Kidney three-dimensional modeling-based MAP scoring method and device

Abstract

The invention discloses an MAP scoring method and device based on kidney three-dimensional modeling, and belongs to the field of medical image processing. The method comprises the following steps: acquiring kidney medical image data, wherein the kidney medical image data comprises kidney CT image data; preprocessing the kidney medical image data; generating a complete kidney three-dimensional model based on the kidney medical image data; calculating the centroid positions of all voxels in the renal vein three-dimensional model, and taking the Z coordinate of the centroid as a standard measurement plane; identifying an anatomical region needing to be measured on the standard measurement plane based on a deep learning segmentation model; extracting an outer contour of the kidney based on the kidney three-dimensional model, determining a posterior abdominal wall contour based on the front edge of the lumbar major muscle, and calculating the shortest distance between the outer contour of the kidney and the posterior abdominal wall contour; and determining a perirenal fat region, enhancing the strip-shaped structure by using a multi-scale filter, and calculating the proportion of the strip-shaped structure in the perirenal fat region. According to the invention, objective quantitative evaluation of perirenal fat adhesion risk is realized.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, three-dimensional medical modeling, and computer-aided diagnosis, and in particular to a MAP scoring method and device based on three-dimensional kidney modeling. Background Technology

[0002] In laparoscopic partial nephrectomy, adherent perinephric fat (APF) is a key factor affecting surgical difficulty and prognosis. Adhesive perinephric fat can lead to: prolonged operation time (30-60 minutes); increased intraoperative blood loss; increased risk of renal capsule tearing; and increased postoperative complication rate. Therefore, accurate preoperative assessment of the degree of perinephric fat adhesion is crucial for surgical planning.

[0003] The MAP score, developed by the Mayo Clinic in the United States, predicts the probability of perirenal fat adhesion using two key indicators: P-value (Posterior perinephric fat thickness): the shortest distance from the posterior abdominal wall at the level of the renal vein to the renal capsule; and perinephric stranding: fibrous, high-density strands within the perirenal fat space.

[0004] The scoring criteria for the P-value are as follows: when the shortest distance from the retroperitoneal wall at the level of the renal vein to the renal capsule is less than 1.0 cm, the score is 0; when the shortest distance from the retroperitoneal wall at the level of the renal vein to the renal capsule is 1.0-1.9 cm, the score is 1; and when the shortest distance from the retroperitoneal wall at the level of the renal vein to the renal capsule is greater than or equal to 2.0 cm, the score is 2.

[0005] The scoring criteria for perirenal fat strands are as follows: 0 points for no strands, in which case the perirenal fat on the CT image appears completely black with no high-density strands; 2 points for mild / moderate strands, in which case fine strands are present on the CT image, but no large inflammatory strands; and 3 points for severe strands, in which case large, dense strand-like structures are present on the CT image.

[0006] Traditional MAP scoring relies on manual measurement by physicians on two-dimensional CT images, which has the following limitations:

[0007] (1) Inaccurate anatomical localization: On two-dimensional CT cross-sectional images, doctors need to rely on experience to determine the central level of the renal vein. Different doctors may choose measurement levels that differ by 5-10 levels, resulting in a difference of 20-30% in the P-value measurement results;

[0008] (2) Difficulty in understanding spatial relationships: The kidney is an irregular three-dimensional organ, and its spatial relationship with the posterior abdominal wall varies greatly at different levels. Two-dimensional images cannot fully reflect this three-dimensional spatial relationship, which can easily lead to measurement errors;

[0009] (3) One-sided assessment of the distribution of cords: Perirenal fat cords are distributed in a three-dimensional network. Observing only a single layer cannot accurately assess their severity and distribution range;

[0010] (4) High subjectivity: The selection of measurement points and the judgment of the severity of the strips depend on the doctor's subjective judgment, and the consistency between different doctors is poor (Kappa value is only 0.65-0.70).

[0011] (5) Inefficient: It takes 10-15 minutes to complete the MAP score of a patient, including steps such as finding the appropriate level, measuring distance, and evaluating the strips;

[0012] (6) Cannot be processed in batches: Manual methods are not suitable for large-scale clinical studies and retrospective analyses;

[0013] (7) The severity of the stripe is subjectively assessed by visually observing the shape, density and distribution of the stripes in the CT image, and it relies entirely on the doctor's visual judgment. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of the prior art and provide a MAP scoring method and device based on three-dimensional modeling of the kidney. By accurately reconstructing the kidney and surrounding anatomical structures in three dimensions, an objective quantitative assessment of the risk of perirenal fat adhesion can be achieved.

[0015] This invention is achieved through the following technical solution:

[0016] The first aspect of this invention discloses a MAP scoring method based on three-dimensional kidney modeling, comprising:

[0017] Acquire renal medical imaging data, including renal CT image data;

[0018] Preprocessing of renal medical imaging data;

[0019] Generate a complete 3D model of the kidney based on renal medical imaging data;

[0020] Calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and use the Z coordinate of the centroid as the standard measurement surface;

[0021] Identify the anatomical regions to be measured on a standard measurement plane based on a deep learning segmentation model;

[0022] The outer contour of the kidney is extracted based on the three-dimensional model of the kidney, the contour of the posterior abdominal wall is determined based on the anterior edge of the psoas major muscle, and the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall is calculated.

[0023] The perirenal fat region was identified, and the cord-like structures were enhanced using a multi-scale filter. The proportion of the cord-like structures in the perirenal fat region was calculated.

[0024] Furthermore, the renal medical imaging data undergoes preprocessing, including:

[0025] Convert kidney CT image data to the RAS coordinate system.

[0026] The kidney CT image data was interpolated and resampled to generate isotropic data.

[0027] Furthermore, a complete three-dimensional model of the kidney is generated based on renal medical imaging data, including:

[0028] A deep learning segmentation model was used to segment renal cortex and renal medulla from renal medical imaging data. Based on the renal cortex and renal medulla data, a three-dimensional model of renal parenchyma was generated using the Marching Cubes algorithm. When generating the three-dimensional model of renal parenchyma using the Marching Cubes algorithm, isosurfaces were extracted by looking up tables, more voxels were added in areas with greater curvature, and triangular patches that met preset conditions were removed by mesh simplification.

[0029] The renal vein region data was segmented from renal medical imaging data using a deep learning segmentation model, and a three-dimensional model of the renal vein was generated by a three-dimensional reconstruction algorithm to determine the spatial course of the renal vein.

[0030] Three-dimensional segmentation data of the psoas major muscle was extracted from renal medical imaging data using a deep learning segmentation model (such as the 3D-U-Net model), and the three-dimensional surface of the posterior abdominal wall was determined based on the three-dimensional segmentation data of the psoas major muscle.

[0031] Furthermore, generating a complete three-dimensional model of the kidney based on renal medical imaging data also includes:

[0032] The surface of a 3D model of kidney parenchyma is smoothed by combining Laplacian smoothing, volume-invariant constrained smoothing, and feature-preserving anisotropic smoothing.

[0033] Further, the spatial course of the renal vein was determined, including:

[0034] The central path of the renal vein was extracted from the three-dimensional model of the renal vein using a blood vessel centerline extraction algorithm.

[0035] Spatial analysis was used to determine the path of the renal vein from the kidney to the inferior vena cava, as well as the spatial orientation of the renal vein within the kidney.

[0036] By combining the three-dimensional centerline and anatomical structure of the renal vein, the course of the renal vein in different regions was analyzed, including its direction, curvature, and relationship with surrounding tissues.

[0037] Furthermore, based on a deep learning segmentation model, the anatomical regions to be measured are identified on a standard measurement plane, including:

[0038] Extracting two-dimensional slice images of standard measurement planes from renal medical imaging data;

[0039] Preprocess the extracted two-dimensional slice image;

[0040] Two-dimensional slice images are input into a deep learning segmentation model. The deep learning segmentation model automatically labels and segments different anatomical regions in the image and generates segmentation results containing labels for each anatomical region.

[0041] Morphological processing is performed on the segmentation results output by the deep learning segmentation model;

[0042] The segmentation results output by the deep learning segmentation model are smoothed using Laplacian smoothing or anisotropic smoothing algorithms that preserve features.

[0043] Extract the anatomical regions that need to be measured from the segmentation results;

[0044] Geometric analysis was performed on the extracted anatomical regions to calculate anatomical parameters.

[0045] Further, the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall is calculated, including:

[0046] Extract all point sets of the posterior abdominal wall contour, and construct a KD tree data structure using the posterior abdominal wall contour point set. Each node of the KD tree stores one point.

[0047] Traverse all points on the outer contour of the kidney. For each kidney capsule point, use the nearest neighbor query method of KD tree to find the nearest point on the posterior abdominal wall contour.

[0048] For each renal capsule point and its nearest posterior abdominal wall point, calculate the Euclidean distance between them;

[0049] Calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall. This shortest distance is the minimum value among all distances between points on the renal capsule and points on the posterior abdominal wall.

[0050] Furthermore, the perirenal fat region was identified, and the cord-like structures were enhanced using a multi-scale filter. The proportion of the cord-like structures in the perirenal fat region was calculated, including:

[0051] The area within a preset range surrounding the kidney is defined as the perirenal fat region;

[0052] Design a multi-scale Frangi filter bank with a scale range of 1.0-3.0 mm;

[0053] A multi-scale Fragi filter was applied to renal medical imaging data, and image processing was performed at each scale. A corresponding response map was generated for each scale to enhance the regions in the image that conform to tubular features.

[0054] The response results of the Frangi filter at different scales are fused;

[0055] Differentiating between strands and blood vessels based on CT values ​​and morphological characteristics;

[0056] Output the strip-like structure region and its location;

[0057] Calculate the volume of the cord-like structures and the volume of the perirenal fat region;

[0058] The proportion of the cord-like structure to the perirenal fat region is calculated based on the volume of the cord-like structure and the volume of the perirenal fat region.

[0059] Furthermore, the MAP scoring method also includes:

[0060] After calculating the shortest distance on the standard measurement surface, supplementary measurements are performed on the two layers above and below the standard measurement surface.

[0061] If the difference between the supplementary measurement results of the upper and lower layers exceeds 10%, the measurement range is expanded, and the median of these measurement results is taken as the final measurement value. This can reduce the impact of outlier data and improve the accuracy of the final result.

[0062] A second aspect of the present invention discloses a MAP scoring device based on three-dimensional kidney modeling, comprising:

[0063] The data acquisition module is used to acquire renal medical imaging data, including renal CT image data.

[0064] The data preprocessing module is used to preprocess renal medical imaging data;

[0065] The 3D modeling module is used to generate a complete 3D model of the kidney based on renal medical imaging data.

[0066] The centroid calculation module is used to calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and uses the Z coordinate of the centroid as the standard measurement surface.

[0067] The anatomical region identification module is used to identify the anatomical regions to be measured on a standard measurement plane based on a deep learning segmentation model.

[0068] The shortest distance calculation module is used to extract the outer contour of the kidney based on the three-dimensional model of the kidney, determine the contour of the posterior abdominal wall based on the anterior edge of the psoas major muscle, and calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall.

[0069] The cord analysis module is used to identify the perirenal fat region, enhance cord-like structures using a multi-scale filter, and calculate the proportion of cord-like structures in the perirenal fat region.

[0070] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0071] (1) Traditional two-dimensional MAP scoring methods rely on cross-sectional slice data of CT images. The spatial relationship between the kidney and its surrounding structures cannot be fully presented by two-dimensional images, especially in the spatial location of complex structures such as renal veins and renal capsule. Two-dimensional CT slice data can only reflect the local situation of a certain slice level. It is easy to produce projection errors due to different slice selections. It cannot fully show the actual spatial relationship between the kidney and its surrounding structures, leading to the risk of misdiagnosis and missed diagnosis.

[0072] This invention generates a complete three-dimensional kidney model by utilizing segmentation information of the renal cortex and renal medulla from CT data. The spatial information of the kidney and its surrounding structures (such as the renal veins and retroperitoneal wall) is fully preserved and presented, avoiding information omissions caused by layer selection in traditional two-dimensional slices. Through comprehensive analysis of the three-dimensional data, the relative position between the kidney and surrounding anatomical structures can be assessed more accurately, improving the comprehensiveness and accuracy of the assessment and ensuring that the assessment results are more objective and reliable.

[0073] (2) Traditional MAP scoring methods often rely on the doctor's experience and visual judgment when selecting the measurement plane. Different doctors may have significant differences in their selection of the location of the renal vein center, resulting in inconsistent measurement planes and thus affecting the accuracy of the measurement results. In terms of the location of the renal vein, traditional MAP scoring methods usually rely on a certain anatomical feature of the kidney to select the measurement plane. However, this method is highly subjective and may be subject to deviations due to individual patient differences.

[0074] This invention proposes a standardized measurement plane selection method based on the three-dimensional centroid of the renal vein. By accurately modeling the three-dimensional renal vein in CT images, the coordinates of all voxels are extracted, and the centroid position of the renal vein is calculated. Then, based on the Z-coordinate of this centroid position, a standard cross-section perpendicular to the long axis of the human body is determined as the measurement plane. This method completely eliminates the subjective factors in traditional methods, ensuring consistent plane selection standards for each measurement, guaranteeing high consistency and repeatability of measurement results, and ensuring more accurate kidney measurements unaffected by differences in physician experience.

[0075] (3) In the traditional manual measurement process, doctors need to rely on experience to select the appropriate layer and measure the anatomical area. This method has a large degree of subjectivity and inconsistency, resulting in a large error in the measurement results. The Kappa value is usually 0.65-0.70. Due to the difference in doctors' operation, the measurement error can reach ±3mm or even greater. In addition, in the traditional two-dimensional method, the measurement of cord-like structures depends on the doctor's interpretation of different CT slices, which can easily lead to missed or misdiagnosis of cord-like structures.

[0076] This invention significantly improves the accuracy and consistency of MAP scores by combining 3D modeling technology with automated processes. A modified Marching Cubes algorithm is used to accurately generate a 3D kidney model, and a KD tree acceleration algorithm is used to optimize the shortest distance calculation of the posterior abdominal wall contour, achieving a high accuracy of ±0.5mm for each measurement. Simultaneously, standardized measurement procedures and automated calculations eliminate human error, greatly improving measurement accuracy and consistency. Experimental data demonstrate that the Kappa value increases from the traditional 0.65-0.70 to over 0.95, significantly improving the consistency and reliability of MAP scores.

[0077] (4) Traditional manual measurement methods have high computational complexity and processing time. Since each case's MAP score requires doctors to perform calibration, select the plane, and measure the distance one by one, the whole process takes 10-15 minutes, which is very inefficient in large-scale clinical data processing and retrospective analysis. Moreover, due to the high computational complexity, traditional methods cannot support high-throughput batch data processing;

[0078] This invention significantly improves computational efficiency through algorithm optimization. Specifically, it employs the KD-tree algorithm for spatial indexing of the posterior abdominal wall contour points, reducing the complexity of shortest distance calculation from O(n²) to O(n log n), thus significantly reducing the computational burden. Furthermore, it enhances model generation speed through an accelerated Marching Cubes algorithm. The processing time for a single case is reduced from 10-15 minutes to less than 30 seconds, supporting high-throughput processing of 100 cases per hour, greatly improving the efficiency of clinical data processing. This automated processing method is of great significance for large-scale data analysis and clinical research.

[0079] (5) In the traditional MAP scoring method, the assessment of cord-like structures usually relies on the doctor's subjective judgment of CT images. Due to the large differences in the thickness and shape of cord-like structures, the single-scale cord detection method is prone to missing small lesions or misjudging blood vessels as cords, leading to missed diagnosis or misdiagnosis;

[0080] This invention employs a multi-scale Frangi filter bank for cord detection, automatically selecting an adaptive scale based on the varying thickness of the cords, thereby effectively improving the detection rate of subtle lesions. Through multi-scale detection, the system can simultaneously identify both fine and coarse cord-like structures at different scales, thus enhancing detection sensitivity and specificity and avoiding missed diagnoses common in traditional methods. The combination of CT values ​​and morphological features further enhances the ability to distinguish cords from blood vessels, ensuring the accuracy of the detection results.

[0081] (6) In traditional methods, the selection of measurement planes and the extraction of anatomical landmarks often rely on the doctor's experience, leading to significant errors in the measurement results. Especially for the complex structure of the kidney, manually selected slice layers may miss key anatomical landmarks;

[0082] This invention calculates the centroid of a three-dimensional model of the kidney and renal vein to accurately determine the standard measurement plane, completely avoiding errors that may arise when doctors select slides. Simultaneously, the method of this invention can automatically identify and extract key anatomical landmarks, such as the center of the renal vein and the outer contour of the kidney, ensuring that each measurement is based on accurate and consistent landmarks, thus improving measurement accuracy and consistency.

[0083] (7) Traditional methods rely on the doctor's experience to judge the relationship between the kidney and the surrounding anatomical structures, which is prone to deviation, especially when there are large anatomical variations in the kidney, the choice of surgical plan may be affected;

[0084] This invention generates a three-dimensional model that more intuitively and accurately reflects the spatial relationship between the kidney and surrounding structures, providing doctors with objective data support. In clinical applications, doctors can select the most suitable surgical approach based on the three-dimensional model and predict the risks and difficulties of the surgery. This method optimizes surgical plans through precise kidney modeling, avoids experience-based errors in traditional methods, and improves surgical safety and success rates.

[0085] (8) The three-dimensional modeling technology of this invention is not only applicable to the kidney, but can also be extended to other organs (such as the liver, pancreas, etc.) to evaluate and plan surgical procedures through three-dimensional modeling of CT data. This method can comprehensively evaluate multiple organs and provide cross-organ diagnostic and treatment decision support. In addition, the three-dimensional model can also be used as input for deep learning models to further improve the accuracy of medical image analysis and promote the development of intelligent medicine;

[0086] CT data, as high-resolution medical imaging data, provides accurate anatomical information for 3D modeling. It can be combined with deep learning algorithms to further improve the detection capability of complex lesions and provide data support for personalized treatment. Attached Figure Description

[0087] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0088] Figure 1 This is a flowchart of the MAP scoring method based on three-dimensional kidney modeling in this invention. Detailed Implementation

[0089] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0090] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0091] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the accompanying drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0092] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0093] like Figure 1 As shown in the figure, this embodiment discloses a MAP scoring method and device based on three-dimensional kidney modeling.

[0094] The first aspect of this embodiment discloses a MAP scoring method based on three-dimensional kidney modeling, such as... Figure 1 As shown, the MAP scoring method includes steps S100 to S700.

[0095] Step S100. Obtain renal medical imaging data.

[0096] The renal medical imaging data includes renal CT image data, which includes continuous cross-sectional slice data obtained through CT scans. The renal CT image data is used to provide detailed imaging information of the kidney and its surrounding anatomical structures.

[0097] The renal medical imaging data is used for subsequent 3D modeling of the kidney.

[0098] Step S200. Preprocess the renal medical imaging data.

[0099] Preprocessing of renal medical imaging data includes standardization, resampling, and denoising.

[0100] In this embodiment, the renal medical imaging data is preprocessed to ensure consistent image quality and suitability for model input.

[0101] In some embodiments of this example, the renal medical imaging data is standardized, including converting the renal CT image data into the RAS (Right-Anterior-Superior) coordinate system.

[0102] In this embodiment, the kidney CT image data is uniformly converted into the RAS coordinate system to avoid the situation where different coordinate systems (such as LPS, RAS, etc.) may affect the consistency of kidney CT image data, thus ensuring the consistency of spatial calculation.

[0103] In some embodiments of this example, resampling of renal medical imaging data includes: interpolating and resampling renal CT image data to generate isotropic data.

[0104] The slice thickness of the original CT image (e.g., 5 mm) is usually greater than the intra-slice resolution (e.g., 0.7 × 0.7 × 5 mm), which may lead to spatial inconsistencies during 3D reconstruction. This embodiment interpolates and resamples the CT image to generate isotropic data (e.g., 1 × 1 × 1 mm). By uniformly adjusting the pixel spacing in three directions, it ensures consistent resolution in each direction in 3D space, thereby reducing errors caused by resolution inconsistencies and significantly improving the quality of 3D reconstruction.

[0105] The isotropic data refers to medical image data with consistent resolution in all directions in three-dimensional space.

[0106] In some embodiments of this example, preprocessing of renal medical imaging data further includes: detecting whether there are image quality problems (such as motion artifacts, metal artifacts, etc.) in the renal CT image data, and issuing warnings or refusing to process the renal CT image data with image quality problems.

[0107] Step S300. Generate a complete three-dimensional model of the kidney based on renal medical imaging data.

[0108] In some embodiments of this example, a complete three-dimensional model of the kidney is generated based on medical imaging data, including steps S310 to S330.

[0109] Step S310. Use a deep learning segmentation model based on 3D-U-Net to segment renal cortex and renal medulla from renal medical imaging data. Based on the renal cortex and renal medulla data, use the Marching Cubes algorithm to generate a three-dimensional model of renal parenchyma. When generating the three-dimensional model of renal parenchyma using the Marching Cubes algorithm, isosurfaces are extracted by looking up tables. When generating the three-dimensional model of renal parenchyma using the Marching Cubes algorithm, more voxels are added in areas with greater curvature. When generating the three-dimensional model of renal parenchyma using the Marching Cubes algorithm, triangular patches that meet preset conditions are removed by mesh simplification.

[0110] In some embodiments of this example, a deep learning segmentation model based on 3D-U-Net is used to segment renal cortex and renal medulla from renal medical imaging data, including:

[0111] Kidney medical imaging data is input into a pre-built and trained deep learning segmentation model (such as the 3D-U-Net model). This deep learning segmentation model learns the spatial features of the renal cortex and medulla through an encoder-decoder structure to perform accurate segmentation. The deep learning segmentation model outputs binary labeled images of the renal cortex and medulla, in which the renal cortex and medulla regions are labeled separately. Morphological processing (such as closing and opening operations) is performed on the segmentation results of the renal cortex and medulla to fill small holes, remove noise, and ensure the coherence of the segmented regions.

[0112] This embodiment generates a three-dimensional model of the renal parenchyma using an improved Marching Cubes algorithm. This algorithm can automatically and accurately obtain renal cortex and medullary data from renal medical imaging data (such as renal CT image data), improving the efficiency and accuracy of renal parenchyma three-dimensional model reconstruction, thereby providing a precise foundation for subsequent three-dimensional modeling of the renal parenchyma.

[0113] The Marching Cubes algorithm constructs 3D surfaces by extracting isosurfaces from each voxel. In the original Marching Cubes algorithm, each voxel needs to find multiple critical points and determine their shape to determine the representation of the isosurface within the voxel. This embodiment introduces lookup tables to accelerate this process. By pre-calculating and storing each possible voxel configuration (i.e., different combinations of critical values) and the corresponding surface topology, the lookup tables can quickly access the pre-calculated results each time isosurfaces are extracted, thereby reducing the computational steps required for each judgment in the algorithm and improving the speed of isosurface extraction, especially when processing large-scale data.

[0114] This embodiment improves the Marching Cubes algorithm with adaptive subdivision. In areas of the kidney 3D model with high curvature or rich detail (such as the kidney edge and near the renal vein), an adaptive subdivision strategy automatically increases the mesh resolution, ensuring that the surface of these areas can be reconstructed more finely. This allows the 3D model to achieve higher accuracy in detailed areas (such as the kidney surface and the junction of the renal vein), avoiding mesh coarseness or surface distortion in these areas. For example, increasing the number of voxels in areas with high curvature achieves the increased resolution.

[0115] The reconstructed mesh in 3D typically contains a large number of triangular facets, which can lead to a heavy rendering burden. This embodiment uses mesh simplification techniques (such as edge folding) to reduce the number of unnecessary triangular facets while maintaining the surface topology and shape of the kidney model. This significantly reduces the file size of the generated model and the amount of rendering computation, improving the rendering speed and processing efficiency, especially in real-time visualization or large-scale dataset processing. For example, the impact of each triangular facet on the overall structure is calculated, and facets with an impact on the shape less than a threshold are removed, i.e., facets with little impact on the shape are removed.

[0116] In some embodiments of this example, the surface of the three-dimensional model of the kidney parenchyma is smoothed by combining Laplacian smoothing, volume-invariant constrained smoothing, and feature-preserving anisotropic smoothing.

[0117] In this embodiment, Laplacian smoothing is applied to reduce step artifacts. Specifically, applying Laplacian smoothing to reduce step artifacts includes: for each vertex of the kidney 3D model, calculating the average position of its neighboring vertices, and moving each vertex to that average position. This method can remove step artifacts caused by mesh coarseness (i.e., obvious segmentation due to mesh roughness), making the surface smoother and more natural. It only affects the position of the vertices and does not change the topology of the mesh, making it suitable for smoothing low-resolution meshes, especially after initial reconstruction.

[0118] In this embodiment, a volume-preserving smoothing method is used to ensure that the model's volume remains constant while smoothing the surface. Specifically, volume-preserving smoothing includes: adding volume constraints with each vertex position update to ensure that adjustments to the vertices do not cause changes in the overall volume; and calculating the local volume of each vertex and adjusting the vertex positions to prevent significant changes in the overall volume. This embodiment uses the Laplacian operator to smooth the surface, while volume-preserving constraints ensure that the overall shape and volume of the kidney parenchyma are preserved after each smoothing process. This method ensures that the smoothed 3D model maintains the overall shape and volume consistency of the kidney while removing surface noise, effectively avoiding anatomical errors caused by shape changes during smoothing. It is particularly suitable for medical applications that require accurate preservation of organ volume.

[0119] In this embodiment, feature-preserving anisotropic smoothing is used to adjust the smoothing operation to retain boundaries and important features during the smoothing process. Specifically, feature-preserving anisotropic smoothing includes: using local gradient or curvature information to guide the smoothing process, ensuring that the smoothing operation has less impact on detailed areas (such as the edges of the kidney, renal veins, etc.) and performs more smoothing in flat areas; smoothing is performed using anisotropic diffusion equations, taking into account changes in surface normal vectors during the smoothing process to maintain sharp edge features; and adjusting the smoothing intensity so that areas with greater curvature (such as the outer contour of the kidney and near the renal veins) retain less smoothing, while flat areas (such as the relatively smooth areas inside the kidney) allow for more smoothing. Feature-preserving anisotropic smoothing is directional during the smoothing process, avoiding the smoothing effect on detailed parts, preserving the features of the kidney model (such as surface contours, renal vein areas, etc., avoiding edge blurring caused by over-smoothing), and improving the detail preservation ability of the 3D model, especially when dealing with complex anatomical structures, effectively avoiding the loss of details.

[0120] Step S320. Use a deep learning segmentation model (such as the 3D-U-Net model) to segment the renal vein region data from the renal medical image data, and use a three-dimensional reconstruction algorithm (such as the Marching Cubes algorithm) to generate a three-dimensional model of the renal vein and determine the spatial course of the renal vein.

[0121] The spatial course of the renal vein refers to the "path" or "direction" of the renal vein in three-dimensional space. Specifically, spatial course describes the path of the renal vein from the kidney to the inferior vena cava, including its direction, shape, and spatial relationship with other anatomical structures. "Spatial course" can be understood as a description of the path of the renal vein in three-dimensional space, used to describe its spatial layout in the kidney and surrounding structures.

[0122] In some embodiments of this example, a deep learning segmentation model is used to segment renal vein region data from renal medical imaging data, including: inputting renal medical imaging data into a pre-constructed and trained deep learning segmentation model; the deep learning segmentation model classifies the renal tissue pixel by pixel to segment renal vein region data.

[0123] In this embodiment, the deep learning segmentation model classifies kidney tissue pixel by pixel, accurately segmenting the renal cortex, medulla, and renal vein regions. This deep learning segmentation model extracts and segments the renal vein structure by optimizing the convolutional and decoding layers. The output of the deep learning segmentation model is a 3D label map, marking each voxel as belonging to the renal vein, renal cortex, or renal medulla.

[0124] In this embodiment, after obtaining the segmentation results of the renal vein, the spatial topological information of the renal vein is extracted to reconstruct its three-dimensional structure, ensuring accurate representation of the course and morphology of the renal vein. Specifically, renal vein region data is extracted from the segmentation results output by the deep learning segmentation model. This renal vein region data has been labeled as renal vein in the segmentation label map. Using the extracted renal vein region data, a three-dimensional mesh model of the renal vein is generated through a three-dimensional reconstruction algorithm. The three-dimensional model of the renal vein represents its morphology and spatial structure.

[0125] In some embodiments of this example, surface smoothing algorithms (such as Laplacian smoothing or feature-preserving smoothing) are applied to remove noise from the three-dimensional model of the renal vein, thereby improving the smoothness and accuracy of the surface of the three-dimensional model of the renal vein and thus improving the quality of the three-dimensional model of the renal vein.

[0126] In some embodiments of this example, determining the spatial course of the renal vein includes: extracting the central path of the renal vein from a three-dimensional model of the renal vein using a vascular centerline extraction algorithm (such as a curvature-based centerline extraction method); determining the path of the renal vein from the kidney to the inferior vena cava, and the spatial orientation of the renal vein within the kidney, through spatial analysis (such as a shortest path algorithm); and analyzing the course of the renal vein in different regions by combining the three-dimensional centerline and anatomical structure of the renal vein, including the direction, curvature, and relationship with surrounding tissues (such as the renal cortex, renal medulla, and renal capsule).

[0127] The central pathway of the renal vein can accurately represent the spatial course of the renal vein.

[0128] The spatial course of the renal vein refers to its path in three-dimensional space. In this embodiment, the course of the renal vein is analyzed by extracting the centerline or by using a path tracing algorithm based on a three-dimensional model of the renal vein.

[0129] In this embodiment, by constructing a three-dimensional model of the renal vein and determining its spatial course, important anatomical information is provided for kidney surgery planning, disease diagnosis, and treatment plans. Especially in complex kidney diseases or surgeries, it can provide doctors with accurate renal vein location and path analysis, which helps to improve the success rate and accuracy of the surgery.

[0130] Step S330. Use a deep learning segmentation model (such as the 3D-U-Net model) to extract the three-dimensional segmentation data of the psoas major muscle from the kidney medical imaging data, and determine the three-dimensional surface of the posterior abdominal wall based on the three-dimensional segmentation data of the psoas major muscle.

[0131] The psoas muscle is located deep in the abdomen and exhibits significant density differences in CT images, making it easy to segment. Accurate segmentation of the psoas muscle can provide essential references for three-dimensional modeling of the posterior abdominal wall.

[0132] The method utilizes a deep learning segmentation model to extract three-dimensional segmentation data of the psoas major muscle from renal medical imaging data. This includes: inputting renal medical imaging data into a trained deep learning segmentation model, and having the deep learning segmentation model output three-dimensional segmentation data of the psoas major muscle, thereby marking the accurate location of the psoas major muscle in three-dimensional space.

[0133] The posterior abdominal wall is mainly composed of the psoas major muscle, spine, ribs, and other deep muscles. By extracting the three-dimensional segmentation data of the psoas major muscle and combining it with anatomical relationships, the three-dimensional surface of the posterior abdominal wall can be determined. The three-dimensional surface of the posterior abdominal wall is usually directly related to the anterior edge of the psoas major muscle (i.e., the part close to the abdominal cavity). By analyzing the position of the anterior edge of the psoas major muscle in CT images, the location of the posterior abdominal wall can be determined.

[0134] The method for determining the three-dimensional surface of the posterior abdominal wall based on the three-dimensional segmentation data of the psoas major muscle includes: inferring the geometry of the posterior abdominal wall based on the three-dimensional segmentation results of the psoas major muscle; and generating the three-dimensional surface of the posterior abdominal wall, which represents the abdominal wall region around the kidneys, using the Marching Cubes algorithm or other mesh generation algorithm through the extracted boundary data of the psoas major muscle.

[0135] Generally, the posterior abdominal wall is located behind the psoas major muscle, particularly in the rib and spinal region. The rough geometry of the posterior abdominal wall is formed by the anterior border of the psoas major muscle and its positional relationship with surrounding structures such as the spine and ribs.

[0136] In some embodiments of this example, the determination of the three-dimensional surface of the posterior abdominal wall based on the three-dimensional segmentation data of the psoas major muscle further includes: smoothing and refining the posterior abdominal wall surface by fusing Laplacian smoothing and feature-preserving anisotropic smoothing.

[0137] The generated posterior abdominal wall surface may contain some irregularities or small artifacts, thus requiring smoothing and refinement to ensure the accuracy and smoothness of the final surface. Applying the Laplacian smoothing method to the generated 3D posterior abdominal wall surface can reduce step-like artifacts caused by mesh coarseness and make the surface smoother. Applying feature-preserving anisotropic smoothing, adjusting the smoothing degree according to the characteristics of the region (such as curvature and boundaries) during the smoothing process, can ensure that details are not over-smoothed and avoid affecting key anatomical structures (such as the spinal margins and rib areas).

[0138] In some embodiments of this example, the determination of the three-dimensional surface of the posterior abdominal wall based on the three-dimensional segmentation data of the psoas major muscle also includes: for important structural regions (such as the junction of the spine and the posterior abdominal wall), the surface is optimized by a detail-preserving algorithm to ensure that important anatomical features are not lost; after the surface is smoothed, there may be too many triangular facets, so the mesh needs to be optimized to reduce the rendering burden and keep the topology unchanged.

[0139] This embodiment further optimizes the generated posterior abdominal wall surface to ensure its accuracy and practicality in anatomical structures.

[0140] Step S400. Calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and use the Z coordinate of the centroid as the standard measurement surface.

[0141] This embodiment analyzes the three-dimensional segmentation results of the renal vein and extracts the coordinates of all voxels belonging to the renal vein region (i.e., voxels whose segmentation results are renal veins). Specifically, the three-dimensional segmentation results of the renal vein are analyzed, and the positions of all voxels in the renal vein region are extracted. The position of each voxel is represented by its coordinates (X, Y, Z) in three-dimensional space. The three-dimensional image data of the renal vein three-dimensional model is traversed, and the coordinates of the voxels in the renal vein region with a value of 1 are extracted to generate a list of voxel coordinates.

[0142] The centroid is a weighted average of the positions of all voxels and can represent the center position of the segmented region. Specifically, the coordinates of all voxels in the renal vein region are calculated, and the weighted average position of these coordinates is obtained. The weighted average is based on the position of each voxel and is achieved by summing the coordinates of all voxels and dividing by the total number of voxels. The centroid position is calculated by obtaining three coordinates: X, Y, and Z coordinates. The Z coordinate is the centroid Z position of the three-dimensional model of the renal vein.

[0143] In this embodiment, the Z-coordinate of the centroid is used as the standard measurement plane. This plane is located at the Z-position of the centroid of the renal vein 3D model and is parallel to the XY plane. This plane can serve as a reference for subsequent measurements and analyses to help determine the position and relationship of other anatomical regions. Using the Z-coordinate of the centroid as the standard measurement plane ensures that the measurements are consistent with the actual position and shape of the renal vein, avoiding deviations from the anatomical orientation of the human body.

[0144] These embodiments automatically determine the standard measurement plane by calculating the three-dimensional centroid of the renal vein. Compared with the traditional method that relies on the doctor's visual judgment to select the measurement plane, the method of this embodiment considers the overall morphology of the renal vein rather than local features. The process of determining the standard measurement plane is completely objective and unaffected by the operator, and can achieve 100% repeatability, with measurement results at different times being completely consistent.

[0145] Step S500. Identify the anatomical regions to be measured on the standard measurement plane based on a deep learning segmentation model.

[0146] In some implementations of this embodiment, the anatomical region to be measured is identified on a standard measurement plane based on a deep learning segmentation model, including steps S510 to S570.

[0147] Step S510. Extract two-dimensional slice images of the standard measurement plane from the renal medical imaging data.

[0148] The two-dimensional slice image contains cross-sectional information of kidney-related anatomical structures (such as renal veins, renal cortex, renal medulla, etc.).

[0149] Step S520. Preprocess the extracted two-dimensional slice image.

[0150] For example, the extracted two-dimensional slice images can be processed by denoising, standardizing, and enhancing contrast so that subsequent deep learning models can accurately distinguish anatomical structures.

[0151] Step S530. Input the two-dimensional slice image into a deep learning segmentation model (such as the 3D-U-Net model). The deep learning segmentation model automatically labels and segments different anatomical regions in the image and generates segmentation results containing labels for each anatomical region (such as renal vein, renal cortex, renal medulla, etc.).

[0152] Step S540. Perform morphological processing on the segmentation results output by the deep learning segmentation model.

[0153] For example, the segmentation results output by the deep learning segmentation model can be processed by dilation, erosion, opening and closing operations to further optimize the segmentation effect, remove artifacts, fill holes, and ensure the integrity and coherence of the anatomical region.

[0154] Step S550. Smooth the segmentation results output by the deep learning segmentation model using Laplacian smoothing or feature-preserving anisotropic smoothing algorithms.

[0155] In this embodiment, the boundaries of the anatomical region are further optimized by Laplacian smoothing or anisotropic smoothing algorithm with feature preservation, removing possible rough or irregular regions and ensuring the smoothness of the segmentation result.

[0156] Step S560. Extract the anatomical regions to be measured (such as renal vein, renal cortex, renal medulla, etc.) from the segmentation results.

[0157] Each anatomical region is assigned a unique label to facilitate subsequent measurements and analysis.

[0158] Step S570. Perform geometric analysis on the extracted anatomical region and automatically calculate the relevant anatomical parameters.

[0159] For example, anatomical parameters include the diameter of the renal vein, the thickness of the renal cortex, and the volume of the renal medulla.

[0160] This embodiment uses deep learning technology to achieve automatic identification and quantification of anatomical regions, providing efficient and accurate data support for medical analysis, diagnosis, and surgical planning.

[0161] This embodiment uses precise anatomical landmarks for automatic positioning, which improves measurement accuracy.

[0162] Step S600. Extract the outer contour of the kidney based on the three-dimensional model of the kidney, determine the contour of the posterior abdominal wall based on the anterior edge of the psoas major muscle, and calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall.

[0163] The shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall is the P-value, which is an important parameter describing the spatial relationship between the kidney and the posterior abdominal wall.

[0164] In some embodiments of this example, the outer contour of the kidney is extracted based on a three-dimensional kidney model, including: extracting the outer contour of the kidney using morphological algorithms or edge detection methods based on the three-dimensional kidney model.

[0165] In some embodiments of this example, the contour of the posterior abdominal wall is determined based on the anterior border of the psoas major muscle, including: inferring the three-dimensional contour of the posterior abdominal wall based on anatomical landmarks such as the anterior border of the psoas major muscle and the spine.

[0166] In some embodiments of this example, distance calculation algorithms (such as KD-trees or nearest neighbor lookup methods) are used to calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall. A KD-tree (K-Dimensional Tree) is a spatial index structure for fast lookups in multidimensional space, suitable for nearest neighbor lookups in this scenario. This embodiment uses a KD-tree to accelerate the calculation of the shortest distance between the outer contour points of the kidney and the contour points of the posterior abdominal wall.

[0167] The shortest distance between the outer contour of the kidney and the posterior abdominal wall is calculated using a distance calculation algorithm, including: constructing a KD tree structure of the posterior abdominal wall contour points; for each renal capsule point (a point on the outer contour of the kidney), the nearest neighbor query is used to find the nearest posterior abdominal wall point, and the minimum value among all the distances between the renal capsule points and the posterior abdominal wall points is the shortest distance between the outer contour of the kidney and the posterior abdominal wall contour.

[0168] Constructing a KD-tree structure for the posterior abdominal wall contour points includes: extracting all point sets of the posterior abdominal wall contour, where each point has three-dimensional coordinates (X, Y, Z); constructing a KD-tree data structure using the posterior abdominal wall contour point set, where each node of the KD-tree stores one point and is partitioned in space according to specific rules, enabling efficient nearest neighbor queries for any point. In this embodiment, by constructing a KD-tree, each point of the posterior abdominal wall contour can be quickly searched within a time complexity of O(log N), significantly improving query efficiency.

[0169] For each renal capsule point (a point on the outer contour of the kidney), a nearest neighbor query is used to find the nearest posterior abdominal wall point. This includes: traversing all points on the outer contour of the kidney (i.e., renal capsule points), each point having three-dimensional coordinates (X, Y, Z); for each renal capsule point, using the nearest neighbor query method of a KD tree to find the nearest posterior abdominal wall contour point; for each renal capsule point and its nearest posterior abdominal wall point, calculating the Euclidean distance between them; and calculating the shortest distance between the outer contour of the kidney and the posterior abdominal wall contour, which is the minimum value among all distances between renal capsule points and posterior abdominal wall points.

[0170] In this embodiment, the P-value can be accurately calculated, providing an important reference for kidney-related anatomical analysis and surgical planning.

[0171] This embodiment successfully reduces the computational complexity of calculating the shortest distance between the outer contour of the kidney and the posterior abdominal wall from O(n²) to O(n log n) by introducing a KD-tree spatial index structure, significantly improving computational speed, especially maintaining real-time processing capabilities when handling high-resolution contours. Simultaneously, the KD-tree structure ensures that the optimal solution is obtained for each query, avoiding redundant calculations and guaranteeing the accuracy and efficiency of the computation.

[0172] Traditional methods directly calculate the distances between all renal capsule points and retroperitoneal wall points, resulting in a computational complexity of O(n²), where n is the number of points in the outer contour of the kidney and the retroperitoneal wall. For example, if there are hundreds of points in each of the renal capsule and retroperitoneal wall contours, the total number of distance calculations for each pair of points would reach tens of thousands. This brute-force calculation method is computationally intensive, especially at high resolutions, and is extremely slow. This embodiment employs a KD-tree spatial indexing structure. A KD-tree organizes data points in three-dimensional space into a tree structure, allowing the search for the nearest neighbor of each point to occur in O(log n) time. For each renal capsule point, using a KD-tree for nearest neighbor lookup, its nearest retroperitoneal wall point can be found in O(log n) time complexity. Therefore, the computational complexity of the entire process is reduced from O(n²) to O(n log n), significantly improving computational efficiency.

[0173] Traditional methods require distance calculations for every pair of points, resulting in a quadratic increase in computational load. This embodiment, however, utilizes a KD-tree index structure to significantly accelerate both querying and distance calculation. When performing nearest neighbor queries, the KD-tree effectively skips irrelevant points, reducing unnecessary computations. Specifically, the KD-tree can find the optimal nearest neighbor in a shorter computation time, thus improving computation speed by more than 10 times compared to traditional methods. By dividing the data into multiple spatial regions, the KD-tree allows for distance calculations to focus only on the nearest regions, greatly reducing the number of points that need to be calculated and improving computational speed. This is especially true for high-resolution contours, where the calculation of hundreds of point pairs is effectively reduced, enabling real-time processing.

[0174] With the development of medical imaging technology, the resolution of generated 3D kidney models has gradually increased, resulting in each contour potentially containing thousands of points. This significantly increases computational load and memory consumption, making traditional point-to-point calculation methods insufficient for real-time processing of high-resolution data. This embodiment organizes and indexes the 3D coordinates of the kidney's outer contour and retroperitoneal wall contour points using a KD-tree, enabling rapid lookup of the shortest distance between each renal capsule point and the retroperitoneal wall point. This ensures high computational speed and real-time processing capabilities even when handling high-resolution contours.

[0175] The structure of a KD-tree ensures that the optimal solution is found in each nearest neighbor query. The data index built using a KD-tree can quickly filter the points closest to the target point in space, guaranteeing that the nearest neighbor obtained in each query is the globally optimal solution, without missing any closer points. Through spatial partitioning and optimized tree structure, the KD-tree avoids redundant distance calculations, thus ensuring that the shortest distance between the renal capsule point and the posterior abdominal wall point is found efficiently without considering all possible combinations of point pairs. In this way, the entire process ensures computational accuracy and consistently yields the globally optimal solution.

[0176] Step S700. Determine the perirenal fat region, enhance the cord-like structures using a multi-scale filter, and calculate the proportion of the cord-like structures in the perirenal fat region.

[0177] In some embodiments of this example, the perirenal fat region is determined, the cord-like structure is enhanced using a multi-scale filter, and the proportion of the cord-like structure in the perirenal fat region is calculated, including:

[0178] Step S710. Define the area within a preset range surrounding the kidney as the perirenal fat region.

[0179] For example, the area within 10 cm of the kidney is defined as the perirenal fat region.

[0180] Step S720. Design a multi-scale Frangi filter bank with a scale range of 1.0-3.0 mm.

[0181] Frangi filters are used to enhance tubular structures in images. By designing multi-scale Frangi filters, it is possible to detect strip-like structures of different sizes at multiple scales.

[0182] Based on the resolution and characteristics of the cord-like structures in kidney CT images, this embodiment designs a multi-scale Frangi filter bank with a scale range of 1.0-3.0 mm. Each scale can enhance cord-like structures of different sizes and adapt to different details.

[0183] The Frangi filter calculates the response of tubular structures in an image based on local curvature and gradient. The filter design (such as the filter parameter design) ensures that it can adapt to strip-like structures of different thicknesses.

[0184] Step S730. Apply a multi-scale Frangi filter to the kidney medical image data, perform image processing at each scale, and generate a corresponding response map for each scale to enhance the regions in the image that conform to tubular features.

[0185] At different scales, the Frangi filter can extract tubular structures of different sizes. By detecting tubular structures at multiple scales, the saliency of filamentous structures can be enhanced, and both fine and coarse filaments can be effectively extracted.

[0186] The response map at each scale highlights the tubular structures in the image, reduces background noise, and improves the visibility of the cord-like structures.

[0187] Step S740. Fuse the response results of the Frangi filter at different scales.

[0188] The detected cord-like structures at each scale may have local response differences. By fusing response results from different scales, more accurate detection of cord-like structures can be obtained.

[0189] The Frangi filter's response is the enhanced image. After processing the input CT image, the Frangi filter outputs an enhanced image of the vegetal structure. In this enhanced image, the value of each pixel represents the probability score (vesselness measure) that the location belongs to a vegetal structure. In this embodiment, the Frangi filter's response results at different scales are fused to obtain multiple enhanced images obtained at different scales. By fusing, a final vegetal enhanced image integrating information from all scales is obtained.

[0190] In this embodiment, the response results of the Frangi filter are fused using a weighted average or maximum response method to integrate the response results at multiple scales.

[0191] This embodiment weights the results based on the response intensity at each scale, ensuring that both detailed areas (e.g., small stripes) and coarse stripes are well enhanced.

[0192] Step S750. Differentiate between strands and blood vessels based on CT values ​​and morphological features.

[0193] Linear structures and blood vessels typically appear differently on CT images. Linear structures appear as higher density (brighter areas), while blood vessels usually appear as lower density (darker areas). Linear structures and blood vessels can be effectively distinguished using CT values ​​and morphological features.

[0194] Based on the grayscale values ​​of CT images, linear structures typically have higher density (higher brightness), while blood vessels appear as lower density. The two can be distinguished by setting an appropriate density threshold.

[0195] Cord-like structures are typically long and narrow, while blood vessels are uniform tubular structures. The morphological characteristics of these structures (such as aspect ratio and curvature) can be used to further distinguish between cords and blood vessels. For example, cord-like structures are usually elongated with a large aspect ratio, while blood vessels are more uniformly round or elliptical; cord-like structures have a higher curvature, while the curvature of blood vessels changes relatively gently.

[0196] Step S760. Output the strip-like structure region and its location.

[0197] Through the above steps, the cord-like structures surrounding the kidneys are finally extracted, and these structures are separated from blood vessels. The output includes the enhanced cord-like structure regions and their locations.

[0198] Step S770. Calculate the volume of the cord-like structure and the volume of the perirenal fat region.

[0199] The volume of the cord-like structure is calculated as follows: after the cord-like structure is extracted, its volume is calculated using the spatial coordinates of each voxel in the 3D image. Cord-like structures appear as high-density areas in CT images; by detecting the 3D coordinates of these areas, their volume in 3D space can be quantified. The volume is calculated by determining the space occupied by these voxels.

[0200] The method for calculating the volume of the perirenal fat region is as follows: The volume of the perirenal fat region is calculated based on its voxel data in three-dimensional space. By spatially quantizing the extracted fat region, the volume occupied by these regions is calculated.

[0201] Step S780. Calculate the proportion of the cord-like structure to the perirenal fat region based on the volume of the cord-like structure and the volume of the perirenal fat region.

[0202] This embodiment utilizes a multi-scale Frangi filter and comprehensive feature analysis to automatically and accurately enhance and extract the cord-like structures surrounding the kidneys, without any human intervention, providing precise data support for subsequent medical analysis.

[0203] The cord-like structures in the perirenal fat vary in thickness, and single-scale detection methods are prone to missing small cord-like structures or misidentifying large blood vessels as cords. This embodiment employs a multi-scale three-dimensional cord detection method. By performing image processing at multiple scales, it can effectively enhance the detection capability of cords of different thicknesses, significantly improve detection sensitivity, maintain high specificity, and adapt to the anatomical variations of different patients.

[0204] This embodiment uses multiple scales (1.0-3.0 mm), each capable of detecting cord-like structures of different sizes. Smaller cords are enhanced at smaller scales, while thicker cords are effectively detected at larger scales. The response at each scale helps identify cord-like structures of different sizes around the kidney, thereby reducing the probability of missing small cords. This multi-scale approach allows for the effective identification of structures of different sizes in the image, significantly improving the detection sensitivity of small cords in perirenal fat. Compared to single-scale detection, the detection sensitivity is improved by 30%, enabling the identification of more cord-like structures, especially small and thin cords.

[0205] High specificity requires minimizing false alarms during the detection process, especially avoiding misidentifying blood vessels around the kidneys as cord-like structures. This embodiment combines CT values ​​and morphological features with multi-scale detection to effectively distinguish between cord-like structures and blood vessels, ensuring high specificity.

[0206] In kidney CT images, cord-like structures and blood vessels have significant differences in CT values. Cord-like structures usually appear as higher density (brighter areas), while blood vessels usually appear as lower density (darker areas). By setting an appropriate CT threshold, cord-like structures and blood vessels can be effectively separated, reducing misjudgment.

[0207] Cord-like structures typically present as slender shapes, while blood vessels are usually more round or elliptical. By analyzing the morphological characteristics of cords, such as their aspect ratio and curvature, we can further distinguish between cords and blood vessels, ensuring that only cord-like structures are detected.

[0208] This embodiment combines CT values ​​with morphological features to form a comprehensive detection model, which excludes blood vessels while ensuring the presence of cord-like structures, thereby guaranteeing high specificity, with a specificity rate of over 95%.

[0209] The renal anatomy may vary from patient to patient, especially the cord-like structures in the perirenal fat region, whose thickness and distribution may differ due to individual variations. To ensure broad applicability of the detection method, this embodiment employs a multi-scale, three-dimensional cord detection method, which can accommodate anatomical variations in different patients.

[0210] This embodiment utilizes multi-scale detection, enabling the algorithm to automatically adapt to variations in the kidneys and perirenal fat regions among different patients. For example, some patients may have finer cord-like structures around the kidneys, while others may have thicker cords. The multi-scale approach ensures effective identification of these cords across these diverse anatomical structures.

[0211] The 3D U-Net model in this embodiment has been trained with a large amount of data on different anatomical structures and has strong generalization ability. The model can automatically adapt to the anatomical differences of different patients, reduce errors caused by individual differences, and ensure that the recognition of various anatomical structures remains efficient and accurate.

[0212] In this embodiment, the severity of cord-like structures is objectively assessed by calculating the percentage of their volume relative to the perirenal fat region. This quantifies the severity of the cords, shifting from subjective judgment to precise measurement. It overcomes the shortcomings of existing technologies that rely on subjective assessment of cord severity from two-dimensional images and lack quantitative standards, providing more accurate and objective measurement results. This has significant clinical application value for surgical planning and disease diagnosis. The percentage of cord-like structures relative to the perirenal fat region indicates the spatial proportion occupied by cord-like structures within the perirenal fat region. A higher percentage indicates more cord-like structures around the kidney, potentially suggesting a higher degree of renal adhesion.

[0213] This embodiment fundamentally improves the MAP (Mean Angiogenesis Assessment) scoring method, providing a quantitative standard for MAP scoring. By automatically calculating the volume percentage of MAP, it achieves objective scoring of MAP, transforming it from "empirical assessment" to "data-driven quantitative analysis." While maintaining clinical applicability, it completely solves the problem of scoring consistency. For example, the grading standard for the severity of MAP is as follows: <5%, indicating very few cord-like structures around the kidney, which can usually be considered as no obvious MAP, with minimal impact on the kidney, score 0; 5-25%, indicating mild to moderate cord-like structures, which may have some impact on kidney health, but the impact is relatively minor, score 2; >25%, indicating a large number of cord-like structures around the kidney, indicating strong adhesion between the kidney and surrounding tissues (such as the retroperitoneal wall and renal capsule), which usually leads to higher surgical risks, score 3.

[0214] This embodiment achieves sub-voxel precision boundary positioning through trilinear interpolation, improving measurement accuracy to the 0.1mm level; supplementary measurements are performed on two layers above and below the measurement layer; if the difference in results exceeds 10%, the measurement range is expanded and the median is taken; a normal P-value range (5-50mm) is established, and results outside the range are verified a second time or marked as abnormal.

[0215] The extraction of the kidney's outer contour and the retroperitoneal wall contour is often limited by the resolution of CT images, leading to certain accuracy errors in the segmentation boundaries. This embodiment employs trilinear interpolation to refine these boundaries. Trilinear interpolation, by weighting the voxel data around each point on the kidney's outer contour and the retroperitoneal wall contour, can obtain more accurate boundary positions. It interpolates in three directions (X, Y, and Z axes) based on the values ​​of adjacent voxels, optimizing boundary localization and achieving sub-voxel level accuracy (i.e., 0.1 mm). After trilinear interpolation, the boundary positions of the original kidney's outer contour and the retroperitoneal wall contour are more accurately determined, avoiding errors caused by voxel resolution limitations. This refined boundary localization method results in higher accuracy for subsequent shortest distance measurements, improving the overall measurement accuracy to the 0.1 mm level. This ensures high accuracy in measuring the distance (P-value) between the kidney and the retroperitoneal wall, providing a high-precision foundation for subsequent quantitative analysis.

[0216] In this embodiment, after calculating the shortest distance on the standard measurement plane, supplementary measurements are performed on the two layers above and below the measurement plane. That is, not only is the distance calculated on the current standard measurement plane, but the shortest distance is also measured on the two layers above and below the plane (i.e., the two adjacent layers in the Z-axis direction). This multi-layer comparison is used to detect and avoid deviations caused by image noise or measurement errors, ensuring the stability and reliability of the measurement results.

[0217] The shortest distance measurement is already based on the distance between the outer contour of the kidney and the contour of the posterior abdominal wall. However, in order to reduce errors caused by voxel resolution or other influencing factors, supplementary measurements can further ensure the stability of the shortest distance and avoid errors that may occur due to single-plane measurements.

[0218] If the difference between the supplementary measurements from the upper and lower layers exceeds 10%, it indicates that the measurement may have been affected by noise or other factors. In this case, expanding the measurement range—that is, re-measuring and taking the median of these measurements as the final value—can reduce the impact of outliers and improve the accuracy of the final result.

[0219] Based on extensive clinical data, the P-value (the shortest distance between the outer contour of the kidney and the contour of the retroperitoneal wall) typically falls within the range of 5-50 mm. Therefore, this embodiment establishes a normal P-value range. Results exceeding this range will undergo secondary verification or be marked as abnormal. This method helps filter out outliers during the measurement process, ensuring the reliability of the final measurement results.

[0220] In some embodiments of this example, the MAP scoring method further includes generating a MAP scoring report based on the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall, and the proportion of cord-like structures in the perirenal fat region.

[0221] The methods implemented in these embodiments can automatically generate standardized MAP scoring reports, effectively improving work efficiency.

[0222] The second aspect of this embodiment discloses a MAP scoring device based on three-dimensional kidney modeling. The MAP scoring device includes a data acquisition module, a data preprocessing module, a three-dimensional modeling module, a centroid calculation module, an anatomical region identification module, a shortest distance calculation module, and a strip analysis module.

[0223] The data acquisition module is used to acquire renal medical imaging data, which includes renal CT image data.

[0224] The data preprocessing module is used to preprocess renal medical imaging data.

[0225] The 3D modeling module is used to generate a complete 3D model of the kidney based on renal medical imaging data.

[0226] The centroid calculation module is used to calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and uses the Z coordinate of the centroid as the standard measurement surface.

[0227] The anatomical region identification module is used to identify the anatomical regions that need to be measured on a standard measurement plane based on a deep learning segmentation model.

[0228] The shortest distance calculation module is used to extract the outer contour of the kidney based on the three-dimensional model of the kidney, determine the contour of the posterior abdominal wall based on the anterior edge of the psoas major muscle, and calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall.

[0229] The cord analysis module is used to identify the perirenal fat region, enhance the cord-like structures using a multi-scale filter, and calculate the proportion of cord-like structures in the perirenal fat region.

[0230] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0231] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A MAP scoring method based on three-dimensional kidney modeling, characterized in that, include: Acquire renal medical imaging data, including renal CT image data; Preprocessing of renal medical imaging data; Generate a complete 3D model of the kidney based on renal medical imaging data; Calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and use the Z coordinate of the centroid as the standard measurement surface; Identify the anatomical regions to be measured on a standard measurement plane based on a deep learning segmentation model; The outer contour of the kidney is extracted based on the three-dimensional model of the kidney, the contour of the posterior abdominal wall is determined based on the anterior edge of the psoas major muscle, and the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall is calculated. The perirenal fat region was identified, and the cord-like structures were enhanced using a multi-scale filter. The proportion of the cord-like structures in the perirenal fat region was calculated.

2. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, Preprocessing of renal medical imaging data includes: Convert kidney CT image data to the RAS coordinate system. The kidney CT image data was interpolated and resampled to generate isotropic data.

3. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, A complete 3D model of the kidney is generated based on renal medical imaging data, including: A deep learning segmentation model was used to segment renal cortex and renal medulla from renal medical imaging data. Based on the renal cortex and renal medulla data, a three-dimensional model of renal parenchyma was generated using the Marching Cubes algorithm. When generating the three-dimensional model of renal parenchyma using the Marching Cubes algorithm, isosurfaces were extracted by looking up tables, more voxels were added in areas with greater curvature, and triangular patches that met preset conditions were removed by mesh simplification. The renal vein region data was segmented from renal medical imaging data using a deep learning segmentation model, and a three-dimensional model of the renal vein was generated by a three-dimensional reconstruction algorithm to determine the spatial course of the renal vein. Three-dimensional segmentation data of the psoas major muscle was extracted from renal medical imaging data using a deep learning segmentation model (such as the 3D-U-Net model), and the three-dimensional surface of the posterior abdominal wall was determined based on the three-dimensional segmentation data of the psoas major muscle.

4. The MAP scoring method based on three-dimensional kidney modeling according to claim 3, characterized in that, Generating a complete 3D model of the kidney based on renal medical imaging data also includes: The surface of a 3D model of kidney parenchyma is smoothed by combining Laplacian smoothing, volume-invariant constrained smoothing, and feature-preserving anisotropic smoothing.

5. The MAP scoring method based on three-dimensional kidney modeling according to claim 3, characterized in that, Determining the spatial course of the renal vein includes: The central path of the renal vein was extracted from the three-dimensional model of the renal vein using a blood vessel centerline extraction algorithm. Spatial analysis was used to determine the path of the renal vein from the kidney to the inferior vena cava, as well as the spatial orientation of the renal vein within the kidney. By combining the three-dimensional centerline and anatomical structure of the renal vein, the course of the renal vein in different regions was analyzed, including its direction, curvature, and relationship with surrounding tissues.

6. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, Based on a deep learning segmentation model, the anatomical regions to be measured are identified on a standard measurement plane, including: Extracting two-dimensional slice images of standard measurement planes from renal medical imaging data; Preprocess the extracted two-dimensional slice image; Two-dimensional slice images are input into a deep learning segmentation model. The deep learning segmentation model automatically labels and segments different anatomical regions in the image and generates segmentation results containing labels for each anatomical region. Morphological processing is performed on the segmentation results output by the deep learning segmentation model; The segmentation results output by the deep learning segmentation model are smoothed using Laplacian smoothing or anisotropic smoothing algorithms that preserve features. Extract the anatomical regions that need to be measured from the segmentation results; Geometric analysis was performed on the extracted anatomical regions to calculate anatomical parameters.

7. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, Calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall, including: Extract all point sets of the posterior abdominal wall contour, and construct a KD tree data structure using the posterior abdominal wall contour point set. Each node of the KD tree stores one point. Traverse all points on the outer contour of the kidney. For each kidney capsule point, use the nearest neighbor query method of KD tree to find the nearest point on the posterior abdominal wall contour. For each renal capsule point and its nearest posterior abdominal wall point, calculate the Euclidean distance between them; Calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall. This shortest distance is the minimum value among all distances between points on the renal capsule and points on the posterior abdominal wall.

8. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, The perirenal fat region was identified, and the cord-like structures were enhanced using a multi-scale filter. The proportion of cord-like structures in the perirenal fat region was calculated, including: The area within a preset range surrounding the kidney is defined as the perirenal fat region; Design a multi-scale Frangi filter bank with a scale range of 1.0-3.0 mm; A multi-scale Fragi filter was applied to renal medical imaging data, and image processing was performed at each scale. A corresponding response map was generated for each scale to enhance the regions in the image that conform to tubular features. The response results of the Frangi filter at different scales are fused; Distinguishing between strands and blood vessels based on CT values ​​and morphological characteristics; Output the strip-like structure region and its location; Calculate the volume of the cord-like structures and the volume of the perirenal fat region; The proportion of the cord-like structure to the perirenal fat region is calculated based on the volume of the cord-like structure and the volume of the perirenal fat region.

9. The MAP scoring method based on three-dimensional kidney modeling according to claim 1, characterized in that, After calculating the shortest distance on the standard measurement surface, supplementary measurements are performed on the two layers above and below the standard measurement surface. If the difference between the supplementary measurement results of the upper and lower layers exceeds 10%, the measurement range is expanded, and the median of these measurement results is taken as the final measurement value. This can reduce the impact of outlier data and improve the accuracy of the final result.

10. A MAP scoring device based on three-dimensional kidney modeling, characterized in that, include: The data acquisition module is used to acquire renal medical imaging data, including renal CT image data. The data preprocessing module is used to preprocess renal medical imaging data; The 3D modeling module is used to generate a complete 3D model of the kidney based on renal medical imaging data. The centroid calculation module is used to calculate the centroid position of all voxels in the three-dimensional model of the renal vein, and uses the Z coordinate of the centroid as the standard measurement surface. The anatomical region identification module is used to identify the anatomical regions to be measured on a standard measurement plane based on a deep learning segmentation model. The shortest distance calculation module is used to extract the outer contour of the kidney based on the three-dimensional model of the kidney, determine the contour of the posterior abdominal wall based on the anterior edge of the psoas major muscle, and calculate the shortest distance between the outer contour of the kidney and the contour of the posterior abdominal wall. The cord analysis module is used to identify the perirenal fat region, enhance cord-like structures using a multi-scale filter, and calculate the proportion of cord-like structures in the perirenal fat region.

Images