Three-dimensional anatomically-based lung cancer lymph node partition system and method

Abstract

The application discloses a lung cancer lymph node partition system and method based on three-dimensional anatomy. A chest CT three-dimensional image is collected, and a personalized density histogram is constructed by extracting mediastinal voxel CT values; fat and mixed soft tissues are segmented by a Gaussian mixture model, and blood vessels and lymph node candidate regions are extracted in combination with a three-dimensional Hessian matrix. A target function is constructed in combination with a contrast noise ratio and information entropy, and a best window width and window level are adaptively optimized to generate an enhanced three-dimensional data body. Dual-channel tensors are formed by fusing the enhanced data and the original image, and are input into a three-dimensional spatial partition model to output lymph node partition results. The application separates overlapping soft tissues by using three-dimensional anatomy, adaptively matches image contrast parameters, fuses dual-channel features and spatial topological constraints, solves the tissue feature confusion problem caused by traditional fixed parameters, and improves the lung cancer lymph node partition positioning accuracy and anatomical rationality.

Description

Technical Field

[0001] This application relates to the technical field of medical image processing, and in particular to a lung cancer lymph node partitioning system and method based on three-dimensional anatomy. Background Technology

[0002] Precise lymph node localization in lung cancer is a crucial prerequisite for TNM staging, surgical planning, and prognostic assessment. Clinically, this assessment primarily relies on three-dimensional anatomical structures obtained from chest computed tomography (CT) scans. In existing image processing and interpretation workflows, both traditional image segmentation algorithms and deep learning-based three-dimensional spatial localization models mostly process image data directly under population-fixed empirical window width and level parameters (e.g., mediastinal window width 350-400 HU, window level 40-60 HU) or uniform global normalization parameters. However, there are significant differences in mediastinal anatomical characteristics (such as mediastinal fat content, vascular enhancement, and muscle tissue volume) among different patients. Furthermore, in uncontrast-enhanced plain or conventional CT scans, lymph nodes, blood vessels, and surrounding muscles are all soft tissues in terms of physical attenuation characteristics, and their CT value ranges highly overlap (all distributed between approximately +30 HU and +60 HU).

[0003] This processing method, which relies on fixed empirical parameters, cannot adapt to the natural differences in voxel density distribution among individual patients, resulting in extremely poor three-dimensional spatial contrast between lymph node boundaries and mediastinal fat and adjacent blood vessels in some patients with specific body types. Due to the lack of feature separation and individualized density mapping enhancement mechanisms for overlapping soft tissues, existing three-dimensional spatial partitioning models struggle to extract clear edge gradient information from the underlying feature map when processing image data with insufficient contrast. This easily leads to misclassification of unenhanced blood vessels or muscles as lymph nodes, or failure to detect lymph nodes integrated into the background fat. Current partitioning systems have not yet established a technical chain for automatically solving for optimal image contrast parameters based on the individual patient's chest three-dimensional tissue density distribution characteristics. In summary, how to break through the limitations of fixed parameters for the population, and automatically optimize the best display parameters for each individual patient based on their three-dimensional density distribution and tubular structure characteristics, in the face of soft tissue overlap, thereby improving the accuracy of lymph node localization in lung cancer, is an objective technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, one objective of this application is to propose a lung cancer lymph node partitioning system and method based on three-dimensional anatomy. By separating overlapping soft tissues through three-dimensional morphology and adaptively optimizing the best contrast parameters for each individual, combined with dual-channel feature fusion and spatial topological constraints, it effectively overcomes feature confusion caused by traditional fixed empirical parameters and significantly improves the localization accuracy and anatomical rationality of lung cancer lymph node partitioning.

[0006] To achieve the above objectives, the first aspect of this application proposes a method for lung cancer lymph node partitioning based on three-dimensional anatomy, comprising:

[0007] Step S10: Receive chest CT three-dimensional image data, extract the CT value set of all voxels in the mediastinal region of the chest CT three-dimensional image data, and construct an individualized three-dimensional tissue density distribution histogram.

[0008] Step S20: Fit and decompose the three-dimensional tissue density distribution histogram using a Gaussian mixture model to generate tissue density mixing parameters; extract the mediastinal fat region and soft tissue mixing region based on the tissue density mixing parameters; extract tubular structure features from the soft tissue mixing region using a three-dimensional Hessian matrix to generate vascular region and lymph node candidate region.

[0009] Step S30: Calculate the contrast-to-noise ratio and regional information entropy of the candidate lymph node region, the mediastinal fat region, and the vascular region; construct an objective function with the contrast-to-noise ratio and the regional information entropy as variables; use an adaptive optimization algorithm to search for the parameter combination that maximizes the objective function, which is used as the optimal display window width and window level parameters.

[0010] Step S40: Apply the optimal display window width and window level parameters to the chest CT three-dimensional image data for density mapping to generate an enhanced three-dimensional data volume;

[0011] Step S50: The enhanced three-dimensional data volume is stitched together with the chest CT three-dimensional image data to generate dual-channel tensor data, which is then input into a three-dimensional spatial partitioning model for processing, and the lung cancer lymph node partitioning result corresponding to the chest CT three-dimensional image data is output.

[0012] In addition, the lung cancer lymph node partitioning system and method based on three-dimensional anatomy proposed in this application may also have the following additional technical features:

[0013] In one embodiment of this application, step S20, generating tissue density mixing parameters, includes:

[0014] Extract the frequency data from the three-dimensional tissue density distribution histogram;

[0015] The Gaussian mixture model is initialized with model parameters including the mean vector and covariance matrix;

[0016] The frequency data is input into the Gaussian mixture model to perform expectation maximization iterative processing.

[0017] Calculate the posterior probability matrix of each frequency data point belonging to each Gaussian component;

[0018] The updated parameters are generated by updating the mean vector and the covariance matrix using the posterior probability matrix.

[0019] When the change in the log-likelihood function of the Gaussian mixture model is less than the convergence threshold, the expectation maximization iteration process is stopped, and the updated parameters are output as the tissue density mixture parameters, which include a first density interval and a second density interval.

[0020] In one embodiment of this application, step S20, generating the vascular region and lymph node candidate region, includes:

[0021] Extract the set of three-dimensional spatial coordinates of all voxels in the soft tissue mixed region;

[0022] Calculate the second-order partial derivatives of the voxels corresponding to the three-dimensional spatial coordinate set, and construct the three-dimensional Hessian matrix for each voxel;

[0023] Calculate the three eigenvalues ​​of the three-dimensional Hessian matrix for each voxel, and calculate the response value of the tubular structure feature based on the numerical relationship and sign attribute of the three eigenvalues.

[0024] Set a vascular response threshold, extract voxels whose response values ​​are greater than or equal to the vascular response threshold, and generate the vascular region;

[0025] Extract voxels whose response values ​​are less than the vascular response threshold to generate non-vascular residual regions;

[0026] A three-dimensional connected component labeling algorithm is applied to the non-vascular remaining region to extract connected components with a volume greater than a set volume threshold and a sphericity greater than a set sphericity threshold, and these components are combined to generate the lymph node candidate region.

[0027] In one embodiment of this application, step S30, constructing an objective function with the contrast-to-noise ratio and the region information entropy as variables, includes:

[0028] The test display parameters are obtained and mapped onto the chest CT three-dimensional image data to obtain a simulated image matrix;

[0029] The first mean of the lymph node candidate region, the second mean of the mediastinal fat region, and the third mean of the vascular region are extracted from the simulated image matrix.

[0030] Calculate the background noise standard deviation, and calculate the contrast noise ratio based on the first mean, the second mean, the third mean, and the background noise standard deviation;

[0031] Calculate the probability distribution data of pixels within the simulated image matrix to obtain the region information entropy;

[0032] The contrast-to-noise ratio is used as a positive optimization term, and the difference between the regional information entropy and the entropy threshold is transformed by a penalty function and used as a negative constraint term. The objective function is generated by merging these terms.

[0033] In one embodiment of this application, step S30 involves using an adaptive optimization algorithm to search for the parameter combination that maximizes the objective function, as the optimal display window width and window level parameters, including:

[0034] Within a preset parameter search space, select multiple sets of initial test parameter combinations, and input the initial test parameter combinations into the objective function to calculate the corresponding set of initial target values;

[0035] A Gaussian process regression model is constructed using the initial test parameter combination and the initial target value set;

[0036] Based on the predicted mean and predicted variance output by the Gaussian process regression model, a confidence upper bound acquisition function is constructed.

[0037] In each round of optimization iteration, a candidate parameter combination that maximizes the confidence upper bound acquisition function is found and input into the objective function to calculate the current target value, and the Gaussian process regression model is updated.

[0038] When the number of iterations reaches the set maximum iteration threshold, the optimization iteration stops, and the parameters corresponding to the maximum target value in all iteration records are extracted as the optimal display window width and window level parameters.

[0039] In one embodiment of this application, step S40, generating the enhanced 3D data volume, includes:

[0040] Extract the optimal window width and optimal window level values ​​from the optimal display window width and window level parameters;

[0041] The lower and upper threshold values ​​of the linear mapping physical interval are calculated using the optimal window width and the optimal window level values.

[0042] The original physical density value of each voxel in the chest CT three-dimensional image data is extracted by iterative process.

[0043] When the original physical density value is less than or equal to the lower limit threshold, the mapped gray value is set to the minimum gray constant.

[0044] When the original physical density value is greater than or equal to the upper limit threshold, the mapped gray value is set to the maximum gray constant;

[0045] When the original physical density value is between the lower threshold and the upper threshold, the corresponding mapped gray value is calculated using a normalization function.

[0046] The enhanced 3D data volume is generated by combining all the mapped grayscale values.

[0047] In one embodiment of this application, step S50, outputting the lung cancer lymph node partitioning results corresponding to the chest CT three-dimensional image data, includes:

[0048] The chest CT three-dimensional image data are subjected to global normalization processing to generate a standard reference data volume;

[0049] The standard reference data volume is used as the first channel matrix, and the enhanced 3D data volume is used as the second channel matrix. They are fused and stitched on the channel axis of the deep learning framework to generate the dual-channel tensor data.

[0050] The dual-channel tensor data is input into the downsampling feature extraction layer of the three-dimensional spatial partitioning model to generate a multi-scale semantic feature map;

[0051] The upsampling decoding layer of the three-dimensional spatial partitioning model is used to perform spatial analysis calculations on the multi-scale semantic feature map to generate probability prediction tensors for each category;

[0052] The maximum probability index of each category probability prediction tensor is extracted to generate the lung cancer lymph node partitioning results.

[0053] In one embodiment of this application, before extracting the set of CT values ​​of all voxels in the mediastinal region of the chest CT three-dimensional image data in step S10, the method further includes:

[0054] Receives the initial format sequence of medical slide files output by the medical scanning equipment;

[0055] Parse the spatial metadata of the initial format sequence medical slide file, the spatial metadata including pixel spacing and layer spacing parameters;

[0056] Using a three-dimensional cubic spline interpolation algorithm, the initial format sequence medical slide file is resampled into isotropic spatial resolution tensor data based on the spatial metadata;

[0057] An anisotropic diffusion filtering algorithm based on partial differential equations is used to perform edge-preserving denoising on the isotropic spatial resolution tensor data to generate smooth reconstructed tensor data.

[0058] The mediastinal anatomical boundary bounding box is cropped from the smoothed reconstructed tensor data using a region growing extraction algorithm to generate the chest CT three-dimensional image data.

[0059] In one embodiment of this application, after generating the lung cancer lymph node partitioning results, the method further includes performing morphological correction calculations on the lung cancer lymph node partitioning results:

[0060] Locate all independent positively predicted connected components in the lung cancer lymph node partitioning results;

[0061] Calculate the three-dimensional spatial distance data for each independent positively predicted connected component, wherein the three-dimensional spatial distance data represents the geometric distance between the independent positively predicted connected component and the preset tracheal centerline;

[0062] Determine whether the three-dimensional spatial distance data is greater than the set anatomical constraint distance threshold;

[0063] When the three-dimensional spatial distance data is greater than the anatomical constraint distance threshold, the corresponding independent positive prediction connected region is marked as an anatomical boundary false positive region;

[0064] The label data of the anatomically crossed false positive regions is deleted from the lung cancer lymph node partitioning results, and the final lung cancer lymph node partitioning results after topology correction are generated and output.

[0065] To achieve the above objectives, a second aspect of this application proposes a lung cancer lymph node partitioning system based on three-dimensional anatomy, which is implemented using a three-dimensional anatomy-based lung cancer lymph node partitioning method.

[0066] The beneficial effects of the lung cancer lymph node partitioning system and method based on three-dimensional anatomy according to this application are as follows:

[0067] 1. To address the issue of high density overlap in soft tissues during plain CT scans, this paper introduces a three-dimensional Hessian matrix for extracting tubular structure features, building upon the Gaussian mixture model for extracting mixed soft tissue regions. This system and method objectively separate vascular regions from candidate lymph node regions in terms of geometric topology through the calculation of spatial second-order partial derivatives and their eigenvalues, overcoming the physical limitations of single-density classification and providing an accurate anatomical benchmark for subsequent optimization of display parameters.

[0068] 2. Abandoning fixed empirical display parameters for the entire population, a target function was constructed with contrast-to-noise ratio and regional information entropy as variables. An adaptive optimization algorithm was then used to solve for the optimal display window width and window level parameters. This optimization mechanism, while utilizing regional information entropy as a negative constraint to avoid loss of anatomical texture, maximizes the three-dimensional spatial gradient contrast between the candidate lymph node region and the surrounding mediastinal fat and vascular regions. This achieves optimal image mapping enhancement tailored to the individual patient's chest three-dimensional tissue density distribution characteristics.

[0069] 3. The standard baseline data volume and the enhanced 3D data volume are concatenated to generate dual-channel tensor data, which is then input into a 3D spatial partitioning model for processing. This dual-channel architecture preserves the absolute physical attenuation baseline of the original chest CT 3D image data while injecting spatial attention features to enhance boundary contrast. This avoids data distribution shifts and feature extraction confusion caused by single mapping enhancement at the underlying data structure level. Combined with anatomical constraint mechanisms such as judging the 3D spatial distance data of the predicted connected components, it can automatically identify and eliminate false positive regions that cross anatomical boundaries at the 3D spatial topology level, ensuring that the output lung cancer lymph node partitioning results accurately conform to the true objective anatomical distribution of the mediastinum.

[0070] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0071] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0072] Figure 1 This is a flowchart illustrating the overall process of the lung cancer lymph node partitioning method based on three-dimensional anatomy proposed in this application.

[0073] Figure 2 This is a flowchart of step S20 of the lung cancer lymph node partitioning method based on three-dimensional anatomy in this application;

[0074] Figure 3 This is a flowchart of step S30 of the lung cancer lymph node partitioning method based on three-dimensional anatomy in this application. Detailed Implementation

[0075] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0076] The following description, in conjunction with the accompanying drawings, illustrates a lung cancer lymph node partitioning system and method based on three-dimensional anatomy, according to embodiments of this application.

[0077] like Figure 1 As shown, the lung cancer lymph node partitioning system and method based on three-dimensional anatomy in this application embodiment may include:

[0078] Step S10: Receive chest CT three-dimensional image data, extract the CT value set of all voxels in the mediastinal region of the chest CT three-dimensional image data, and construct an individualized three-dimensional tissue density distribution histogram.

[0079] It should be noted that the chest CT three-dimensional image data mentioned in this step is isotropic spatial resolution volume data obtained after preprocessing the original medical images. Its data format is a three-dimensional matrix, in which each element corresponds to a voxel, and the element value is the CT value of that voxel.

[0080] The three-dimensional tissue density distribution histogram is a statistical histogram constructed with CT value on the horizontal axis and the voxel frequency (i.e., number of voxels) with that CT value in the mediastinal region on the vertical axis. This three-dimensional tissue density distribution histogram reflects the distribution of tissues with different densities in the mediastinal region.

[0081] For example, mediastinal adipose tissue is typically distributed in the CT value range of -100 HU to -50 HU, soft tissue (including lymph nodes, blood vessels, and muscles) in the range of +30 HU to +50 HU, and bone or calcification in the range of +200 HU and above. Constructing the individualized three-dimensional tissue density distribution histogram allows subsequent processing to reflect the actual tissue composition characteristics of the patient's mediastinal region, avoiding biases caused by fixed empirical values ​​for a group.

[0082] In one embodiment of this application, before extracting the set of CT values ​​of all voxels in the mediastinal region of the chest CT three-dimensional image data in step S10, the method further includes:

[0083] Receives the initial format sequence of medical slide files output by the medical scanning equipment;

[0084] The initial format sequence medical slice file is a CT scan output file sequence that conforms to the DICOM (Digital Imaging and Communication in Medicine) standard protocol. Each slice file corresponds to a two-dimensional tomographic image and its associated header file information.

[0085] Parse the spatial metadata of the initial format sequence medical slide file, the spatial metadata including pixel spacing and layer spacing parameters;

[0086] The pixel spacing refers to the actual physical size (in millimeters) represented by a single pixel in a two-dimensional tomographic image, and the layer spacing refers to the actual physical distance (in millimeters) between two adjacent slices.

[0087] Since the original CT slices usually have inconsistent inter-slice resolution (determined by the inter-slice spacing) and intra-slice resolution (determined by the pixel spacing), that is, the physical size of voxels is different in the three spatial dimensions, direct use will lead to geometric distortion in subsequent three-dimensional spatial feature extraction.

[0088] Using a three-dimensional cubic spline interpolation algorithm, the initial format sequence medical slide file is resampled into isotropic spatial resolution tensor data based on the spatial metadata;

[0089] The three-dimensional cubic spline interpolation algorithm constructs cubic spline functions in three dimensions, resampling the original non-isotropic voxel mesh into an isotropic mesh with the same physical size in all three dimensions, so that subsequent three-dimensional spatial operations (such as three-dimensional Hessian matrix calculation and three-dimensional connected component labeling) have the correct spatial geometric meaning.

[0090] An anisotropic diffusion filtering algorithm based on partial differential equations is used to perform edge-preserving denoising on the isotropic spatial resolution tensor data to generate smooth reconstructed tensor data.

[0091] The mediastinal anatomical boundary bounding box is cropped from the smoothed reconstructed tensor data using a region growing extraction algorithm to generate the chest CT three-dimensional image data.

[0092] The anisotropic diffusion filtering (such as the Perona-Malik diffusion model) is a nonlinear filtering method. Its core idea is to perform strong diffusion in regions with small image gradients (uniform regions) to suppress noise, and perform weak diffusion or stop diffusion in regions with large image gradients (edge ​​regions) to preserve edge structure.

[0093] This processing removes quantum and electronic noise from the original CT images while preserving the anatomical boundaries between adipose tissue, lymph nodes, and blood vessels within the mediastinal region. The region growth extraction algorithm starts with a seed point in the central region of the mediastinum and grows outwards gradually based on the similarity criterion of adjacent voxel CT values. It stops growing at the junction of the mediastinum and lung tissue or chest wall, thus obtaining a minimal three-dimensional bounding box containing the complete mediastinal structure.

[0094] This process reduces the amount of data required for subsequent calculations, while removing interference from irrelevant lung tissue and chest wall structures on the statistics of the three-dimensional tissue density distribution histogram.

[0095] like Figure 2As shown, in step S20, the three-dimensional tissue density distribution histogram is fitted and decomposed using a Gaussian mixture model to generate tissue density mixing parameters.

[0096] In one embodiment of this application, step S20, generating tissue density mixing parameters, includes:

[0097] Extract the frequency data from the three-dimensional tissue density distribution histogram;

[0098] The Gaussian mixture model is initialized with model parameters including the mean vector and covariance matrix;

[0099] The frequency data is input into the Gaussian mixture model to perform expectation maximization iterative processing.

[0100] Calculate the posterior probability matrix of each frequency data point belonging to each Gaussian component;

[0101] The updated parameters are generated by updating the mean vector and the covariance matrix using the posterior probability matrix.

[0102] When the change in the log-likelihood function of the Gaussian mixture model is less than the convergence threshold, the expectation maximization iteration process is stopped, and the updated parameters are output as the tissue density mixture parameters, which include a first density interval and a second density interval.

[0103] The convergence threshold ranges from 1 × 10⁻⁶. -6 Up to 1×10 -4 Preferably 1×10 -6 This threshold is used to ensure that the model parameters have converged sufficiently to a stable solution during the iteration process.

[0104] In the GMM fitting process described above, the frequency data refers to the voxel frequency corresponding to each CT value in the three-dimensional tissue density distribution histogram, forming the data sample set to be fitted. During initialization, the number of components in the Gaussian mixture model is set to three, corresponding to adipose tissue, soft tissue, and bone / calcification, respectively. Specifically, the K-means clustering algorithm can be used to initially cluster the sample data of the three-dimensional tissue density distribution histogram. The CT values ​​of the three cluster centers are used as the initial values ​​of the mean vectors of each Gaussian component, and the variance of the CT values ​​within each cluster is used as the initial value of the covariance matrix.

[0105] Understandably, K-means clustering is a well-known unsupervised clustering algorithm used to provide reasonable initial parameters for the EM algorithm, preventing it from getting trapped in local optima and reducing the number of iterations. Specifically, during clustering, the number of cluster centers is set to three, corresponding to the initial center positions of the density peaks in adipose tissue, soft tissue, and bone / calcification, respectively.

[0106] The expectation-maximization algorithm comprises two alternating steps: the expectation step calculates the posterior probability that each CT value sample belongs to the fat component, soft tissue component, and bone / calcification component based on the current model parameters; the maximization step uses the calculated posterior probabilities to perform a weighted maximum likelihood estimate of the model parameters, updating the mean, variance, and weights of each component. The log-likelihood function is used to measure the model's fit to the data; when its change is less than a preset threshold (e.g., 10⁻⁶), it indicates that the model has converged.

[0107] The first density interval corresponds to the CT value range covered by the fat density peak component, and the second density interval corresponds to the CT value range covered by the soft tissue density peak component.

[0108] It should be noted that the first density range is preferably -150HU to -10HU, and the second density range is preferably +10HU to +80HU. The above ranges are set based on the typical CT value distribution range of adipose tissue and soft tissue in the mediastinal region. Specifically, they can be calculated by adding or subtracting 1.96 times the standard deviation from the mean parameter of each component output by the Gaussian mixture model, so as to cover about 95% of the confidence interval of each tissue density peak.

[0109] The mediastinal fat region and soft tissue mixed region are extracted based on the tissue density mixing parameters; the tubular structure features of the soft tissue mixed region are extracted using the three-dimensional Hessian matrix to generate vascular region and lymph node candidate region;

[0110] The purpose of this step is to separate tissues of different densities within the mediastinal region and further distinguish blood vessels and lymph nodes based on their geometric morphology within the soft tissue.

[0111] The Gaussian Mixture Model is a probabilistic model used to describe data distribution. It assumes that the data is generated by a weighted combination of multiple Gaussian distributions (i.e., normal distributions), with each Gaussian distribution corresponding to a CT value distribution for a tissue type.

[0112] The three-dimensional tissue density distribution histogram of the mediastinal region is essentially a superposition of the CT value distributions of three tissue types: adipose tissue, soft tissue, and bone / calcification. Therefore, three Gaussian components are used to fit the three-dimensional tissue density distribution histogram. The mean (representing the typical CT value of the tissue) and variance (representing the fluctuation range of the CT value of the tissue) of each Gaussian component are estimated iteratively through the Expectation-Maximization algorithm.

[0113] After fitting, components with smaller mean values ​​correspond to adipose tissue, components with medium mean values ​​correspond to soft tissue (including lymph nodes, blood vessels, and muscles), and components with larger mean values ​​correspond to bone or calcification. Density intervals are set based on the mean parameters of the adipose and soft tissue components. Voxels whose CT values ​​fall within the adipose density interval are extracted as mediastinal adipose regions, and voxels whose CT values ​​fall within the soft tissue density interval are extracted as mixed soft tissue regions.

[0114] However, the soft tissue mixed area contains lymph nodes, blood vessels and muscles, and the CT values ​​of these three are highly overlapping, making it impossible to further distinguish them using one-dimensional density information.

[0115] Therefore, this step introduces a three-dimensional Hessian matrix for tubular structure feature extraction. The three-dimensional Hessian matrix is ​​a 3×3 symmetric matrix composed of the second-order partial derivatives of the voxel grayscale function in three-dimensional space, and its mathematical form is: ;

[0116] In the formula, The grayscale value (i.e., CT value) of the voxel in the soft tissue mixed region. , , denoted as the three-dimensional spatial coordinates of the voxel.

[0117] The three eigenvalues ​​of the matrix (denoted as...) , , (Sorted by absolute value) reflects the gray curvature changes of the local neighborhood of a voxel in different directions.

[0118] Specifically, the three feature values , , By solving the characteristic equation We obtain the formula, is the identity matrix. In the formula... The matrix is ​​a 3×3 identity matrix. Substituting these three eigenvalues ​​into the tubular response function R quantifies the degree to which each voxel belongs to a tubular structure. Subsequent steps, based on the above mathematical principles, will calculate the second-order partial derivatives of each voxel using the finite difference method and construct the Hessian matrix, thereby calculating the response value R and ultimately achieving the separation of the vascular region from the lymph node candidate region.

[0119] Specifically, the regions where blood vessels are generated and the candidate lymph nodes include:

[0120] Extract the set of three-dimensional spatial coordinates of all voxels in the soft tissue mixed region;

[0121] Calculate the second-order partial derivatives of the voxels corresponding to the three-dimensional spatial coordinate set, and construct the three-dimensional Hessian matrix for each voxel;

[0122] Calculate the three eigenvalues ​​of the three-dimensional Hessian matrix for each voxel, and calculate the response value of the tubular structure feature based on the numerical relationship and sign attribute of the three eigenvalues.

[0123] Set a vascular response threshold, extract voxels whose response values ​​are greater than or equal to the vascular response threshold, and generate the vascular region;

[0124] Extract voxels whose response values ​​are less than the vascular response threshold to generate non-vascular residual regions;

[0125] A three-dimensional connected component labeling algorithm is applied to the non-vascular remaining region to extract connected components with a volume greater than a set volume threshold and a sphericity greater than a set sphericity threshold, and these components are combined to generate the lymph node candidate region.

[0126] As one possibility, the vascular response threshold ranges from 0.3 to 0.7, preferably 0.5. The volume threshold ranges from 3 mm³ to 10 mm³, preferably 5 mm³, and this volume threshold is used to exclude noise voxel clusters that are too small.

[0127] The sphericity threshold value ranges from 0.5 to 0.8, preferably 0.6, and the formula for calculating sphericity is as follows:

[0128] ;

[0129] in, Let the volume of the connected component be . This represents the surface area of ​​the connected domain. For an ideal sphere, this value is 1. A higher value indicates that the shape is closer to a sphere.

[0130] The above thresholds are set based on the fact that lymph nodes in CT images are usually spherical soft tissue nodules with a volume of more than 5 mm³, while blood vessels, although spherical in cross-section, extend in tubular form in three-dimensional space. By filtering by both volume and sphericity, the cross-section of blood vessels and noise points can be effectively excluded.

[0131] In the above Hessian matrix tubular feature extraction process, the second-order partial derivatives of discrete voxels in a 3D digital image are approximated using the finite difference method. For example, for voxels... its along The second-order partial derivatives in the direction can be obtained from Approximately, the mixed partial derivatives are calculated from the corresponding difference combination.

[0132] After constructing the Hessian matrix, solve for its eigenvalues ​​and then... , , Sort. The tubular response function R is defined as:

[0133] ;

[0134] In the formula, , , Preset parameters for sensitivity are selected to control tubular features. Response value. The higher the value, the greater the probability that the voxel belongs to a tubular structure.

[0135] , , The sensitivity values ​​used to control the detection of tubular structures are as follows: The preferred values ​​are α=0.5, β=0.5, and γ=1.0;

[0136] The above parameters are set based on the classic parameter configuration of the Frangi vascular enhancement filter and combined with the statistical characteristics of the vascular diameter in the mediastinal region. Specifically, they can be obtained by optimizing on the labeled dataset through grid search.

[0137] For ideal tubular structures like blood vessels, the three eigenvalues ​​satisfy... ≈ ≪0 and ≈0, its grayscale changes drastically in the cross-sectional direction (corresponding to a large negative feature value), and its grayscale changes gently along the lumen direction (corresponding to a feature value close to zero), at which point R→1;

[0138] For spherical structures like lymph nodes, the three characteristic values ​​satisfy... ≈ ≈ The grayscale changes in the three orthogonal directions are all quite drastic and the amplitudes are similar (all three feature values ​​are negative and the absolute values ​​are similar), at which point R→0;

[0139] After setting the vascular response threshold, Voxels with values ​​greater than or equal to this threshold are extracted as vascular regions. Voxels with values ​​less than the threshold are extracted as non-vascular residual regions. The three-dimensional connected component labeling algorithm scans the voxel space, grouping adjacent (e.g., 26-neighborhood connected) voxels with the same category label into the same connected component. Sphericity is defined as the ratio of the surface area of ​​a sphere with the same volume as the connected component to the actual surface area of ​​the connected component; spheroidal lymph nodes have a higher sphericity than irregular muscle tissue. Through dual screening using volume and sphericity, noise points with excessively small volumes and irregularly shaped muscle fiber bundles are eliminated, thereby accurately locating the candidate lymph node regions.

[0140] For example, in the three-dimensional chest CT image data of a lung cancer patient, there is a suspected metastatic lymph node located on the right side of the trachea and behind the superior vena cava. Because its CT value falls within the soft tissue density range (approximately +35HU) and its shape is spherical (all three Hessian matrix eigenvalues ​​are negative and their absolute values ​​are similar) rather than tubular, it was successfully included in the candidate lymph node region. The surrounding superior vena cava, due to its tubular geometry (one eigenvalue is close to zero and the other two are negative), was correctly classified into the vascular region.

[0141] like Figure 3 As shown, in step S30, the contrast-to-noise ratio and regional information entropy of the candidate lymph node region, the mediastinal fat region, and the vascular region are calculated, and an objective function with the contrast-to-noise ratio and the regional information entropy as variables is constructed.

[0142] An adaptive optimization algorithm is used to search for the combination of parameters that maximizes the objective function, which is then used as the optimal display window width and window level parameters.

[0143] It should be noted that the purpose of this step is to use the identified candidate lymph node regions, mediastinal fat regions, and vascular regions as anatomical anchor points to automatically search for individualized window width and window level parameters that maximize the contrast between lymph nodes and surrounding tissues.

[0144] The contrast-to-noise ratio (CNR) is a quantitative indicator that measures the contrast between the target region and the background region in an image relative to the background noise. It is calculated by dividing the absolute value of the difference between the average gray level of the target region and the average gray level of the background region by the standard deviation of the background noise.

[0145] In this application, the contrast-to-noise ratio (CNR) comprehensively reflects the grayscale differences between the candidate lymph node region and the mediastinal fat region, as well as between the candidate lymph node region and the vascular region. A higher CNR value indicates a clearer boundary between the lymph node and the surrounding structures. The regional information entropy is a statistic describing the degree of disorder in the grayscale distribution of an image; the more uniform the grayscale distribution, the lower the entropy value, and the more dispersed the grayscale distribution, the higher the entropy value.

[0146] The construction logic of the objective function is as follows: the contrast-to-noise ratio (CNR) is used as a positive optimization term, that is, the larger the CNR value, the larger the objective function value, guiding the optimization algorithm to search for window parameters that maximize the contrast of lymph node-fat and lymph node-blood vessel boundaries; at the same time, the regional information entropy is used as a negative constraint term, and when the regional information entropy exceeds the preset entropy threshold, the objective function value is reduced through a penalty function.

[0147] The physical significance of adding regional information entropy constraints is that when the window width is too narrow, although the boundary contrast may be high, a large amount of tissue information is truncated and lost, the image grayscale distribution is too concentrated (entropy value is too low), and clinicians cannot observe the complete morphology of lymph nodes and the relationship with surrounding structures; when the window width is too wide, the grayscale distribution is too dispersed (entropy value is too high), and the boundary contrast is diluted.

[0148] Therefore, the objective function seeks the optimal balance between maximizing contrast and ensuring information integrity.

[0149] The adaptive optimization algorithm employs Bayesian optimization, where the analytical expression of the objective function is unknown (i.e., the analytical solution for the maximum value cannot be directly obtained), and each evaluation of the objective function value requires performing a complete image mapping and CNR calculation, resulting in high computational costs. In contrast, Bayesian optimization approximates the true objective function by constructing a Gaussian process surrogate model. It utilizes the sampling function to guide sampling in the most promising region of the parameter space. Compared to brute-force grid search or random search, it can approach the global optimum in fewer iterations, improving computational efficiency.

[0150] Specifically, constructing an objective function with the contrast-to-noise ratio and the regional information entropy as variables includes:

[0151] The test display parameters are obtained and mapped onto the chest CT three-dimensional image data to obtain a simulated image matrix;

[0152] The first mean of the lymph node candidate region, the second mean of the mediastinal fat region, and the third mean of the vascular region are extracted from the simulated image matrix.

[0153] Calculate the background noise standard deviation, and calculate the contrast noise ratio based on the first mean, the second mean, the third mean, and the background noise standard deviation;

[0154] Calculate the probability distribution data of pixels within the simulated image matrix to obtain the region information entropy;

[0155] The contrast-to-noise ratio is used as a positive optimization term, and the difference between the regional information entropy and the entropy threshold is transformed by a penalty function and used as a negative constraint term. The objective function is generated by merging these terms.

[0156] The test display parameters refer to the window width and window level parameter combinations tried during the optimization process, i.e., the candidate parameters selected by the Bayesian optimization algorithm based on the acquisition function in each iteration. A linear mapping from density values ​​to grayscale values ​​is performed on the original 3D CT data using this set of parameters to obtain the simulated image matrix used to calculate CNR and regional information entropy.

[0157] The background noise standard deviation is typically measured over a homogeneous region of the image (such as within the mediastinal fat region) to reflect the noise level of the image. Specifically, at least three regions of interest (each region of interest is 5mm×5mm×5mm) are randomly selected within the mediastinal fat region, and the average of the standard deviations of the voxel gray values ​​within each region of interest is calculated as the standard deviation of the background noise.

[0158] The specific formula for calculating the contrast-to-noise ratio is as follows:

[0159] ;

[0160] In the formula, It is the first mean of the lymph node candidate region (i.e., the average value of the gray values ​​of all voxels within the lymph node candidate region). It is the second mean of the mediastinal fat region (i.e., the average value of the gray values ​​of all voxels in the mediastinal fat region). This is the third mean value of the vascular region (i.e., the average value of the gray values ​​of all voxels within the vascular region). The standard deviation of the background noise is given. The larger the CNR value, the greater the grayscale difference between the lymph node and the surrounding mediastinal fat and blood vessels, and the clearer and more distinguishable the boundary.

[0161] The regional information entropy The calculation formula is:

[0162] ;

[0163] In the formula, The grayscale value in the simulated image matrix is ​​the first... The probability of a pixel at a certain gray level (i.e., the proportion of pixels with that gray value to the total number of pixels). This represents the total number of gray levels (e.g., 256 levels). The objective function... The general form is:

[0164] ;

[0165] In the formula, For window width parameters, For window position parameters, This is the regularization coefficient, used to balance the weights of the contrast-to-noise ratio and information entropy constraint in the objective function. The preset entropy threshold, This indicates that a positive penalty term is generated when the regional information entropy exceeds the preset entropy threshold.

[0166] The regularization coefficient The value ranges from 0.1 to 10, preferably 1.0, and it is used to balance the weight relationship between the contrast-to-noise ratio and the regional information entropy constraint. The larger the value, the higher the proportion of the penalty term in the objective function, and the more the algorithm tends to choose the window parameter with lower information entropy.

[0167] The preset entropy threshold The value ranges from 4.0 to 6.0, preferably 5.5. This threshold corresponds to the empirical critical value for the transition of the gray-level distribution in the simulated image matrix from an "information-rich and not overly dispersed" state to an "overly dispersed" state. The above parameters can be obtained by performing grid search optimization on the validation set.

[0168] The objective function is designed to ensure that the optimization process maximizes boundary contrast without excessively sacrificing the overall information integrity of the image.

[0169] Specifically, in this step, an adaptive optimization algorithm is used to search for the parameter combination that maximizes the objective function, which serves as the optimal display window width and window level parameters, including:

[0170] Within a preset parameter search space, select multiple sets of initial test parameter combinations, and input the initial test parameter combinations into the objective function to calculate the corresponding set of initial target values;

[0171] The preset parameter search space refers to the search range of window width and window position. For example, the window width search range is 100HU to 600HU, and the window position search range is -50HU to +100HU. This range covers the conventional value range of longitudinal diaphragm window parameters.

[0172] A Gaussian process regression model is constructed using the initial test parameter combination and the initial target value set;

[0173] Before constructing a Gaussian process regression model, a set of initial sample points needs to be collected within the parameter search space to serve as the initial training data for the Gaussian process regression model. Initial sampling typically employs Latin hypercube sampling or random sampling to achieve uniform coverage of the search space.

[0174] Latin hypercube sampling is a stratified sampling method that divides the search space into multiple intervals in each dimension and selects one point in each interval to ensure that the sample points are evenly distributed throughout the search space. Specifically, the window width search range [100HU, 600HU] and the window level search range [-50HU, +100HU] are each divided into 10 equally spaced intervals. This generates 10 × 10 = 100 sets of initial test parameter combinations in both dimensions. Each set of parameters is substituted into the objective function to calculate the corresponding initial target value, forming the initial training dataset. Gaussian process regression is a non-parametric Bayesian regression method that assumes the objective function follows a multidimensional Gaussian distribution. Using the initial sampling points as training data, a kernel function (usually a Matern kernel or a squared exponential kernel) is used to calculate the covariance matrix between each sampling point, thus providing the predicted mean (reflecting the expected value) and predicted variance (reflecting uncertainty) of the objective function at the unsampled parameter locations.

[0175] Specifically, the mathematical form of the Matern kernel function is as follows:

[0176] ;

[0177] in For input parameters, For length scale parameters, For smoothness parameters, For the Gamma function, This is a modified Bessel function of the second kind. The mathematical form of the squared exponent kernel is:

[0178] ;

[0179] In the formula This is the signal variance. This application preferably uses the Matern kernel (…). To achieve a balance between smoothness and computational efficiency, the hyperparameters of the kernel function (such as length scale) are used. Signal variance Noise variance It is obtained by performing maximum likelihood estimation on the training data.

[0180] Based on the predicted mean and predicted variance output by the Gaussian process regression model, a confidence upper bound acquisition function is constructed.

[0181] The upper confidence bound (UCB) collection function takes the following form:

[0182] ;

[0183] In the formula, The predicted mean of the Gaussian process regression model is given. To predict the standard deviation, K is an adjustment coefficient that balances exploration (selecting regions with large predicted variance) and utilization (selecting regions with large predicted mean), where the adjustment coefficient... The value ranges from 1.0 to 5.0, with 2.0 being preferred.

[0184] In each round of optimization iteration, a candidate parameter combination that maximizes the confidence upper bound acquisition function is found and input into the objective function to calculate the current target value, and the Gaussian process regression model is updated.

[0185] In each iteration of the optimization process, the solution is obtained in a continuous parameter space. The maximum value will be the value obtained when the maximum value is reached. As candidate sampling points for this iteration, they are substituted into the objective function to obtain the true target value. Then, the new sampling points are added to the training dataset and the Gaussian process regression model is retrained to achieve the posterior update of the model.

[0186] When the number of iterations reaches the set maximum iteration threshold, the optimization iteration stops, and the parameters corresponding to the maximum target value in all iteration records are extracted as the optimal display window width and window level parameters.

[0187] For example, after 40 Bayesian optimization iterations, the objective function value reaches its peak and no longer increases significantly. At this point, the corresponding window width and window level parameters (such as window width) are... Window position This refers to the optimal display window width and window level parameters for that individual patient, compared to fixed empirical values ​​in existing technologies (such as septal window). This individualized parameter is better suited to the actual mediastinal fat content and vascular enhancement of this example patient.

[0188] Step S40: Apply the optimal display window width and window level parameters to the chest CT three-dimensional image data for density mapping to generate an enhanced three-dimensional data volume;

[0189] It should be noted that the purpose of this step is to apply the optimized individualized window width and window level parameters to the original CT value data in a linear mapping manner, generating a contrast-enhanced three-dimensional grayscale data volume. The essence of CT image display is to map the CT value of the voxel (unit: HU, typically ranging from -1000HU to +1000HU) to the grayscale value of the display device (typically 0 to 255).

[0190] This mapping relationship is determined by the window width and the window level: the window width determines the coverage range of the mapped CT values, and the window level determines the center position of that coverage range.

[0191] Specifically, half the window width minus the window level is used as the lower bound of the mapping (i.e., voxels with CT values ​​below this lower bound are all mapped to black), and half the window width plus the window level is used as the upper bound of the mapping (i.e., voxels with CT values ​​above this upper bound are all mapped to white). CT values ​​between the upper and lower bounds are linearly mapped to a grayscale range of 0 to 255. After this mapping, lymph nodes and adipose tissue, which originally had small differences in CT values, are significantly separated in grayscale, making the boundaries of lymph nodes clearly distinguishable both visually and numerically.

[0192] In one embodiment of this application, step S40, generating the enhanced 3D data volume, includes:

[0193] Extract the optimal window width and optimal window level values ​​from the optimal display window width and window level parameters;

[0194] The lower and upper threshold values ​​of the linear mapping physical interval are calculated using the optimal window width and the optimal window level values.

[0195] The original physical density value of each voxel in the chest CT three-dimensional image data is extracted by iterative process.

[0196] When the original physical density value is less than or equal to the lower limit threshold, the mapped gray value is set to the minimum gray constant.

[0197] When the original physical density value is greater than or equal to the upper limit threshold, the mapped gray value is set to the maximum gray constant;

[0198] When the original physical density value is between the lower threshold and the upper threshold, the corresponding mapped gray value is calculated using a normalization function.

[0199] The enhanced 3D data volume is generated by combining all the mapped grayscale values.

[0200] Specifically, let the optimal window width value be... The optimal window size is [value missing]. The lower threshold is The upper limit threshold is For the raw CT value of each voxel The mapping rules are as follows:

[0201] when At that time, output grayscale value (Black, i.e., the minimum grayscale constant);

[0202] when At that time, output grayscale value (White, i.e., the maximum grayscale constant);

[0203] when hour, ;

[0204] The gray values ​​obtained by calculating all voxels The original voxels are rearranged according to their spatial coordinates to form the enhanced 3D data volume with the same size as the original 3D data volume.

[0205] For example, using the optimal window width Window position After mapping the patient's original CT values, the lymph node voxels within the candidate lymph node region (original CT values ​​approximately...) The image was mapped to a grayscale value of approximately 155, while the surrounding mediastinal adipose tissue (original CT value approximately...) was... The data is mapped to a gray value of approximately 35, and the difference between the two is 120 gray levels, which is significantly greater than the difference when using fixed empirical values ​​(such as a window width of 350 HU and a window level of 50 HU). This provides clear 3D input data for subsequent partitioning models.

[0206] Step S50: The enhanced three-dimensional data volume is stitched together with the chest CT three-dimensional image data to generate dual-channel tensor data, which is then input into a three-dimensional spatial partitioning model for processing, and the lung cancer lymph node partitioning result corresponding to the chest CT three-dimensional image data is output.

[0207] It should be noted that this step involves deep fusion of the original CT value information and the enhanced contrast information, and uses a deep learning model to perform automated three-dimensional lymph node partitioning.

[0208] The enhanced 3D data volume provides high-contrast boundary information, which is helpful for the model to accurately locate the boundary between lymph nodes and surrounding tissues; however, the grayscale value loses the original CT density value information during the enhancement process (because different CT values ​​may be mapped to the same grayscale value). Therefore, the original chest CT 3D image data is used as a supplementary channel to provide the model with the original physical density numerical information.

[0209] The dual-channel tensor data is concatenated along the channel dimension of the deep learning framework, which preserves the original density information and provides enhanced contrast information, enabling the three-dimensional spatial partitioning model to utilize both types of features simultaneously for more accurate voxel-level classification prediction.

[0210] The three-dimensional spatial partitioning model is an end-to-end three-dimensional convolutional neural network. Its input is a two-channel three-dimensional tensor, and its output is a three-dimensional probability tensor with the same size as the input. Each voxel position outputs a probability vector, and the length of the vector is the total number of partition categories (e.g., 14 partition categories plus 1 background category as defined by the IASLC (International Association for the Study of Lung Cancer) lymph node partitioning standard).

[0211] The 3D spatial partitioning model adopts the UNet 3D variant architecture. This architecture consists of two 3D convolutional layers (3×3×3 kernel size, stride 1) followed by a 3D max-pooling layer (2×2×2 pooling window size, stride 2) at each downsampling stage of the encoder, for a total of four downsampling stages. The number of channels is as follows: 64 channels in the first stage, 128 channels in the second stage, 256 channels in the third stage, and 512 channels in the fourth stage. The decoder is symmetrical to the encoder. Each upsampling stage includes a 3D transposed convolutional layer (2×2×2 kernel size, stride 2) followed by two 3D convolutional layers (3×3×3 kernel size, stride 1). Skip-layer connections are used to concatenate and fuse the feature maps of the corresponding stages of the encoder along the channel dimension. All convolutional layers are followed by batch normalization layers and ReLU activation functions. This layer design and channel configuration are optimized structures verified through multiple experiments, ensuring segmentation accuracy while controlling the number of model parameters and avoiding overfitting.

[0212] The structure of the three-dimensional spatial partitioning model adopts an encoder-decoder architecture. The training process of the three-dimensional spatial partitioning model is as follows: obtain chest CT three-dimensional image data containing the mediastinal region and labeled with lymph node partitioning labels as the training sample set. The lymph node partitioning labels are manually labeled layer by layer on the original CT images according to the IASLC (International Association for the Study of Lung Cancer) lymph node partitioning standard by radiologists with more than five years of experience in chest imaging diagnosis.

[0213] During training, the chest CT 3D image data in the training sample set are processed sequentially through steps S10 to S40 to obtain corresponding dual-channel tensor data, which serves as the input to the 3D spatial partitioning model. Lymph node partition labels annotated by radiologists are used as supervision signals. A weighted combination of the Dice loss function and the cross-entropy loss function is used as the total loss function. The Adam optimizer is used with a preset learning rate (e.g., 1×10⁻⁶). -4 The model parameters are iteratively updated until the change in the total loss function value on the validation set is less than a preset threshold (e.g., the change over 10 consecutive training epochs is less than 1 × 10). -4 Training is stopped when the time is right, and the trained three-dimensional spatial partitioning model is obtained.

[0214] The encoder extracts high-dimensional semantic features step by step through 3D convolution and pooling operations, while the decoder restores the feature map to its original resolution through 3D deconvolution and skip connection operations, while preserving spatial detail information.

[0215] In the encoder path, the downsampling feature extraction layer consists of multiple 3D convolutional blocks and max pooling layers stacked alternately. Each convolutional block contains a 3D convolutional layer, a batch normalization layer, and a ReLU activation function.

[0216] In the decoder path, the upsampled decoding layer gradually restores the spatial resolution through three-dimensional transposed convolution, and combines skip connections to stitch and fuse the feature maps of the corresponding layers of the encoder in the channel dimension to supplement high-resolution spatial detail information.

[0217] After training with a large amount of 3D CT image data labeled with lymph node partitions, the 3D spatial partitioning model can automatically learn the mapping relationship from image features to partition categories.

[0218] Specifically, the training data comes from de-identified enhanced chest CT images from multiple medical centers, and the lymph node partition labels are manually annotated layer by layer on the original CT images by radiologists with more than five years of experience in chest imaging diagnosis, according to the IASLC lymph node partitioning standard.

[0219] During training, the training samples are processed sequentially through steps S10 to S40 to obtain dual-channel tensor data, which is then input into the three-dimensional spatial partitioning model. The lymph node partitioning labels annotated by radiologists are used as supervision signals. A combination of the Dice loss function and the weighted cross-entropy loss function is used as the total loss function. The Adam optimizer is used to iteratively update the model parameters at a preset learning rate until the change in the total loss function value on the validation set is less than a preset threshold, at which point training stops, and the trained three-dimensional spatial partitioning model is obtained.

[0220] Specifically, the output includes the lung cancer lymph node partitioning results corresponding to the chest CT three-dimensional image data, including:

[0221] The chest CT three-dimensional image data are subjected to global normalization processing to generate a standard reference data volume;

[0222] The standard reference data volume is used as the first channel matrix, and the enhanced 3D data volume is used as the second channel matrix. They are fused and stitched on the channel axis of the deep learning framework to generate the dual-channel tensor data.

[0223] The dual-channel tensor data is input into the downsampling feature extraction layer of the three-dimensional spatial partitioning model to generate a multi-scale semantic feature map;

[0224] The upsampling decoding layer of the three-dimensional spatial partitioning model is used to perform spatial analysis calculations on the multi-scale semantic feature map to generate probability prediction tensors for each category;

[0225] The maximum probability index of each category probability prediction tensor is extracted to generate the lung cancer lymph node partitioning results.

[0226] It is understandable that CT values ​​generated by different CT devices and different scanning parameters may have systematic biases. Global normalization maps CT values ​​to a standard range (such as a standard normal distribution with a mean of 0 and a variance of 1), eliminating the impact of differences between devices on model inference.

[0227] The channel axis fusion refers to stacking two three-dimensional matrices of the same size along the channel dimension to obtain a matrix of size [missing information]. The four-dimensional tensor, in which These are the height, width, and depth, respectively.

[0228] The downsampling feature extraction layer consists of multiple three-dimensional convolutional blocks and max pooling layers stacked alternately. Each convolutional block contains a convolutional layer, a batch normalization layer, and an activation function. By gradually reducing the spatial resolution and increasing the number of channels, semantic features at different levels of abstraction, from edge textures to organ structures, are extracted.

[0229] The upsampling decoding layer gradually restores spatial resolution through 3D deconvolution or interpolation upsampling operations, and incorporates feature maps from the corresponding encoder layers via skip connections to supplement high-resolution spatial information. The final output layer uses the Softmax activation function to convert the network output of each voxel into a probability value, resulting in a layer of size [size missing]. The probability prediction tensor This represents the total number of partition categories.

[0230] For each voxel, from The category index corresponding to the highest predicted probability value among the predicted probabilities of each category is selected as the final partition label for that voxel, resulting in a size of [missing value]. The label matrix, i.e., the three-dimensional partitioning result of lymph nodes.

[0231] Furthermore, after generating the lung cancer lymph node partitioning results, the method further includes performing morphological correction calculations on the lung cancer lymph node partitioning results:

[0232] Locate all independent positively predicted connected components in the lung cancer lymph node partitioning results;

[0233] It should be noted that the positive prediction connected region refers to a connected region composed of adjacent voxels, having the same partition label, and the label belongs to the lymph node region (not the background).

[0234] Calculate the three-dimensional spatial distance data for each independent positively predicted connected component, wherein the three-dimensional spatial distance data represents the geometric distance between the independent positively predicted connected component and the preset tracheal centerline;

[0235] Determine whether the three-dimensional spatial distance data is greater than the set anatomical constraint distance threshold;

[0236] When the three-dimensional spatial distance data is greater than the anatomical constraint distance threshold, the corresponding independent positive prediction connected region is marked as an anatomical boundary false positive region;

[0237] The label data of the anatomically crossed false positive regions is deleted from the lung cancer lymph node partitioning results, and the final lung cancer lymph node partitioning results after topology correction are generated and output.

[0238] Specifically, the pre-segmented tracheal tree structure is extracted, the three-dimensional spatial trajectory of the tracheal centerline is calculated, and then the shortest Euclidean distance from the centroid of each connected domain to the tracheal centerline is calculated.

[0239] The three-dimensional spatial trajectory of the tracheal centerline is extracted by performing a skeletonization algorithm on the tracheal tree structure. Specifically, a three-dimensional morphological thinning algorithm is used to iteratively erode the binary mask of the tracheal tree until a centerline skeleton of a single voxel width is obtained. Then, the vertices of each branch of the skeleton line are connected to form a complete centerline trajectory. The shortest Euclidean distance from the centroid to the tracheal centerline is calculated as follows: Let the set of three-dimensional spatial coordinates of all voxels within the connected domain be... ,in Let be the total number of voxels within the connected region, then the coordinates of the centroid of the connected region are:

[0240] ;

[0241] The trajectory of the tracheal centerline consists of a series of discrete points Composition, where T is the total number of discrete points on the centerline, then the three-dimensional spatial distance data

[0242] That is, the shortest Euclidean distance from the center of mass to the center line of the trachea.

[0243] Based on the anatomical distribution of mediastinal lymph nodes, each station's lymph nodes are distributed within a specific distance range around central structures such as the trachea, esophagus, and major blood vessels. Areas too far from the tracheal centerline should not anatomically contain lymph nodes. The anatomical constraint distance threshold ranges from 20mm to 40mm, preferably 30mm. This threshold is determined based on statistical analysis of the anatomical distribution of mediastinal lymph nodes: areas within 30mm of the tracheal centerline cover more than 95% of mediastinal lymph node locations. Areas beyond this distance mostly belong to the chest wall, peripheral lung fields, or subcutaneous tissue; positive lymph node predictions in these areas are usually due to model misjudgment.

[0244] The false positive area of ​​the anatomical overreach may be due to the three-dimensional spatial partitioning model misidentifying cross-sections of chest wall muscles, esophagus, or blood vessels as lymph nodes.

[0245] This post-processing step uses explicit prior knowledge of human anatomy to constrain the output of the three-dimensional spatial partitioning model, eliminating false positive predictions that do not conform to anatomical rules, and further improving the accuracy and clinical reliability of lymph node partitioning results.

[0246] For example, in the original output of the three-dimensional spatial partitioning model, there is a small connected region located on the left edge of the mediastinum, 3.5 cm away from the center line of the trachea. Since this distance exceeds the preset anatomical constraint distance threshold (e.g., 2.5 cm), the connected region is judged as an anatomical boundary false positive area and deleted. The corrected partitioning result is more accurate, retaining only the positive prediction area within 2.5 cm of the center line of the trachea.

[0247] This embodiment also discloses a lung cancer lymph node partitioning system based on three-dimensional anatomy, which is implemented using a lung cancer lymph node partitioning method based on three-dimensional anatomy.

[0248] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0249] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0250] 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 method for lung cancer lymph node partitioning based on three-dimensional anatomy, characterized in that, include: Step S10: Receive chest CT three-dimensional image data, extract the CT value set of all voxels in the mediastinal region of the chest CT three-dimensional image data, and construct an individualized three-dimensional tissue density distribution histogram. Step S20: Fit and decompose the three-dimensional tissue density distribution histogram using a Gaussian mixture model to generate tissue density mixing parameters; extract the mediastinal fat region and soft tissue mixing region based on the tissue density mixing parameters; extract tubular structure features from the soft tissue mixing region using a three-dimensional Hessian matrix to generate vascular region and lymph node candidate region. Step S30: Calculate the contrast-to-noise ratio and regional information entropy of the candidate lymph node region, the mediastinal fat region, and the vascular region; construct an objective function with the contrast-to-noise ratio and the regional information entropy as variables; use an adaptive optimization algorithm to search for the parameter combination that maximizes the objective function, which is used as the optimal display window width and window level parameters. Step S40: Apply the optimal display window width and window level parameters to the chest CT three-dimensional image data for density mapping to generate an enhanced three-dimensional data volume; Step S50: The enhanced three-dimensional data volume is stitched together with the chest CT three-dimensional image data to generate dual-channel tensor data, which is then input into a three-dimensional spatial partitioning model for processing, and the lung cancer lymph node partitioning result corresponding to the chest CT three-dimensional image data is output.

2. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S20, tissue density mixing parameters are generated, including: Extract the frequency data from the three-dimensional tissue density distribution histogram; The Gaussian mixture model is initialized with model parameters including the mean vector and covariance matrix; The frequency data is input into the Gaussian mixture model to perform expectation maximization iterative processing. Calculate the posterior probability matrix of each frequency data point belonging to each Gaussian component; The updated parameters are generated by updating the mean vector and the covariance matrix using the posterior probability matrix. When the change in the log-likelihood function of the Gaussian mixture model is less than the convergence threshold, the expectation maximization iteration process is stopped, and the updated parameters are output as the tissue density mixture parameters, which include a first density interval and a second density interval.

3. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S20, the vascular region and lymph node candidate region are generated, including: Extract the set of three-dimensional spatial coordinates of all voxels in the soft tissue mixed region; Calculate the second-order partial derivatives of the voxels corresponding to the three-dimensional spatial coordinate set, and construct the three-dimensional Hessian matrix for each voxel; Calculate the three eigenvalues ​​of the three-dimensional Hessian matrix for each voxel, and calculate the response value of the tubular structure feature based on the numerical relationship and sign attribute of the three eigenvalues. Set a vascular response threshold, extract voxels whose response values ​​are greater than or equal to the vascular response threshold, and generate the vascular region; Extract voxels whose response values ​​are less than the vascular response threshold to generate non-vascular residual regions; A three-dimensional connected component labeling algorithm is applied to the non-vascular remaining region to extract connected components with a volume greater than a set volume threshold and a sphericity greater than a set sphericity threshold, and these components are combined to generate the lymph node candidate region.

4. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S30, constructing an objective function with the contrast-to-noise ratio and the regional information entropy as variables includes: The test display parameters are obtained and mapped onto the chest CT three-dimensional image data to obtain a simulated image matrix; The first mean of the lymph node candidate region, the second mean of the mediastinal fat region, and the third mean of the vascular region are extracted from the simulated image matrix. Calculate the background noise standard deviation, and calculate the contrast noise ratio based on the first mean, the second mean, the third mean, and the background noise standard deviation; Calculate the probability distribution data of pixels within the simulated image matrix to obtain the region information entropy; The contrast-to-noise ratio is used as a positive optimization term, and the difference between the regional information entropy and the entropy threshold is transformed by a penalty function and used as a negative constraint term. The objective function is generated by merging these terms.

5. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S30, an adaptive optimization algorithm is used to search for the parameter combination that maximizes the objective function, which serves as the optimal display window width and window level parameters. This includes: Within a preset parameter search space, select multiple sets of initial test parameter combinations, and input the initial test parameter combinations into the objective function to calculate the corresponding set of initial target values; A Gaussian process regression model is constructed using the initial test parameter combination and the initial target value set; Based on the predicted mean and predicted variance output by the Gaussian process regression model, a confidence upper bound acquisition function is constructed. In each round of optimization iteration, a candidate parameter combination that maximizes the confidence upper bound acquisition function is found and input into the objective function to calculate the current target value, and the Gaussian process regression model is updated. When the number of iterations reaches the set maximum iteration threshold, the optimization iteration stops, and the parameters corresponding to the maximum target value in all iteration records are extracted as the optimal display window width and window level parameters.

6. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S40, the enhanced 3D data volume is generated, including: Extract the optimal window width and optimal window level values ​​from the optimal display window width and window level parameters; The lower and upper threshold values ​​of the linear mapping physical interval are calculated using the optimal window width and the optimal window level values. The original physical density value of each voxel in the chest CT three-dimensional image data is extracted by iterative process. When the original physical density value is less than or equal to the lower limit threshold, the mapped gray value is set to the minimum gray constant. When the original physical density value is greater than or equal to the upper limit threshold, the mapped gray value is set to the maximum gray constant; When the original physical density value is between the lower threshold and the upper threshold, the corresponding mapped gray value is calculated using a normalization function. The enhanced 3D data volume is generated by combining all the mapped grayscale values.

7. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, In step S50, the lung cancer lymph node partitioning results corresponding to the chest CT three-dimensional image data are output, including: The chest CT three-dimensional image data are subjected to global normalization processing to generate a standard reference data volume; The standard reference data volume is used as the first channel matrix, and the enhanced 3D data volume is used as the second channel matrix. They are fused and stitched on the channel axis of the deep learning framework to generate the dual-channel tensor data. The dual-channel tensor data is input into the downsampling feature extraction layer of the three-dimensional spatial partitioning model to generate a multi-scale semantic feature map; The upsampling decoding layer of the three-dimensional spatial partitioning model is used to perform spatial analysis calculations on the multi-scale semantic feature map to generate probability prediction tensors for each category; The maximum probability index of each category probability prediction tensor is extracted to generate the lung cancer lymph node partitioning results.

8. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 1, characterized in that, Before extracting the set of CT values ​​of all voxels in the mediastinal region of the chest CT three-dimensional image data in step S10, the method further includes: Receives the initial format sequence of medical slide files output by the medical scanning equipment; Parse the spatial metadata of the initial format sequence medical slide file, the spatial metadata including pixel spacing and layer spacing parameters; Using a three-dimensional cubic spline interpolation algorithm, the initial format sequence medical slide file is resampled into isotropic spatial resolution tensor data based on the spatial metadata; An anisotropic diffusion filtering algorithm based on partial differential equations is used to perform edge-preserving denoising on the isotropic spatial resolution tensor data to generate smooth reconstructed tensor data. The mediastinal anatomical boundary bounding box is cropped from the smoothed reconstructed tensor data using a region growing extraction algorithm to generate the chest CT three-dimensional image data.

9. The lung cancer lymph node partitioning method based on three-dimensional anatomy according to claim 7, characterized in that, After generating the lung cancer lymph node partitioning results, the method further includes performing morphological correction calculations on the lung cancer lymph node partitioning results: Locate all independent positively predicted connected components in the lung cancer lymph node partitioning results; Calculate the three-dimensional spatial distance data for each independent positively predicted connected component, wherein the three-dimensional spatial distance data represents the geometric distance between the independent positively predicted connected component and the preset tracheal centerline; Determine whether the three-dimensional spatial distance data is greater than the set anatomical constraint distance threshold; When the three-dimensional spatial distance data is greater than the anatomical constraint distance threshold, the corresponding independent positive prediction connected region is marked as an anatomical boundary false positive region; The label data of the anatomically crossed false positive regions is deleted from the lung cancer lymph node partitioning results, and the final lung cancer lymph node partitioning results after topology correction are generated and output.

10. A lung cancer lymph node partitioning system based on three-dimensional anatomy, characterized in that, It is implemented based on the lung cancer lymph node partitioning method based on three-dimensional anatomy as described in any one of claims 1-9.

Images