A liver fibrosis intelligent evaluation method and system based on abdominal CT images

Abstract

The application provides a liver fibrosis intelligent evaluation method and system based on an abdominal CT image, and the method comprises the following steps: extracting a liver region image from the abdominal CT image to construct 3D body data of the liver region; performing region segmentation on the body data to form a whole liver mask, a target region mask, and calculating the proportion of the volume of the target region to the volume of the liver, the target region being a left outer lobe upper segment and a left outer lobe lower segment region; when the proportion of the volume of the target region to the volume of the liver meets a preset threshold interval, extracting effective surface vertices of the target region and constructing a surface mesh; marking the nodule vertices in the surface mesh and weighting the heights of the nodule vertices, performing quadratic polynomial surface fitting on the weighted surface mesh to obtain a smooth reference surface; performing multi-dimensional surface feature extraction on the real surface mesh based on the smooth reference surface; and performing liver surface nodule scoring and liver fibrosis staging evaluation according to the multi-dimensional surface features.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an intelligent assessment method and system for liver fibrosis based on abdominal CT images. Background Technology

[0002] Hepatic fibrosis (HF) is the liver's repair response to chronic damage. It is the common end stage of all chronic liver diseases and the single most important factor determining long-term clinical outcomes related to the liver. Liver surface nodularity (LSN) refers to irregularly shaped nodules on the liver surface caused by fibrosis or regeneration. Their size and number increase with disease severity. Liver fibrosis leads to hepatocyte necrosis, regeneration, and collagen deposition, resulting in an uneven liver surface and eventually nodule formation. If liver fibrosis is not addressed promptly, it will progress to end-stage liver diseases such as cirrhosis, liver failure, and liver cancer, seriously threatening patients' lives. HF is a pre-cirrhotic lesion and is histologically reversible. Therefore, timely and accurate assessment of the degree of liver fibrosis is crucial for the treatment and prognosis of chronic hepatitis B.

[0003] Currently, various methods for assessing liver fibrosis have been developed. Liver biopsy histopathology is the "gold standard" for HF diagnosis, but it is an invasive procedure with risks of intrusion, sampling errors, and complications. Non-invasive methods fall into two main categories: hematological markers and imaging techniques. Hematological markers, such as FIB-4 and APRI, have limited assessive value as a single indicator, lack highly specific indicators, and have low diagnostic efficacy for all stages of fibrosis, especially intermediate stages. Imaging methods include conventional imaging examinations such as ultrasound, CT, and MRI, as well as elastography methods such as transient elastography (TE) and magnetic resonance elastography (MRE). Conventional imaging examinations have low sensitivity for fibrosis, rely on physician experience and subjective judgment, and are mainly used for the diagnosis of cirrhosis. TE has high diagnostic accuracy, but is significantly affected by factors such as obesity and ascites, and requires highly skilled operators. MRE has high quantitative accuracy, but relies on specialized equipment, is complex to operate, and is expensive, making it difficult to popularize. Abdominal CT, as one of the most commonly used examination methods, is frequently used for the diagnosis and screening of various abdominal diseases and traumas. Quantitative assessment of liver surface nodules using abdominal CT has been proven useful for the diagnosis and staging of liver fibrosis. However, most existing techniques analyze and measure within 2D CT images, leading to a lack of spatial information and a limited quantification dimension, which affects the accuracy of the assessment results. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a method and system for intelligent assessment of liver fibrosis based on abdominal CT images to overcome the above problems.

[0005] This invention provides an intelligent assessment method for liver fibrosis based on abdominal CT images, the method comprising:

[0006] Obtain abdominal CT images including the liver region;

[0007] Liver region images were extracted from abdominal CT images, and 3D volumetric data of the liver region were constructed based on the liver region images;

[0008] The 3D volume data is segmented using a pre-trained image segmentation model. Based on the segmentation results, a 3D whole liver mask and a target region mask are formed. The volume of the target region is calculated as a proportion of the total volume of the liver. The target regions are the upper and lower segments of the left lateral lobe.

[0009] When the volume of the target region accounts for a proportion of the total volume of the liver that meets the preset threshold range, the effective surface vertices of the target region are extracted and the real surface mesh of the target region is constructed based on the effective surface vertices.

[0010] Identify the nodal vertices in the real surface mesh, mark the nodal vertices, and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting. Perform quadratic polynomial surface fitting on the dynamically weighted surface mesh to obtain a smooth reference surface of the target region.

[0011] Based on a smooth reference surface, multi-dimensional surface features are extracted from the real surface mesh to form a feature vector;

[0012] Liver surface nodules are scored based on feature vectors, and liver fibrosis is staged based on the scores.

[0013] Another aspect of the present invention provides an intelligent assessment system for liver fibrosis based on abdominal CT images, the system comprising:

[0014] The data acquisition module is used to acquire abdominal CT images containing the liver region;

[0015] The volume data generation module is used to extract liver region images from abdominal CT images and construct 3D volume data of the liver region based on the liver region images;

[0016] The region segmentation module is used to segment the 3D volume data into regions using a pre-trained image segmentation model, and to form a 3D whole liver mask and a target region mask based on the segmentation results. It also calculates the proportion of the target region's volume to the total volume of the liver. The target regions are the upper and lower segments of the left lateral lobe.

[0017] The real surface construction module is used to extract the effective surface vertices of the target region and construct the real surface mesh of the target region based on the effective surface vertices when the proportion of the volume of the target region to the total volume of the liver meets the preset threshold range.

[0018] The reference surface construction module is used to identify nodal vertices in the real surface mesh, mark the nodal vertices, and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting. The dynamically weighted surface mesh is then fitted with a quadratic polynomial to obtain a smooth reference surface for the target region.

[0019] The feature extraction module is used to extract multi-dimensional surface features from the real surface mesh based on a smooth reference surface and form a feature vector;

[0020] The assessment module is used to score liver surface nodules based on feature vectors and to assess liver fibrosis staging based on the scores.

[0021] The present invention provides a method and system for intelligent assessment of liver fibrosis based on abdominal CT images. This method extends the existing quantitative measurement of liver fibrosis based on abdominal CT images from 2D images to 3D surfaces. It expands the quantitative indicators from a single liver surface nodule score to multi-dimensional surface feature indicators, enriching the quantitative dimensions and realizing quantitative scoring of liver surface nodules. This enables precise staging of liver fibrosis. The present invention can provide a more comprehensive and multi-dimensional analytical perspective for clinical practice, improving the efficiency and accuracy of intelligent assessment of liver fibrosis.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings:

[0024] Figure 1 This is a flowchart of an intelligent assessment method for liver fibrosis based on abdominal CT images, according to an embodiment of the present invention.

[0025] Figure 2 This is a structural block diagram of an intelligent assessment system for liver fibrosis based on abdominal CT images, according to an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined.

[0028] To address the shortcomings of existing liver fibrosis measurement and assessment methods, such as insufficient utilization of the three-dimensional information of CT volume and limited evaluation parameters, this invention provides an intelligent assessment method for liver fibrosis based on abdominal CT images, such as... Figure 1 As shown, the intelligent assessment method for liver fibrosis based on abdominal CT images proposed in this invention includes the following steps:

[0029] S11. Obtain abdominal CT images containing the liver region.

[0030] This step involves importing the original abdominal enhanced CT portal venous phase DICOM sequence data to obtain abdominal CT images containing the liver region.

[0031] S12. Extract liver region images from abdominal CT images and construct 3D volumetric data of the liver region based on the liver region images.

[0032] S13. The 3D volume data is segmented using a pre-trained image segmentation model. Based on the segmentation results, a 3D whole liver mask and a target region mask are formed. The volume of the target region is calculated as a proportion of the total volume of the liver. The target region is the upper segment of the left lateral lobe and the lower segment of the left lateral lobe.

[0033] S14. When the volume of the target region accounts for the proportion of the total volume of the liver within a preset threshold range, extract the effective surface vertices of the target region and construct the real surface mesh of the target region based on the effective surface vertices.

[0034] In this embodiment, a trigger condition is preset, that is, when the II+III segment mask is generated and the volume ratio is within a preset threshold range, the surface vertex extraction process is triggered. The threshold range can be selected as 10%-25%. When the ratio is abnormal, a prompt "there may be a segmentation error" is displayed so that the user can confirm whether to continue triggering the surface vertex extraction process.

[0035] S15. Identify the nodal vertices in the real surface mesh, mark the nodal vertices, and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting. Perform quadratic polynomial surface fitting on the dynamically weighted surface mesh to obtain a smooth reference surface of the target region.

[0036] In this embodiment, based on the pathological characteristics of "increased local height difference due to nodule protrusion" and "more pronounced enhancement of fibrotic nodules (higher HU value)," the nodule vertices in the surface mesh are accurately identified through local height difference calculation and CT value verification. By reducing the weight of nodule vertices, the "lifting" effect of large nodules on the reference surface is reduced, ensuring that the fit reflects the normal surface morphology. The final output includes a vertex weighting coefficient matrix (N×1), a nodule vertex label matrix (N×1), and statistics on nodule vertex height differences (mean and maximum). This invention can accurately distinguish between normal and nodule vertices, reduce the interference of nodules on the reference surface, and avoid the subsequent underestimation of the liver surface nodule score LSN.

[0037] S16. Based on the smooth reference surface, multi-dimensional surface features are extracted from the real surface mesh to form a feature vector.

[0038] S17. Score liver surface nodules based on feature vectors, and assess liver fibrosis staging based on the scores.

[0039] The intelligent assessment method for liver fibrosis based on abdominal CT images provided by this invention extends the existing quantitative measurement of liver fibrosis based on abdominal CT images from 2D images to 3D surfaces. It expands the quantitative indicators from a single liver surface nodule score to multi-dimensional surface feature indicators, enriching the quantitative dimensions and realizing quantitative scoring of liver surface nodules. This enables precise staging of liver fibrosis. This invention can provide a more comprehensive and multi-dimensional analytical perspective for clinical practice, improving the efficiency and accuracy of intelligent assessment of liver fibrosis.

[0040] In this embodiment of the invention, step S12, which involves extracting liver region images from abdominal CT images and constructing 3D CT data of the liver region based on the liver region images, specifically includes the following steps;

[0041] S121. Based on the HU value characteristics of liver tissue, the abdominal CT image is cropped in grayscale using a preset HU threshold range to remove voxel regions of other irrelevant tissues in the image besides the liver region.

[0042] In this embodiment, a HU threshold range is set based on the HU value characteristics of liver tissue, such as 50 to 70 HU for normal liver parenchyma and 100 to 50 HU for fat. The abdominal CT image is grayscale cropped according to the HU threshold range, and only voxels whose CT values ​​fall within the preset range are retained. The rest are excluded to remove bones, air, etc., and eliminate interference from irrelevant tissues, thereby achieving a rough extraction of organs or lesions.

[0043] S122. Use trilinear interpolation to resample voxels to a uniform size of a preset space.

[0044] In this embodiment, trilinear interpolation to 1mm×1mm×1mm is used to achieve equal voxel resampling, eliminate the layer thickness difference between different devices, and ensure uniform vertex spacing on the 3D surface.

[0045] S123. Gaussian filtering is applied to the left hepatic lobe region in the resampled CT images to avoid over-smoothing of the right hepatic lobe through directional filtering.

[0046] In this embodiment, 3D Gaussian filtering is performed only on the left lobe of the liver, that is, the region where the x-coordinate value is less than the x-coordinate value of the image center point. This reduces the impact of respiratory artifacts on surface extraction while preserving the small nodules in the left lobe.

[0047] S124. Perform HU value normalization on the filtered CT images to obtain 3D CT data with standardized volume and voxel spacing information. The normalization formula is: normalized=(HU-μ) / σ.

[0048] Where μ is the mean value of liver parenchyma and σ is the standard deviation of liver parenchyma. This invention reduces measurement bias caused by differences in equipment and scanning parameters by standardizing data quality, thus providing stable input for subsequent segmentation.

[0049] In this embodiment of the invention, a pre-trained lightweight SwinUNETR segmentation model is used as the image segmentation model for 3D whole liver segmentation, and a post-processing module is used to eliminate segmentation noise and remove mis-segmented regions.

[0050] Specifically, the training steps of the image segmentation model include: extracting liver region images from CT image data samples in a preset abdominal CT image dataset, constructing 3D volume data samples of the liver region based on the liver region images, and labeling the whole liver region and Couinaud segmented regions for each 3D volume data sample to obtain a training dataset; and using the training dataset to learn and train the preset SwinUNETR segmentation model to obtain the image segmentation model.

[0051] The SwinUNETR model combines the global perspective of Transformer with the local features of CNN. It takes a standardized volume as input and outputs 8 Couinaud segmented probability maps (0-1). It performs binarization with a 0.5 threshold and enhances the segmentation of the target region through a region-weighted loss.

[0052] Post-processing operations are used to eliminate segmentation noise based on morphological principles, specifically including the following implementations.

[0053] Connectivity filtering: retain the maximum connectivity of each segment, and remove gallbladder and small blood vessels;

[0054] Morphological repair: Perform 3D closing operation (3×3×3 spherical kernel) on the II+III segment mask; remove mis-segmented areas based on anatomical priors, such as the density difference between the gallbladder and liver;

[0055] Boundary smoothing: Linear interpolation corrects the jagged edges at the boundary between segments II+III and IV.

[0056] The final output includes a 3D whole liver mask, a mask for segments II and III, liver centroid coordinates, and the volume percentage of segments II and III, enabling precise division of the anatomical range of segments II and III. This defines the target area for subsequent surface extraction and avoids interference from irrelevant liver segments.

[0057] This invention applies the lightweight SwinUNETR segmentation model, which can intelligently analyze multimodal medical image data such as CT and MRI, automatically segment relevant regions and perform 3D reconstruction, and calculate multidimensional parameters based on the fitted surface. The SwinUNETR model is a 3D medical image segmentation network based on Swin Transformer. Its core advantages are: (1) hierarchical feature extraction: capturing features at different scales through multi-stage downsampling and upsampling processes; (2) local attention mechanism: modeling long-distance dependencies while maintaining computational efficiency; (3) shift window strategy: enhancing the information interaction capability between different windows.

[0058] In this embodiment of the invention, step S14, extracting the effective surface vertices of the target region, includes the following steps:

[0059] S141. Locate the target area using a bounding box, and determine the three-dimensional range of the target area based on the coordinates of the bounding box.

[0060] Specifically, the target area is labeled using bounding boxes, the bounding boxes are extracted, and the three-dimensional range of segment II+III (x_{min}-x_{max}, y_{min}-y_{max}, z_{min}-z_{max}) is determined.

[0061] S142. Based on the preset voxel grayscale threshold, extract isosurfaces to generate initial surface vertices and patches.

[0062] Specifically, isosurfaces are extracted based on voxel grayscale thresholding to generate initial vertices, forming surface vertices and patches. The grayscale threshold can be selected as 0.5, and the initial number of vertices can be selected as 60,000 to 80,000.

[0063] S143. Based on the anatomical orientation constraints, tissue contrast constraints, and / or normal vector constraints corresponding to the target region, invalid points in the initial surface vertices are eliminated to obtain the effective surface vertices and patches of the target region.

[0064] This invention selects effective surface vertices by combining anatomical orientation (left anterior) and tissue contrast, obtaining effective surface vertices (N×3 coordinates), triangular patches (M×3 indices), and vertex normal vectors (N×3 directions) for segments II+III. In a specific example, vertices containing fat in their neighborhood can be retained by verifying tissue contrast, such as those with a HU value of -100 to -50. Vertices whose normal vectors point to the left anterior can be retained by constraining the normal vectors, with the following constraints: the angle with the negative x-axis < 45° and the angle with the positive y-axis < 30°. N represents the first dimension in "effective surface vertices (N×3 coordinates)," representing the total number of effective vertices, and "×3" indicates that each vertex contains three spatial coordinate values: x, y, and z. M is the first dimension in "triangular patches (M×3 indices)," representing the total number of triangular patches, and "×3" indicates that each triangular patch consists of the indices of three vertices.

[0065] Specifically, tissue contrast is a "density difference attribute between different tissues," and the process of implementing tissue contrast constraints is as follows:

[0066] (1) Identify the target tissue: the parenchyma and adipose tissue of segment II+III of the liver, which are the source of the liver surface boundary;

[0067] (2) Vertex neighborhood extraction: For each initial surface vertex, extract a 3*3*3 voxel neighborhood centered on the vertex to avoid misjudgment based solely on the HU value of a single vertex;

[0068] (3) Neighborhood HU value statistics: Determine whether two conditions are met simultaneously: first, the presence of liver parenchyma voxels, and second, the presence of fat voxels. If both conditions are not met simultaneously, the vertex is judged as invalid.

[0069] (4) Threshold optimization: If the patient is thin, the fat HU can be relaxed to -120 to -30 to avoid missed detection.

[0070] Contrast verification is not simply a matter of HU value screening. It requires the condition of "coexistence of target tissues in the neighborhood". The HU value of a single vertex cannot reflect the contrast. The boundary relationship between "liver parenchyma and target tissue" must be verified through neighborhood analysis in order to achieve true contrast constraint.

[0071] This invention can eliminate invalid vertices, reduce the amount of computation in subsequent processing, and at the same time ensure that surface vertices only come from the outer surface of segment II+III, avoiding interference from internal structures in LSN calculation.

[0072] Furthermore, the step S14 of constructing a realistic surface mesh of the target region based on effective surface vertices includes the following steps:

[0073] S144. For a surface mesh formed by effective surface vertices, bilateral filtering is implemented by simultaneously calculating the spatial distance weight and the normal similarity weight, so as to preserve the nodal edges with abrupt changes in normal vectors while smoothing surface mesh noise.

[0074] Specifically, bilateral filtering: spatial domain σ1=1.2mm, normal vector domain σ2=0.1rad, preserving nodule edges.

[0075] This invention employs joint weighting of the spatial and normal domains to smooth noise while preserving nodule edges with abrupt changes in normal vectors. Spatial weights suppress the influence of distant pixels, while normal vector weights protect regions with abrupt changes in normal vectors, such as nodule edges and ridges. Furthermore, adaptive parameter adjustment can be implemented to dynamically adjust the filtering radius based on local curvature, avoiding over-smoothing of features.

[0076] S145. Construct a triangular mesh topology based on the surface vertices after bilateral filtering. Merge distorted triangles by optimizing and adjusting the vertex spacing of degenerate mesh cells in the triangular mesh topology. Fill the holes in the triangular mesh topology with curved surfaces. Perform region retopology on self-intersecting regions to eliminate non-manifold structures and obtain the repaired surface mesh.

[0077] This invention is based on the principle of triangular mesh topology to eliminate defects such as degradation, voids, and self-intersections, thus ensuring surface continuity.

[0078] S146. The repaired surface mesh is uniformly resampled using the Poisson disk sampling algorithm to obtain the real surface mesh.

[0079] This application controls vertex density using a Poisson disk sampling algorithm, balancing accuracy and computational efficiency. The output is an optimized surface mesh with a surface continuity index ≥ 0.95.

[0080] This invention obtains an optimized II+III segment surface mesh (vertex V, facet F, normal vector N) and vertex curvature annotations through surface optimization and topology repair, which improves the integrity, smoothness and uniformity of the surface mesh and provides a high-quality foundation for subsequent reference surface fitting and distance calculation.

[0081] In an optional embodiment of the present invention, step S15, which involves performing quadratic polynomial surface fitting on the dynamically weighted surface mesh, specifically includes: dividing the surface mesh into an upper segment mesh region of the left lateral leaf and a lower segment mesh region of the left lateral leaf; and, based on the wedge-shaped anatomical features of the target region, performing surface fitting on the upper segment mesh region of the left lateral leaf and the lower segment mesh region of the left lateral leaf using a quadratic polynomial.

[0082] Specifically, this invention is based on the wedge-shaped anatomical features of segment II+III, which are "narrow on the left and wide on the right, steep at the top and gentle at the bottom," and uses a 3D quadratic polynomial for fitting. The quadratic polynomial is as follows:

[0083]

[0084] To accommodate the curvature difference between segment II (steeper) and segment III (gentler), this invention divides segment II+III into a first segment (segment II) and a second segment (segment III) for separate fitting, avoiding local deviations caused by single fitting. The reliability of the fitting is verified by calculating the fitting residual and surface smoothness, ensuring that the reference surface is not significantly distorted. Finally, the coordinates of the reference smooth surface of segment II+III, the fitting coefficients, the fitting residuals, and the smoothness score are output.

[0085] In an optional embodiment of the present invention, step S16, which involves extracting multi-dimensional surface features from the real surface mesh based on a smooth reference surface, specifically includes the following steps not shown in the accompanying drawings:

[0086] S161. Calculate the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface to obtain the vertex-level distance matrix.

[0087] S162. Calculate the median Euclidean distance from each vertex in the real surface mesh to the smooth reference surface, and use the obtained median as a characterization feature of the overall bulge degree of the nodule. This embodiment uses the median distance. As a characteristic representing the overall protrusion of the nodule, it directly reflects the overall protrusion of the nodule in segment II+III.

[0088] S163. Count the number of target vertices whose Euclidean distance is greater than a preset distance threshold, and use the proportion of the target vertices to the total number of vertices as the characterization feature of nodule distribution density. This embodiment uses nodule density. As a characteristic representing the distribution density of nodules, it reflects the density of nodule distribution and supplements the shortcomings of single distance features.

[0089] Furthermore, step S17, which involves scoring liver surface nodules based on feature vectors, specifically includes:

[0090] The overall protrusion degree and distribution density of nodules were standardized separately, with the following features:

[0091] Standardization of the overall protrusion characteristic of nodules refers to mapping the original values ​​(in mm) to a score of 0-10. The mapping rule is based on the range of normal / cirrhotic populations (0=0mm, 10=5mm, and scores above 5mm are counted as 10). The standardization formula for the overall protrusion characteristic of nodules is as follows:

[0092] ;

[0093] Standardization of nodule distribution density characteristics refers to mapping the original values ​​(percentages) to a score of 0-10, with the mapping rule (0=0%, 10=50%, values ​​exceeding 50% are counted as 10). The standardization formula for nodule distribution density characteristics is as follows:

[0094] ;

[0095] in, Standardized characterization features for the overall protrusion degree of the nodule. This characterizes the overall protrusion of the nodule. To standardize the characterization features of nodule distribution density, Characterized by the distribution density of nodules;

[0096] Standardized characterization features of the overall protrusion degree of the nodule and A linearly weighted summation is performed, and the summation result is used as the liver surface nodule score. The linearly weighted summation formula is as follows:

[0097] .

[0098] Specifically, fixed weighting can be based on small clinical sample acceptance, setting a fixed weighting ratio. Weight 70%, Weighting: 30%.

[0099] This invention focuses on core features (characteristics of overall nodule protrusion and nodule distribution density) and uses fixed-weight linear fusion to calculate liver surface nodule scores.

[0100] In another optional embodiment of the present invention, step S16, which involves extracting multi-dimensional surface features from the real surface mesh based on a smooth reference surface, specifically includes the following steps not shown in the accompanying drawings:

[0101] S161' Calculate the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface to obtain the vertex-level distance matrix.

[0102] S162' Calculate the median of the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface, and use the obtained median as a characterization feature of the overall protrusion degree of the nodule.

[0103] Vertex-level distance: (Euclidean distance, unit) ).

[0104] Calculate median distance , Let i be the coordinates of vertex i in the real surface mesh. , Let be the coordinates of the point corresponding to vertex i in the smooth reference surface.

[0105] S163' Count the number of target vertices whose Euclidean distance is greater than a preset distance threshold, and use the proportion of the number of target vertices in the total number of vertices as the characteristic of nodule distribution density.

[0106] Optionally, the distance threshold can be set to 1.8. 1.8 Optimize the threshold for viral hepatitis. Understandably, the distance threshold can be dynamically adjusted based on the actual application.

[0107] In this embodiment, the density of nodule distribution is reflected by statistically analyzing the proportion of high-deviation vertices. ;

[0108] ;

[0109] S164' Calculate the slope difference between each vertex in the real surface mesh and its corresponding point on the preset standard surface, form a dataset from all slope differences and calculate the mean of the dataset, and use the mean as a morphological deviation characterization feature.

[0110] Among them, morphological deviation is based on the normal wedge-shaped morphology (standard surface) of segment II+III, and quantifies the structural distortion caused by fibrosis.

[0111] Ideal wedge surface: Based on the morphological template of normal population segment II+III, construct an ideal wedge curve in the "xz plane":

[0112] formula: . in The slope The intercept;

[0113] Deviation calculation: ;

[0114] in, This represents the number of sampling points in the xz plane. For the actual surface at the first The slope of the point The ideal slope reflects the morphological distortion caused by fibrosis in segments II and III.

[0115] S165' Calculate the angle between the normal vector of each vertex in the real surface mesh and the normal vector of its corresponding point on the smooth reference surface, form a dataset from all the normal vector angles, and calculate the standard deviation of the dataset. Use the standard deviation as a surface roughness characterization feature.

[0116] Surface roughness reflects the microscopic irregularities of the surface, supplementing the shortcomings of macroscopic distance characteristics. Specifically, the normal vector of each vertex is first obtained, and the angle between this normal vector and the normal vector of its corresponding point on the smooth reference surface is calculated. All angle values ​​are compiled into a dataset, and finally the standard deviation of the dataset is calculated.

[0117] Calculate the angle between the vertex normal vector of the original surface and the normal vector of the reference surface. ;

[0118] Roughness: ,

[0119] in The average included angle reflects the degree of microscopic irregularity on the surface. This represents the number of valid vertices.

[0120] Finally, the output feature vector is... , , , and vertex set distance matrix .

[0121] The embodiments of the present invention comprehensively characterize nodule features from four dimensions: "macroscopic deviation, distribution density, morphological distortion, and microscopic roughness," providing multi-dimensional basis for LSN scoring and improving diagnostic efficiency.

[0122] Furthermore, step S17, which involves scoring liver surface nodules based on feature vectors, specifically includes:

[0123] First, identify the core features in the feature vector and increase their weight values. Specifically, a pre-set random forest weight learning model can be used to identify the core features in the feature vector and increase their weight values. The random forest model can be pre-trained based on pathological annotation data to assign higher weights to the core features.

[0124] Then, the etiology information input by the user is obtained, and the weight adjustment parameters corresponding to the etiology information are determined according to a preset mapping table. The weights of each feature in the feature vector are adjusted according to the weight adjustment parameters. The mapping table includes the correspondence between different etiologies and weight adjustment parameters. This invention adjusts the weights based on the pathological characteristics of different etiologies, such as denser nodules in viral hepatitis and more obvious morphological distortion in NAFLD, to improve specificity.

[0125] In a specific example, the determination of core feature weights is implemented as follows: Based on training data from 300 clinical patients with pathological staging diagnoses, a random forest model is used to calculate feature importance, and the final weight allocation is as follows:

[0126] 45% (Core indicator, with the highest correlation to pathological staging)

[0127] 30% (reflecting nodule distribution density)

[0128] 20% (specific index of wedge morphology)

[0129] 5% (auxiliary reflection of microscopic nodules).

[0130] Further etiological adaptive adjustment was conducted, and literature data indicated that patients with viral hepatitis had a higher density of segment II+III nodules, therefore... The weighting has been increased to 35%. The percentage dropped to 15%; NAFLD patients showed more pronounced segmental distortion in segments II and III, therefore... The weighting has been increased to 25%. It dropped to 25%.

[0131] Finally, liver surface nodules are scored based on each feature and the weight corresponding to each feature.

[0132] The scoring model is shown below: ;

[0133] The specific formula is as follows: ;

[0134] The scores are then mapped to a scale of 0 to 10, where 0 = normal and 10 = severe.

[0135] This invention integrates multi-dimensional features, combines etiological optimization weights, generates an objective and stable LSN score, and quantifies uncertainty to improve reliability, thereby achieving accurate staging of liver fibrosis.

[0136] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0137] This invention provides an intelligent assessment system for liver fibrosis based on abdominal CT images. The system includes functional modules for implementing the intelligent assessment method for liver fibrosis based on abdominal CT images as described in any of the preceding embodiments. Figure 2 This schematic diagram illustrates the structure of an intelligent assessment system for liver fibrosis based on abdominal CT images provided by an embodiment of the present invention. (Refer to...) Figure 2 The system described in this embodiment of the invention includes:

[0138] Data acquisition module 201 is used to acquire abdominal CT images containing the liver region;

[0139] The volume data generation module 202 is used to extract liver region images from abdominal CT images and construct 3D volume data of the liver region based on the liver region images;

[0140] The region segmentation module 203 is used to segment the 3D volume data using a pre-trained image segmentation model, form a 3D whole liver mask and a target region mask based on the segmentation results, and calculate the proportion of the volume of the target region to the total volume of the liver. The target region is the upper segment of the left lateral lobe and the lower segment of the left lateral lobe.

[0141] The real surface construction module 204 is used to extract the effective surface vertices of the target region and construct the real surface mesh of the target region based on the effective surface vertices when the proportion of the volume of the target region to the total volume of the liver meets the preset threshold range.

[0142] The reference surface construction module 205 is used to identify the nodal vertices in the real surface mesh, mark the nodal vertices and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting, and perform quadratic polynomial surface fitting on the dynamically weighted surface mesh to obtain a smooth reference surface of the target region.

[0143] Feature extraction module 206 is used to extract multi-dimensional surface features from the real surface mesh based on a smooth reference surface and form a feature vector;

[0144] Assessment module 207 is used to score liver surface nodules based on feature vectors and to assess liver fibrosis staging based on the scores.

[0145] In this embodiment of the invention, the volume data generation module 202 includes:

[0146] The grayscale cropping unit is used to crop the abdominal CT image in grayscale according to the HU value characteristics of liver tissue using a preset HU threshold range, so as to remove voxel regions of other irrelevant tissues in the image besides the liver region.

[0147] The first resampling unit is used to resample voxels to a uniform size of a preset space using trilinear interpolation.

[0148] The filtering unit is used to perform Gaussian filtering on the left hepatic lobe region in the resampled CT image in order to avoid over-smoothing of the right hepatic lobe through directional filtering;

[0149] The normalization processing unit is used to perform HU value normalization processing on the filtered CT images to obtain 3D CT data with standardized volume and voxel spacing information.

[0150] In this embodiment of the invention, the real surface construction module 204 includes:

[0151] The annotation unit is used to locate the annotation box of the target area and determine the three-dimensional range of the target area based on the coordinates of the annotation box;

[0152] The isosurface extraction unit is used to extract isosurfaces based on a preset voxel grayscale threshold to generate initial surface vertices and patches.

[0153] The vertex optimization unit is used to eliminate invalid points in the initial surface vertices based on the anatomical orientation constraints, tissue contrast constraints, and / or normal vector constraints corresponding to the target region, so as to obtain the effective surface vertices and patches of the target region.

[0154] Furthermore, the real surface building module 204 also includes:

[0155] The bilateral filtering unit is used to perform bilateral filtering on the surface mesh formed by effective surface vertices by simultaneously calculating the spatial distance weight and the normal similarity weight, so as to preserve the nodal edges with abrupt changes in normal vectors while smoothing surface mesh noise;

[0156] The topology repair unit is used to construct a triangular mesh topology based on the surface vertices after bilateral filtering. It optimizes the vertex spacing of degenerate mesh cells in the triangular mesh topology to merge distorted triangles, fills the holes in the triangular mesh topology with curved surfaces, and performs region retopology on self-intersecting regions to eliminate non-manifold structures, thus obtaining the repaired surface mesh.

[0157] The second resampling unit is used to uniformly resample the repaired surface mesh using the Poisson disk sampling algorithm to obtain the real surface mesh.

[0158] In this embodiment of the invention, the reference surface construction module 205 is specifically used to divide the surface mesh into the upper segment mesh region of the left lateral leaf and the lower segment mesh region of the left lateral leaf; based on the wedge-shaped anatomical features of the target region, a quadratic polynomial is used to perform surface fitting on the upper segment mesh region of the left lateral leaf and the lower segment mesh region of the left lateral leaf respectively.

[0159] In one embodiment of the present invention, the feature extraction module 206 is specifically used to calculate the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface to obtain a vertex-level distance matrix; to count the median of the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface, and to use the obtained median as the overall protrusion characterization feature of the nodule; and to count the number of target vertices whose Euclidean distance is greater than a preset distance threshold, and to use the proportion of the number of target vertices in the total number of vertices as the nodule distribution density characterization feature.

[0160] In this embodiment, the evaluation module 207 is specifically used to perform feature standardization processing on the overall protrusion degree characterization feature and the nodule distribution density characterization feature, respectively, wherein:

[0161] The standardized formula for characterizing the overall protrusion of a nodule is as follows:

[0162] = ;

[0163] The standardized formula for characterizing nodule distribution density is as follows:

[0164] ;

[0165] in, Standardized characterization features for the overall protrusion degree of the nodule. This characterizes the overall protrusion of the nodule. To standardize the characterization features of nodule distribution density, Characterized by the distribution density of nodules;

[0166] The standardized characteristics of the overall protrusion of the nodules and the standardized characteristics of the nodule distribution density were linearly weighted and summed. The summation result was used as the liver surface nodule score.

[0167] .

[0168] In another embodiment of the present invention, the feature extraction module 206 is specifically used to calculate the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface, thereby obtaining a vertex-level distance matrix; to calculate the median of the Euclidean distances from each vertex in the real surface mesh to the smooth reference surface, and to use the obtained median as a characterization feature of the overall convexity of the nodules; to calculate the number of target vertices whose Euclidean distance is greater than a preset distance threshold, and to use the proportion of the number of target vertices in the total number of vertices as a characterization feature of the nodule distribution density; to calculate the slope difference between each vertex in the real surface mesh and its corresponding point on a preset standard surface, to form a dataset from all slope differences and to calculate the mean of the dataset, and to use the mean as a characterization feature of the morphological deviation; and to calculate the angle between the normal vector of each vertex in the real surface mesh and its adjacent points, to form a dataset from all normal vector angles and to calculate the standard deviation of the dataset, and to use the standard deviation as a characterization feature of the surface roughness.

[0169] In this embodiment, the evaluation module 207 is specifically used to identify the core features in the feature vector and increase the weight value of the core features; obtain the etiology information input by the user, determine the weight adjustment parameter corresponding to the etiology information according to the preset mapping relationship table, and adjust the weight of each feature in the feature vector according to the weight adjustment parameter; the mapping relationship table includes the correspondence between different etiology information and weight adjustment parameters; and score the liver surface nodules according to each feature and the weight corresponding to each feature.

[0170] As the system implementation is basically similar to the method implementation, the description is relatively simple, and relevant parts can be found in the description of the method implementation.

[0171] Furthermore, another embodiment of the present invention provides a computer program product storing a computer program that, when executed by a processor, implements the steps described in the above embodiment of the intelligent assessment method for liver fibrosis based on abdominal CT images, for example... Figure 1 Steps S11-S17 are shown.

[0172] Furthermore, another embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps described in the above embodiment of the intelligent assessment method for liver fibrosis based on abdominal CT images, for example... Figure 1 Steps S11-S17 are shown.

[0173] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, any of the claimed embodiments can be used in any combination.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart assessment method for liver fibrosis based on abdominal CT images, characterized in that, The method includes: Obtain abdominal CT images containing the liver region; Liver region images were extracted from abdominal CT images, and 3D volumetric data of the liver region were constructed based on the liver region images; The 3D volume data is segmented using a pre-trained image segmentation model. Based on the segmentation results, a 3D whole liver mask and a target region mask are formed. The volume of the target region is calculated as a proportion of the total volume of the liver. The target regions are the upper and lower segments of the left lateral lobe. When the volume of the target region accounts for a proportion of the total volume of the liver that meets the preset threshold range, the effective surface vertices of the target region are extracted and the real surface mesh of the target region is constructed based on the effective surface vertices. Identify the nodal vertices in the real surface mesh, mark the nodal vertices, and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting. Perform quadratic polynomial surface fitting on the dynamically weighted surface mesh to obtain a smooth reference surface of the target region. Based on a smooth reference surface, multi-dimensional surface features are extracted from the real surface mesh to form a feature vector; Liver surface nodules are scored based on feature vectors, and liver fibrosis is staged based on the scores.

2. The method according to claim 1, characterized in that, Liver region images were extracted from abdominal CT images, and 3D CT data of the liver region were constructed based on the liver region images, including; Based on the HU value characteristics of liver tissue, the abdominal CT images are cropped in grayscale using a preset HU threshold range to remove voxel regions of other irrelevant tissues in the images, except for the liver region. The voxels are resampled to a uniform size within a preset spatial area using trilinear interpolation. Gaussian filtering was applied to the left hepatic lobe region in the resampled CT images to avoid over-smoothing of the right hepatic lobe through directional filtering; The filtered CT images were normalized by HU value to obtain 3D CT data with standardized volume and voxel spacing information.

3. The method according to claim 1, characterized in that, The training steps of the image segmentation model include: extracting liver region images from CT image data samples in a preset abdominal CT image dataset, constructing 3D volume data samples of the liver region based on the liver region images, and labeling the whole liver region and Couinaud segmented regions for each 3D volume data sample to obtain a training dataset; and using the training dataset to learn and train a preset SwinUNETR segmentation model to obtain an image segmentation model.

4. The method according to claim 1, characterized in that, Extract the effective surface vertices of the target region, including: The target area is located by defining a bounding box, and the three-dimensional range of the target area is determined based on the coordinates of the bounding box. Isosurfaces are extracted based on a preset voxel grayscale threshold to generate initial surface vertices and patches; Based on the anatomical orientation constraints, tissue contrast constraints, and / or normal vector constraints corresponding to the target region, invalid points in the initial surface vertices are eliminated to obtain the effective surface vertices and patches of the target region.

5. The method according to claim 4, characterized in that, Constructing a realistic surface mesh for the target region based on effective surface vertices includes: For a surface mesh formed by effective surface vertices, bilateral filtering is implemented by simultaneously calculating spatial distance weights and normal similarity weights to preserve nodule edges with abrupt changes in normal vectors while smoothing surface mesh noise; A triangular mesh topology is constructed based on the surface vertices after bilateral filtering. The degenerate mesh cells in the triangular mesh topology are optimized by adjusting the vertex spacing to merge distorted triangles. The holes in the triangular mesh topology are filled with curved surfaces. The self-intersecting regions are retopologically reconstructed to eliminate non-manifold structures, resulting in the repaired surface mesh. The repaired surface mesh is uniformly resampled using the Poisson disk sampling algorithm to obtain the true surface mesh.

6. The method according to claim 1, characterized in that, Multi-dimensional surface feature extraction is performed on the real surface mesh based on a smooth reference surface, including: Calculate the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface to obtain the vertex-level distance matrix; The median of the Euclidean distance from each vertex in the real surface mesh to the smooth reference surface is calculated, and the obtained median is used as a characterization feature of the overall convexity of the nodules. The number of target vertices whose Euclidean distance is greater than a preset distance threshold is counted, and the proportion of the number of target vertices to the total number of vertices is used as a characteristic of nodule distribution density.

7. The method according to claim 6, characterized in that, Liver surface nodules are scored based on feature vectors, including: The overall protrusion degree and distribution density of nodules were standardized separately, with the following features: The standardized formula for characterizing the overall protrusion of a nodule is as follows: ； The standardized formula for characterizing nodule distribution density is as follows: ； in, Standardized characterization features for the overall protrusion degree of the nodule. This characterizes the overall protrusion of the nodule. To standardize the characterization features of nodule distribution density, Characterized by the distribution density of nodules; The standardized characteristics of the overall protrusion of the nodules and the standardized characteristics of the nodule distribution density were linearly weighted and summed. The summation result was used as the liver surface nodule score. 。 8. The method according to claim 6, characterized in that, The process of extracting multi-dimensional surface features from the real surface mesh based on a smooth reference surface also includes: Calculate the slope difference between each vertex in the real surface mesh and its corresponding point on the preset standard surface, combine all slope differences into a dataset and calculate the mean of the dataset, and use the mean as a morphological deviation characterization feature. Calculate the angle between the normal vector of each vertex in the real surface mesh and the normal vector of its corresponding point on the smooth reference surface. Compile all the normal vector angles into a dataset and calculate the standard deviation of the dataset. Use the standard deviation as a surface roughness characterization feature.

9. The method according to claim 8, characterized in that, Liver surface nodules are scored based on feature vectors, including: Identify the core features in the feature vector and increase the weight value of the core features; The system obtains the etiology information input by the user, determines the weight adjustment parameters corresponding to the etiology information according to a preset mapping table, and adjusts the weights of each feature in the feature vector according to the weight adjustment parameters. The mapping table includes the correspondence between different etiologies and weight adjustment parameters. Liver surface nodules are scored based on each feature and the weight corresponding to each feature.

10. A smart assessment system for liver fibrosis based on abdominal CT images, characterized in that, The system includes: The data acquisition module is used to acquire abdominal CT images containing the liver region; The volume data generation module is used to extract liver region images from abdominal CT images and construct 3D volume data of the liver region based on the liver region images; The region segmentation module is used to segment the 3D volume data into regions using a pre-trained image segmentation model, and to form a 3D whole liver mask and a target region mask based on the segmentation results. It also calculates the proportion of the target region's volume to the total volume of the liver. The target regions are the upper and lower segments of the left lateral lobe. The real surface construction module is used to extract the effective surface vertices of the target region and construct the real surface mesh of the target region based on the effective surface vertices when the proportion of the volume of the target region to the total volume of the liver meets the preset threshold range. The reference surface construction module is used to identify nodal vertices in the real surface mesh, mark the nodal vertices, and dynamically weight the height of each nodal vertex to reduce the weight of the nodal vertex in the surface fitting. The dynamically weighted surface mesh is then fitted with a quadratic polynomial to obtain a smooth reference surface for the target region. The feature extraction module is used to extract multi-dimensional surface features from the real surface mesh based on a smooth reference surface and form a feature vector; The assessment module is used to score liver surface nodules based on feature vectors and to assess liver fibrosis staging based on the scores.

Images