image segmentation method based on a 2.5D model and watershed analysis according to claim 1, characterized in that, The credibility score is determined by the mean of the model heatmap response within the lesion area and the variance of the model heatmap response. A higher credibility score is achieved when the model heatmap response is concentrated and consistent.

Images