A method and system for electrical impedance state assessment with fusion of CT features

Abstract

The present application relates to the technical field of image processing, and more particularly to a method and system for evaluating electrical impedance state by fusing CT features. The method comprises the following steps: CT scanning of a subject to obtain CT scan data; generating an anatomical mask according to the CT scan data to obtain anatomical mask data; collecting three-dimensional impedance data by an electrical impedance scanning device; extracting multi-scale decomposition features according to the three-dimensional impedance data to obtain decomposition feature data; mapping an elastic harmonics network according to the decomposition feature data and the anatomical mask data to obtain a shape-calibrated impedance map data; extracting regional impedance features from the shape-calibrated impedance map data according to the anatomical mask data to obtain regional impedance feature data; and obtaining regional anomaly probability value data according to the regional impedance feature data. The present application fuses CT anatomical information and multi-scale electrical impedance features, realizes shape accurate alignment and regional anomaly recognition, and improves the spatial resolution of electrical impedance imaging.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for assessing electrical impedance status by fusing CT features. Background Technology

[0002] Electrical impedance tomography (ETM) is a non-invasive imaging method that reconstructs images based on the distribution characteristics of conductivity or resistivity within the body. Its basic principle involves placing multiple electrodes on the human body surface, injecting a weak alternating current into the body, and measuring the surface voltage response. By solving the inverse problem of conductivity distribution, an electrical impedance image of the tissue is reconstructed. As a non-invasive, low-cost, and radiation-free medical imaging method, ETM has significant application value in the detection of tissue abnormalities. However, its low imaging resolution makes it difficult to accurately represent complex anatomical structures. CT images are tomographic images that scan different tissue densities in the human body using the differences in X-ray penetration. CT generates high-resolution two-dimensional or three-dimensional tissue density maps through surround scanning and computational reconstruction, clearly displaying the structural features of bones, organs, and lesions. Therefore, existing technologies attempt to incorporate CT images for assisted registration and structural guidance, but most only use static registration methods, which are difficult to effectively address the positional differences between tissues, resulting in limited image fusion accuracy and affecting the accuracy of abnormality identification. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method and system for evaluating electrical impedance status by incorporating CT features, thereby resolving at least one of the aforementioned technical problems.

[0004] This application provides a method for assessing electrical impedance status by fusing CT features, the method comprising:

[0005] The subject is subjected to a CT scan to obtain CT scan data; an anatomical mask is generated based on the CT scan data to obtain anatomical mask data.

[0006] Three-dimensional impedance data is acquired using an electrical impedance scanning device; multi-scale decomposition feature extraction is performed based on the three-dimensional impedance data to obtain decomposed feature data; and elastic harmonic network mapping is performed based on the decomposed feature data and anatomical mask data to obtain configuration calibration impedance map data.

[0007] Based on the anatomical mask data, the regional impedance features of the configuration calibration impedance map data are extracted to obtain the regional impedance feature data;

[0008] The probability value of regional anomalies is calculated based on the regional impedance characteristic data to obtain the regional anomaly probability value data.

[0009] This invention achieves collaborative modeling of anatomical structures and electrophysiological states by fusing CT scan data and three-dimensional electrical impedance data; precise structural alignment of impedance images can be achieved through the generation and configuration calibration of anatomical masks, improving the fusion quality of cross-modal data; spatial resolution of impedance changes is enhanced through multi-scale feature decomposition and elastic harmonic network mapping; impedance-structure joint graphs are constructed by extracting regional impedance features and combining topological and density information, which helps to reveal potential local anomaly patterns; the calculation of regional anomaly probability introduces state template offset and functional consistency analysis mechanisms, thereby improving the sensitivity and interpretability of hidden anomalies.

[0010] Optionally, the CT scan includes:

[0011] The subject was subjected to CT scans under preset first radiation intensity conditions and preset second radiation intensity conditions, respectively, and first CT scan data and second CT scan data were obtained respectively.

[0012] Response difference maps were constructed from the first CT scan data and the second CT scan data to obtain response difference map data.

[0013] Structure-fidelity region estimation is performed on the response difference map data to obtain structure-fidelity region data;

[0014] Structural gradient harmonization is performed based on the structural fidelity region data to obtain CT scan data.

[0015] This invention effectively integrates tissue response characteristics under different radiation intensities by introducing imaging data under dual-dose conditions, enabling sensitive capture of structural details. The construction of dose difference maps significantly enhances the separation of high-frequency details from low-dose noise interference regions; structural fidelity region estimation helps achieve a better trade-off between structural integrity and radiation suppression; and structural gradient harmonization enables precise restoration of key anatomical boundaries, effectively improving blurring and artifact issues in low-dose imaging. This enhances the diagnostic value of CT images while reducing radiation risks to patients, balancing accuracy and safety.

[0016] Optionally, the structure-fidelity region estimation includes:

[0017] The dose intensity change rate was calculated from the response difference map data to obtain the response intensity change rate data;

[0018] Image gradient flow field data is obtained by calculating the image gradient flow field based on the response intensity change rate data and the response difference map data.

[0019] Phase perturbation entropy is calculated based on image gradient flow field data to obtain phase perturbation entropy data;

[0020] Based on the response intensity change rate data, image gradient flow field data, and phase perturbation entropy data, a non-convex consistent feature space projection is performed to obtain the structure-fidelity region data.

[0021] In this invention, the dose intensity change rate reflects the response differences under different radiation intensities, which helps identify structural regions sensitive to dose changes; the image gradient flow field reveals the continuity of image edges and textures, providing support for preserving structural boundaries; and the phase perturbation entropy measures the perturbation consistency within a region, effectively eliminating non-structural interference. By projecting non-convex consistency feature space and integrating multi-source structural credibility information, the fidelity region is accurately extracted, which not only improves the image structure restoration quality but also provides a stable structural support foundation for CT image enhancement and multimodal registration.

[0022] Optionally, the structure gradient harmonization includes:

[0023] Structural gradient field map data is generated based on structural fidelity region data.

[0024] Fidelity region structure diffusion weights are calculated from the structure gradient field map data to obtain fidelity region structure map data.

[0025] Structural gradient harmonic diffusion was performed on the fidelity region structure map data to obtain CT scan data.

[0026] This invention clarifies the directional variation characteristics of high-confidence structural regions by constructing structural gradient field maps. Then, based on the structural diffusion weight calculation of fidelity regions, the participation intensity of different regions in the harmonic diffusion process is dynamically adjusted, allowing true structural information to dominate. Through structural gradient harmonic diffusion, fidelity structural features are smoothly transferred to low-contrast or blurred regions, effectively suppressing artifact diffusion and restoring tissue texture continuity. This invention improves the edge sharpness and tissue identification of CT images, providing a higher-quality image foundation for downstream anatomical modeling and state analysis.

[0027] Optionally, the anatomical mask generation includes:

[0028] Multi-directional structural flow field maps are constructed based on CT scan data to obtain structural flow field map data;

[0029] directional consistency path extraction is performed on the structural flow field map data to obtain directional consistency path data;

[0030] Based on the directional consistency path data, a path aggregation domain graph is constructed to obtain the anatomical mask data.

[0031] This invention achieves high-precision identification and boundary preservation of anatomical structures through the construction of multi-directional structural flow field maps and the extraction of directional consistency paths. The structural flow field map fully depicts the spatial extension direction and local structural continuity between tissues in CT images, overcoming the failure problem of traditional edge detection in low-contrast regions. By extracting directional consistency paths, it is possible to capture the dominant structural trajectories with topological coherence between organs or tissues. The construction of path aggregation domain maps effectively aggregates highly consistent regions and suppresses isolated noise responses, thereby generating anatomical mask data with high structural reliability.

[0032] Optionally, the multi-scale decomposition feature extraction includes:

[0033] Multi-resolution spherical harmonic transformation and three-dimensional non-equidistant wavelet decomposition are performed on the three-dimensional impedance data to obtain the first impedance decomposition data and the second impedance decomposition data, respectively.

[0034] The contrast change rate data is obtained by calculating the multi-scale contrast change rate based on the first impedance decomposition data.

[0035] The local response maximum level is calculated based on the second impedance decomposition data to obtain the local response maximum level data;

[0036] Cross-scale trend calculations are performed based on contrast change rate data and local response maximum level data to obtain decomposed feature data.

[0037] This invention effectively enhances the detail representation and local response sensitivity of impedance data at different structural scales by combining multi-resolution spherical harmonic transform and three-dimensional non-equidistant wavelet decomposition. Spherical harmonic transform possesses excellent spherical structure modeling capabilities and is suitable for frequency domain analysis of global morphological differences, while non-equidistant wavelet decomposition can flexibly adapt to non-uniform sampling and local perturbations, thus more accurately capturing fine-grained changes in anomalous regions. Multi-scale contrast change rate calculation of the first impedance decomposition data quantifies the impedance contrast intensity at different scales; local response maximum level analysis of the second impedance decomposition data identifies key scale levels in anomalous response concentrations. Through cross-scale fusion processing of contrast change rate and local response maximum level, the system can effectively characterize the fine-grained trend of structural changes and the consistency of global scale responses in three-dimensional impedance data, overcoming the limitations of single-scale features in identifying tissue heterogeneity. The contrast change rate reflects the degree of impedance transition between the interior and edges of local regions, while the local response maximum level captures the dominant level position of impedance response at different scales. When the two are combined for trend calculation, decomposed feature data with multi-resolution response sensitivity and spatial stability can be generated, thereby improving the response saliency and partition resolution of abnormal tissue regions in the mapping and identification process.

[0038] Optionally, the resilient harmonic network mapping includes:

[0039] Functional-anatomical dual-image processing is performed based on decomposed feature data and anatomical mask data to obtain functional-anatomical dual-image data.

[0040] Elastic harmonic configuration mapping is performed based on functional-anatomical dual-map data to obtain configuration mapping data;

[0041] Based on the configuration mapping data, a topology-conserving configuration transformation is performed to obtain the configuration calibration impedance map data.

[0042] This invention achieves high-precision spatial alignment and morphological consistency fusion between electrical impedance features and CT anatomical structures by constructing a functional-anatomical dual-map model and introducing elastic mapping and topological conservation. The functional-anatomical dual-map processing fully integrates the response patterns of decomposed impedance features with the structural distribution of the CT mask, constructing a set of nested representations of the map domain that are correlated between functional response and anatomical structure, enhancing the understanding of the correlation before mapping. Elastic harmonic configuration mapping, while preserving the matching accuracy of key regions, allows non-rigid structures to flexibly align in confined spaces, effectively improving the local fidelity of heterogeneous data fusion. Topological conservation configuration transformation ensures that structural continuity and connectivity are not disrupted, preventing feature mismatch or geometric distortion caused by deformation.

[0043] Optionally, the regional impedance feature extraction includes:

[0044] Impedance voxel clustering was performed on the configuration calibration impedance map data based on the anatomical mask data to obtain impedance voxel clustering data;

[0045] Topological feature extraction and local density feature extraction are performed on impedance voxel clustering data to obtain topological feature data and local density feature data, respectively.

[0046] Impedance-structure joint graphs are constructed based on topological feature data and local density feature data to obtain impedance-structure joint graph data.

[0047] The impedance-structure joint diagram data is processed by structural principal mode impedance to obtain regional impedance characteristic data.

[0048] This invention significantly improves the accuracy and discriminative power of regional features in impedance images through structure-guided multidimensional clustering and joint graph modeling. Impedance voxel clustering guided by anatomical masks ensures the clustering process occurs within anatomical boundaries, enhancing the consistency between functional zoning and anatomical structures. Extracting topological and local density features separately characterizes the spatial organization and response aggregation degree among voxels, providing a structure-function coupled multidimensional description for anomaly identification. Constructing an impedance-structure joint graph effectively integrates topological relationships and density distribution, making feature expression more regionally relevant and context-sensitive. Structural principal mode impedance processing focuses on high-contribution structural patterns, combating noise interference and highlighting key anomalous regions.

[0049] Optionally, the calculation of the regional anomaly probability value includes:

[0050] The regional feature offset is calculated based on the regional impedance characteristic data and the preset state template builder to obtain the regional feature offset data. The state template builder is a regional reference template generated by multimodal clustering and structural consistency modeling based on the impedance characteristic distribution of prior normal samples.

[0051] Structural-functional coordination weighting is performed on the regional feature offset data to obtain offset-weighted data;

[0052] Based on the offset-weighted data, the anomaly probability of the anatomical mask data is mapped to the region anomaly probability value data.

[0053] This invention achieves precise quantification and spatial visualization of regional-level electrical impedance anomalies by introducing a state template builder and a structural-functional coordination weighting mechanism. The state template builder, constructed based on prior normal samples, extracts impedance feature distribution patterns under multimodal clustering and generates reference templates by combining anatomical structural consistency, ensuring co-modeling of the structure and function of the normal state. By comparing and analyzing the current regional impedance features, feature offsets are calculated, effectively characterizing the degree of difference between abnormal regions and healthy references. The structural-functional coordination weighting strategy dynamically integrates the coordination relationship between morphological features and impedance response, improving the robustness and physiological relevance of anomaly localization. It also enables the probability mapping of anomalies in anatomical regions, making the abnormal regions present a probabilistic and continuous expression, which helps improve the interpretability and credibility of clinical judgment.

[0054] Optionally, this application also provides a system for assessing the electrical impedance status of fused CT features, used to perform the method for assessing the electrical impedance status of fused CT features as described above, wherein the system for assessing the electrical impedance status of fused CT features includes:

[0055] The CT structure perception module is used to perform CT scans on the subject to obtain CT scan data; and to generate anatomical mask data based on the CT scan data.

[0056] Impedance feature alignment module is used to acquire three-dimensional impedance data through an impedance scanning device; perform multi-scale decomposition feature extraction based on the three-dimensional impedance data to obtain decomposed feature data; and perform elastic harmonic network mapping based on the decomposed feature data and anatomical mask data to obtain configuration calibration impedance map data.

[0057] The structural guided impedance feature extraction module is used to extract regional impedance features from the configuration calibration impedance map data based on the anatomical mask data, and obtain regional impedance feature data.

[0058] The regional anomaly probability value calculation module is used to calculate the regional anomaly probability value based on the regional impedance characteristic data, and obtain the regional anomaly probability value data.

[0059] The purpose of this invention is to achieve high-precision state assessment of electrical impedance information under the semantic guidance of anatomical structures by constructing a structure-function integrated cross-modal fusion process. Anatomical mask data generated from CT scans provides a high-confidence structural reference for spatial alignment of impedance images, effectively alleviating the problems of low resolution and blurred boundaries in impedance images. Local contrast changes and response hierarchy features in three-dimensional impedance data are extracted through multi-scale decomposition, enhancing the expressive power of impedance changes at different spatial scales. Topologically conserved registration between impedance maps and CT anatomical structures is achieved through configuration alignment under structural constraints using elastic harmonic network mapping. An impedance-structure joint map is constructed to extract topological and density co-features, and the representativeness and separation of regional feature expressions are enhanced through structural principal modality analysis. A state template offset and structural-functional coordination weighting strategy are combined to accurately calculate the probability of regional anomalies. This invention enhances the regional sensitivity and anatomical consistency of impedance assessment while maintaining structural integrity, significantly improving the accuracy and interpretability of non-invasive tissue function monitoring. Attached Figure Description

[0060] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0061] Figure 1 A flowchart illustrating the steps of an embodiment of an impedance state assessment method incorporating CT features is shown.

[0062] Figure 2 A flowchart illustrating the steps of a CT scanning method according to an embodiment is shown;

[0063] Figure 3A flowchart illustrating the steps of a multi-scale decomposition feature extraction method according to an embodiment is shown.

[0064] Figure 4 A flowchart illustrating the steps of a method for extracting regional impedance features according to an embodiment is shown.

[0065] Figure 5 A flowchart illustrating the steps of a method for calculating the probability value of regional anomalies according to an embodiment is shown.

[0066] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0068] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0069] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0070] This embodiment is based on experimental data from an imaging center of a research-oriented tertiary hospital. The aim is to verify the effectiveness and accuracy of tissue status assessment and zonal identification after fusing CT image structural information with electrical impedance tomography features. In the sample preparation stage, an adult subject with suspected liver lesions was selected. A combined chest and abdominal scan was performed using low-dose spiral CT. Scanning parameters were set as follows: first dose group (low dose) 120kVp / 15mAs, second dose group (standard dose) 120kVp / 65mAs, slice thickness 1.5mm, and reconstruction matrix 512×512. After dose difference map comparison and analysis, the system automatically identified structurally faithful regions and performed structural gradient harmonization processing to generate CT data with enhanced structural consistency.

[0071] In terms of anatomical mask construction, the main density gradient directions are first extracted from the CT volume and a structure-guided flow field map is constructed. Then, the consistency of the flow field direction is used for path tracing, and path aggregation is completed in the local intensity stable region, thereby segmenting the main organ regions such as liver lobe, kidney, and stomach cavity, and outputting a three-dimensional anatomical mask as a structural reference frame.

[0072] Three-dimensional impedance data of the examined region is acquired using a multi-channel electrical impedance scanning device, with scanning frequencies set at 50kHz, 100kHz, and 250kHz. The acquired impedance data is then fed into a multi-scale decomposition module. First, a spherical harmonic transformation is performed to separate the spherical components. Then, non-equidistant three-dimensional wavelets are used to complete multi-layer decoupling of local details. In this stage, the system calculates two indices: the maximum level of the local response and the rate of change of scale contrast. The former is used to identify the active layer of the impedance response, and the latter is used to measure the response discontinuity in the lesion's peripheral region. The system integrates these indices and constructs an impedance decomposition feature map through cross-scale trend fitting.

[0073] After fusing the aforementioned decomposition features with the anatomical mask, the system constructs a function-structure dual-map atlas. While maintaining the structural topology, it performs configurational harmonization on the impedance map, ensuring high-fit impedance mapping accuracy even when the examined tissue undergoes minor deformations due to anatomical variations or respiratory states. A configuration-calibrated impedance map is output through a deformable network constrained by topological conservation.

[0074] Based on the structural regions defined by the anatomical mask, the system performs voxel-level clustering on the impedance map, calculating the topological connectivity index and local density variation curve for each region. By constructing a joint structure-impedance map and using structural principal mode processing to compress redundant features, the system extracts discriminative regional impedance feature representations.

[0075] During the anomaly detection phase, the system invokes a state template builder pre-constructed from 42 normal samples to perform region-level offset estimation. The impedance characteristics of each region are matched with the multimodal GMM distribution in its corresponding template, and an offset score is output. The system then calculates structural-functional coordination based on the neighborhood structure map, integrating local anomalies into an overall offset risk value. This risk value is projected onto the original anatomical space after sigmoid mapping, forming a three-dimensional anomaly probability map. In this example, the system detected a high-probability region (probability 0.91, volume approximately 1.6 cm³) in the posterior segment of the right lobe of the liver.

[0076] Please see Figures 1 to 5 This application provides a method for assessing electrical impedance status by fusing CT features, the method comprising:

[0077] S1. Perform a CT scan on the subject to obtain CT scan data; generate an anatomical mask based on the CT scan data to obtain anatomical mask data;

[0078] Specifically, a low-dose CT scanner is used to scan the examined area (e.g., the abdomen) under two preset radiation intensities, acquiring corresponding first and second CT image data. Based on the acquired image data, the system constructs a dose difference map reflecting the difference in dose response, and calculates various local structural indicators such as the dose intensity change rate, image gradient flow field, and phase perturbation entropy. The dose intensity change rate is the difference in grayscale response of the same area under different radiation intensities; the image gradient flow field is the degree of directional change of local structural edges; and the phase perturbation entropy is the degree of perturbation of the image's frequency domain information. Specifically, a three-dimensional Hilbert transform is used to extract the instantaneous phase of the original image, constructing a phase map, and statistically calculating the entropy value of the phase distribution within the local neighborhood. The system inputs these multi-source structural indicators into a non-convex uniform feature space, performs joint projection modeling, and extracts regions with high dose response uniformity and strong structural stability to construct structurally faithful region data. Based on this structurally faithful region data, a structural gradient harmonicization operation is performed, using structural diffusion to smooth the image gradient field within the region and enhance its boundaries, resulting in a high-quality CT image that preserves the main tissue structures. The system performs multi-directional structural flow field analysis on the enhanced CT images, such as extracting structural path responses in each direction. This involves calculating the three-dimensional gradient vector of the gradient-harmonized image to obtain the voxel-level directional field. For each voxel, the system examines neighboring voxels within a 26-neighborhood whose angle with its direction vector is less than a set threshold (e.g., 30°), retaining those with similar orientations as path candidate points. Starting from the point of maximum local response intensity in the image, forward tracking is performed along the gradient direction, while simultaneously constraining directional deviation and intensity decrease thresholds. The tracked path satisfies: ,in Let be any pair of continuum elements in the path. Let be the three-dimensional coordinates of the i-th voxel in the path. Let i be the three-dimensional coordinates of the (i+1)th voxel in the path. This represents the current structural direction path obtained through tracking. It is a cosine function. The angle between the gradient directions of the current point and the next point. For the image at points The gradient vector at that point, For the image at points The gradient vector at that point, The cosine threshold of the direction angle. For the image at points The grayscale value or intensity value at that location The response retention factor (e.g., 0.8) is used. For the image at points The grayscale value or intensity value at the location is used. All qualified paths are integrated according to their response intensity to generate a multi-directional structural path response map. Each voxel value represents its frequency in a significant structural path or the superposition of response values. Through path consistency analysis, consistent pathways in the edge directions within local regions are identified, and a directional consistency path map is constructed accordingly. Based on this path map, a path aggregation operation is performed to aggregate structural paths with strong edge responses and high directional concentration to form a path aggregation domain map. The system performs boundary segmentation of tissues and organs based on the path aggregation domain map and outputs anatomical mask data, such as the extracted three-dimensional contour boundary information of organs like the liver and kidneys.

[0079] S2. Acquire three-dimensional impedance data using an impedance scanning device; extract multi-scale decomposition features based on the three-dimensional impedance data to obtain decomposition feature data; perform elastic harmonic network mapping based on the decomposition feature data and anatomical mask data to obtain configuration calibration impedance map data.

[0080] Specifically, with the aid of the acquired anatomical mask for localization, an electrical impedance tomography (EIT) device (such as a three-dimensional acquisition system based on electrical impedance tomography (EIT) or electrical impedance spectroscopy (EIS)) is activated to perform multi-channel scanning of the examined area, obtaining three-dimensional impedance voxel data containing the conductivity distribution of the examined tissue. This voxel data presents the complex impedance response value of each point in a spatial lattice, covering the tissue space consistent with the anatomical region. The system performs multi-scale feature decomposition processing on the aforementioned three-dimensional impedance data. On one hand, based on spherical harmonic transformation, its frequency domain structure pattern in spherical coordinates is extracted, transforming the three-dimensional impedance voxel data Z(x,y,z) to a spherical coordinate system representation. ,in For radius, Polar angle, This is the azimuth angle. Apply a spherical harmonic transformation to this spherical coordinate voxel data: ,in Let be the order of the spherical harmonic function. The maximum order is denoted by , which is the highest order of the spherical harmonic expansion. Let be the degree of the spherical harmonic function. Let be the coefficient function for the corresponding frequency channel, representing the impedance frequency response at radius r. Let be spherical harmonic basis functions. The system calculates the coefficient distribution under multiple orders l and statistically analyzes its contrast rate of change. ,in It is the maximum value among all spherical harmonic coefficients of order l and degree m. It is the minimum value of the spherical harmonic coefficients of all orders m under the current order l. The average value of the spherical harmonic coefficients at all orders m of the current order l is used as an indicator of impedance mode variation at different spatial frequencies, representing the spectral characteristics of the functional organization structure. On the other hand, three-dimensional non-equidistant wavelet decomposition is performed, i.e., the system processes the three-dimensional impedance data... Three-dimensional non-equidistant wavelet decomposition is performed, and a multi-level feature map is constructed using a structure such as DT-CWT. The wavelet basis is an asymmetric complex wavelet basis. Local response features are extracted at scale j and direction k to construct a hierarchical response map: The system performs operations within each level j. The system calculates local maxima, extracts their spatial coordinates and response values ​​to form a local maximum response point map, which is then used to construct a hierarchical structure of local response intensity. After completing the dual-path feature extraction, the system calculates the contrast change rate for each spherical harmonic component; simultaneously, it extracts the local maximum response intensity points for the wavelet response hierarchy, constructing a local maximum response hierarchy map. The two feature channels are fused, and cross-scale trend modeling is performed to establish a consistent mapping between the dynamic distribution trend of impedance within the region and the anatomical structure background, forming complete decomposed feature data. After acquiring the decomposed feature data, combined with the aforementioned anatomical mask information, the system constructs an elastic harmonic network mapping module. The system first constructs a set of functional-anatomical dual-map structures, where the functional map constructs a node map based on the spatial pattern of impedance change, and the anatomical map constructs a structural map based on path aggregation and organ boundaries. The system uses the graph attention mechanism in graph neural networks to perform joint embedding learning on the two maps, extracting feature distributions with consistency between the functional and anatomical layers. Based on this embedding result, the system introduces an elastic mesh mapping model, using each structural node as a control point to perform non-rigid deformation registration, constructing a function in the form of... The elastic configuration transformation function, where For any point on the impedance diagram, The output is the registered spatial location. This transformation maintains both anatomical topological consistency and functional map response integrity, ensuring that the impedance response region deforms synchronously with the anatomical boundary during the mapping process. The system completes the impedance map configuration calibration operation and outputs a set of configuration-calibrated impedance map data aligned with the CT structure, serving as the basis for regional feature extraction and anomaly identification.

[0081] This invention provides a formula for calculating spherical harmonic transform. Spherical harmonic expansion analyzes the global and detailed structures along the spherical surface, while wavelet decomposition captures local textures and boundary anomalies. Higher-order spherical harmonic terms and wavelet detail components are more sensitive to subtle perturbations and lesion structures, enhancing anomaly detection capabilities. Simultaneously considering the spherical spectrum and spatial local response, it provides complete features for subsequent elastic harmonic mapping and anomaly localization. The highest-order spherical harmonic term L and the wavelet scale... Both methods offer flexibility, achieving a balance between accuracy and computational cost. This method is based on signal decomposition and multi-scale analysis theory. Spherical harmonic expansion is used to describe smooth patterns and symmetries on a sphere, while wavelet decomposition emphasizes local spatial perturbations and directionality. The combination of these two approaches can comprehensively capture the structure-function information of the impedance field from macroscopic to microscopic levels, ensuring a reliable input basis for subsequent configuration calibration and anomaly detection.

[0082] S3. Based on the anatomical mask data, extract the regional impedance features from the configuration calibration impedance map data to obtain the regional impedance feature data;

[0083] Specifically, the system uses anatomical mask data to spatially segment the configuration-calibrated impedance map data, dividing the entire image into multiple regions of interest (ROIs) based on organ boundaries or anatomical partitioning features, such as the liver, kidney, and pancreas regions. Each region is treated as an independent analysis unit for feature modeling. Within each segmented anatomical region, the system performs K-nearest neighbor-based clustering on the impedance voxel data within the region. The system clusters and merges voxels with similar impedance values ​​and spatial adjacency, forming multiple impedance clumps. Each clump is considered a potential functional unit with preliminary structural consistency. For the aforementioned impedance clump set, the system extracts its spatial topological and statistical features. Spatial topological features include connectivity indices, such as the Betti number (representing the number of connected regions and the topological structure of loops). Statistically, the system calculates the local density peak and its distribution range for each clump. The system constructs an impedance-structure joint graph using impedance clumps as nodes. The edges in the graph are set based on spatial connectivity measures, such as Euclidean distance thresholds, contact surface sharing, or voxel overlap. The system extracts principal modes from the joint graph structure, employing maximum flow mode analysis to extract subgraphs of nodes in the entire graph that possess the highest information flow capacity, the most stable propagation path, or the highest density. The output regional impedance feature data includes high-density impedance cluster identifiers within each anatomical region, their corresponding topological connectivity patterns, density index distribution maps, and principal mode annotations.

[0084] S4. Calculate the regional anomaly probability value based on the regional impedance characteristic data to obtain the regional anomaly probability value data.

[0085] Specifically, the system first invokes a pre-defined state template builder. This builder is constructed based on historical impedance data from a large number of healthy individuals, employing a multimodal Gaussian mixture model to model the regional impedance characteristics of the healthy population. It then uses structural consistency constraints for clustering and classification, generating regional reference templates with anatomical specificity and statistical stability. Each template includes common impedance distribution centers, variance-covariance matrices, and modal weights for that anatomical region. After acquiring the regional impedance characteristic data of the current subject, the system compares it with the template features corresponding to each region, calculating its offset vector in the impedance feature space. This offset vector represents the statistical distance between the current subject and the standard healthy distribution, and Mahalanobis distance or cosine distance can be used as the evaluation index. Based on the offset, the system introduces a structural-functional compatibility factor as a weighting term to compensate for the anatomical structural matching degree ignored by the pure impedance offset. For example, the system uses the region's tissue density compatibility score, structural alignment residual value, or functional response distribution consistency index as compatibility weights to weight and correct the offset vector. Normalization is then applied to the errors caused by structural anomalies to improve the stability and anatomical reliability of the anomaly assessment. The system maps the weighted offset values ​​back to the regions marked by the anatomical mask, and combines the offset magnitude, direction, and multimodal density function to calculate the anomaly probability value for each region using an anomaly probability function (such as posterior probability estimation based on GMM). Each anatomical region will obtain a numerical anomaly probability index, representing the likelihood of abnormal tissue impedance characteristics (i.e., the degree of difference from prior empirical data) occurring in that region.

[0086] Optionally, the CT scan includes:

[0087] S11. Perform CT scans on the subject under preset first radiation intensity conditions and preset second radiation intensity conditions respectively, and obtain first CT scan data and second CT scan data respectively.

[0088] Specifically, the subject lies supine on the scanning table, maintaining a static posture to minimize image distortion caused by body movement or respiratory motion. A multi-slice computed tomography (CT) system with multi-dose imaging capabilities (e.g., dual-source CT, adjustable tube current CT system) is used to perform two consecutive scans of the target anatomical region. The first scan is set to a preset low-radiation-intensity mode, for example, a tube current of 40 mAs and a voltage of 100 kVp, to acquire low-dose image data; the second scan is set to a high-radiation-intensity mode, for example, a tube current of 120 mAs and a voltage of 120 kVp, to acquire high-dose image data. The two scans should be completed consecutively within a short time interval to avoid anatomical mismatches caused by subject displacement, respiratory phase changes, or physiological state changes. The raw image data obtained from the scans are stored and output in the medical imaging standard format DICOM. The first set of data is labeled as low-dose scan data (D1), and the second set of data is labeled as high-dose scan data (D2). Each set of data includes complete metadata such as slice thickness, reconstruction matrix, spatial resolution, and scan parameters.

[0089] S12. Construct a response difference map from the first CT scan data and the second CT scan data to obtain response difference map data;

[0090] Specifically, spatial registration is performed on low-dose CT image data (denoted as D1) and high-dose CT image data (denoted as D2). The system uses the high-dose CT image D2 as a reference and the low-dose image D1 as the object to be registered. Affine parameters are initialized, and mutual information is used as the optimization objective. Through iterative optimization, the affine transformation matrix with maximum mutual information is obtained, thereby achieving registration. After registration, intensity difference calculations are performed on the voxels (x, y, z) at each corresponding position in the registered D1 and D2 images to construct a response difference map. Its calculation method is defined as follows: ,in, This represents the dose response difference at three-dimensional coordinates (x, y, z).

[0091] S13. Estimate the structure-fidelity region from the response difference map data to obtain the structure-fidelity region data;

[0092] Specifically, the system performs intensity change rate analysis on the constructed response difference map. Within each three-dimensional voxel neighborhood, the system calculates its local statistical variance and first-order gradient response amplitude to obtain the change rate index map. The system constructs a structure-guided anisotropic flow field map. Using the structure tensor method, the system analyzes the gradient direction and amplitude covariance matrix of each voxel neighborhood in the image, extracting eigenvalues ​​and eigenvectors of the principal directions to generate a structure flow field vector map. After completing the intensity change rate and structure flow field construction, the system also performs phase perturbation entropy analysis on the image. A three-dimensional Hilbert transform is used to extract the instantaneous phase of the original image, constructing a phase map, and statistically calculating the entropy value of the phase distribution within the local neighborhood. Regions with high phase perturbation entropy indicate significant noise or artifact interference and are unsuitable for subsequent structure analysis. The system jointly maps the above three types of feature data (change rate map, structure flow field map, and phase perturbation entropy map) to a consistent feature space. The mapping process employs non-convex embedding methods, such as t-SNE based on probabilistic nearest neighbor preservation or UMAP based on manifold preservation, to achieve low-dimensional distribution fitting and aggregation at the feature level. Based on feature clustering density and feature stability evaluation, the system generates structure-fidelity region masks, retaining only regions that simultaneously meet the following three conditions: ① stable intensity response; ② high structural orientation consistency; ③ low phase perturbation.

[0093] S14. Perform structural gradient harmonization based on the structural fidelity region data to obtain CT scan data.

[0094] Specifically, the system performs gradient operator processing on the image within the region identified by the aforementioned structure-fidelity region mask to generate a structure gradient field map. During this processing, the system uses the Sobel operator or a higher-order gradient filter to calculate the first derivative of the 3D image along three orthogonal axes to obtain the local brightness change rate of each voxel. The system combines the structure gradient direction, fidelity region boundary information, and principal direction tensor features to construct a structure diffusion weight matrix. This matrix guides subsequent diffusion operations, and its specific weights are formed by weighting the following factors: ① the difference in gray-level gradients between voxels (measuring local consistency); ② the cosine of the angle between the principal direction vectors (measuring directional consistency); ③ whether it is located in a fidelity edge region (controlling the diffusion range). This diffusion weight matrix can be learned and optimized using a graph convolutional network. In the diffusion stage, the system performs harmonic processing on the original CT image through edge-preserving partial differential diffusion calculation, which is as follows: ,in Represents the image intensity function. For time steps (diffusion time). For divergence operators, To control the edge suppression function of the diffusion coefficient, for example ,in For gradient magnitude, It is a natural exponential function. This is the edge sensitivity threshold. Through this calculation, the system can suppress and fill in local noise and scanning artifacts while maintaining the integrity of tissue structure edges within the fidelity region. The system fuses the harmonized image block within the structural fidelity region with the original CT image. At the fusion boundary, the system applies a Gaussian transition window function or a multi-resolution Laplacian pyramid fusion method to achieve smooth stitching of image intensity and gradient.

[0095] This application provides an edge-preserving partial differential diffusion calculation formula. This formula smooths noise and artifacts within the region while preserving tissue edges; it suppresses diffusion at gradient abrupt changes (true boundaries) to ensure that structural contours are not blurred; after diffusion, the harmonized image is smoothly stitched with the original CT data, thereby obtaining image patches with clear details and low noise, improving subsequent accuracy. The edge suppression function controlling the diffusion coefficient acts as follows: When it is large (near the edge), Approaching 0, diffusion almost stops, protecting the edge; when When the area is smaller (flat), The diffusion process is fast, which helps to smooth noise. Within the structurally faithful region, the diffusion process can significantly reduce high-frequency disturbances such as scanning artifacts and X-ray noise. Since the diffusion intensity is negatively correlated with the gradient amplitude, the true boundary is protected and structural details remain clear. When the diffused image is fused with the original CT data, a result with natural transition and consistent brightness can be obtained.

[0096] Optionally, the structure-fidelity region estimation includes:

[0097] The dose intensity change rate was calculated from the response difference map data to obtain the response intensity change rate data;

[0098] Specifically, the system registers CT images under different dose conditions (e.g., 50 mAs and 150 mAs) and constructs corresponding response difference map data. The system calculates the dose intensity change rate and divides the image into fixed-size sliding windows (e.g., ...). (Voxel), within each window, the local mean difference and variance difference under different dose conditions are calculated and normalized to obtain the rate of change value. This rate of change is recorded as a three-dimensional tensor, representing the sensitivity of the local dose response.

[0099] Image gradient flow field data is obtained by calculating the image gradient flow field based on the response intensity change rate data and the response difference map data.

[0100] Specifically, based on the aforementioned rate-of-change data, image gradient flow field calculations are performed on the response difference map. The gradient flow field is constructed using a three-dimensional structure tensor, where the structure tensor is defined as: ,in Let be the structure tensor, representing the three-dimensional local gradient covariance matrix at position (x, y, z). For Gaussian kernel, For along Image gradient in direction, In response to the difference image, For along Image gradient in direction, For along Image gradient in the direction.

[0101] Phase perturbation entropy is calculated based on image gradient flow field data to obtain phase perturbation entropy data;

[0102] Specifically, based on the short-time Fourier transform (STFT) of the image, the local frequency domain phase spectrum of the image is extracted using a sliding window method, and its perturbation entropy is calculated: ,in Let be the phase perturbation entropy, representing the degree of uncertainty or disorder of the local frequency domain phase spectrum at voxel (x,y,z). For phase frequency component index, To the first in the local frequency domain The energy percentage of each phase component.

[0103] This invention provides a perturbation entropy calculation formula, which quantitatively measures the degree of disorder in local phase distribution and characterizes the perturbation intensity of noise or artifacts on the structural phase. Complementary to features such as structural gradient and dose change rate, it helps identify stable structural regions and eliminate interference regions with severe phase fluctuations. In subsequent "non-convex consistent feature space projection," it serves as a weighting factor to enhance the fidelity of the true structure, contributing to the construction of structurally faithful regions. Relying solely on gradient or intensity information cannot effectively distinguish between "real edges" and "random textures." Phase information is sensitive to local geometry and texture patterns, supplementing the blind spots of amplitude features. Entropy measurement provides a quantitative standard, automatically assessing the stability of the phase spectrum. Information entropy is widely used in image analysis to measure texture complexity and signal uncertainty. Combined with the frequency domain locality of STFT, it can suppress global noise interference and enhance robustness to detailed structures. In the scenario described in this application, using entropy to screen for highly stable regions can significantly improve the recognition accuracy of structurally faithful regions, providing reliable support for subsequent gradient harmonic reconciliation and EIT-CT fusion.

[0104] Based on the response intensity change rate data, image gradient flow field data, and phase perturbation entropy data, a non-convex consistent feature space projection is performed to obtain the structure-fidelity region data.

[0105] Specifically, the system inputs the three features mentioned above—dose intensity change rate tensor, gradient flow tensor, and phase perturbation entropy tensor—into the non-convex consistency embedding module to perform feature space projection. This embedding process uses tensor sparse nonnegative matrix factorization as its core framework, introducing a Tikhonov regularization term to construct a non-convex optimization objective, thereby maintaining structural consistency and boundary stability between feature spaces. The optimization objective is defined as follows: for the feature vector corresponding to each voxel position… ,in , , Representing the dose intensity change rate, the principal direction index of the gradient structure tensor, and the phase perturbation entropy, respectively, the system optimization uses the following non-convex objective function: ,in To be the minimum value, Let be the low-rank embedding decomposition matrix to be determined, and k denote the dimension of the embedding space. Representing matrices in a low-dimensional structure. The input feature matrix is, i.e. The set, The regularization coefficient is... It is a Laplace matrix, constructed on a spatial adjacency graph. It is the Frobenius norm. The L2 norm is used. Through this embedding operation, the system maps all voxels to a unified low-dimensional structure-fidelity space and extracts the residual error distribution of their embedding responses. Using reconstruction error as a consistency indicator, the system sets a threshold based on statistical distribution (e.g., selecting the region with the smallest reconstruction error within a 95% confidence interval) to identify a set of regions with high structural stability and consistent image responses, forming a three-dimensional structure-fidelity mask image.

[0106] This invention provides a non-convex objective function. The mathematical essence of this non-convex objective function lies in extracting the most structurally representative feature representation under low-rank constraints and modeling the local consistency of the graph structure through a Laplacian term, thus avoiding isolated outliers interfering with the low-dimensional representation. After embedding, the system calculates the reconstruction residual for each voxel. The system uses residual values ​​as a structural consistency index; smaller residuals indicate better projection fitting in the original multi-source structural feature space and a higher likelihood of belonging to a structurally stable region. Based on this, the system constructs a statistical distribution of residual values ​​and sets a threshold (e.g., selecting the region with the smallest residual within a 95% confidence interval). The corresponding voxels of this region are then labeled as structurally faithful regions, and a 3D structurally faithful mask image is output. Through this method, the system achieves joint modeling of multimodal structural features, automatically identifying regions with consistency and robustness in intensity response, boundary stability, and phase information.

[0107] Optionally, the structure gradient harmonization includes:

[0108] Structural gradient field map data is generated based on structural fidelity region data.

[0109] Specifically, the original CT image I(x,y,z) is masked using a structure-fidelity region mask M(x,y,z) to preserve voxels within the fidelity region. A 3D Sobel operator or a 3D Scharr operator is then applied to the masked image to calculate the gradients in three directions: ,in The gradient component is in the x-direction. For partial derivative operators, The original CT image. x is the x-axis in space. The gradient component is in the y-direction. The y-axis is the coordinate axis in space. Let z be the z-space coordinate axis. Construct the structure gradient field plot. This serves as the regional edge structure intensity map. Simultaneously, the principal direction tensor is constructed. .

[0110] Fidelity region structure diffusion weights are calculated from the structure gradient field map data to obtain fidelity region structure map data.

[0111] Specifically, at each voxel location (x, y, z), the structural sensitivity index is calculated based on its gradient field magnitude and directionality: ,in For the diffusion weight map values, It is a natural exponential function. For spatial gradient operators, The structure gradient field map, The scaling factor is a smoothing factor, ranging from 0.5 to 1.5. If a graph neural network (GNN) is used, with voxels as nodes, the adjacency weights are: ,in The weight of the adjacent edge. As the directional similarity factor, The angle between the two gradient vectors. For nodes gradient vector, For nodes The gradient vector. The output structure graph is a three-dimensional tensor that identifies the weight distribution of each voxel in the spatial structural continuity.

[0112] This invention provides a formula for calculating a structural sensitivity index, which is applicable at locations of drastic structural changes (i.e., edge regions). Large values ​​lead to Approaching 0 inhibits diffusion; in regions with smooth structures, Small value, A value close to 1 allows for sufficient diffusion. This mechanism effectively avoids boundary blurring and enhances the structural consistency of the harmonic image. Adjacent edge weights are constructed using each voxel as a graph node. , representing the degree of structural consistency between voxel i and its neighboring voxel j. The diffusion intensity is automatically adjusted based on local gradient changes, making diffusion more "structure-sensitive" and avoiding edge blurring. Combining direction and amplitude information effectively suppresses diffusion across the boundaries of different anatomical structures, enhancing the anatomical consistency of the harmonic image. From a derivation perspective, this formula originates from the design of the diffusion coefficient term in the diffusion partial differential equation. The core idea is to determine the diffusion intensity through the gradient change rate, i.e., a "deformation-based heat conduction" model. The exponential function ensures that the diffusion coefficient decays rapidly at abrupt gradient changes, which is crucial for structure preservation.

[0113] Structural gradient harmonic diffusion was performed on the fidelity region structure map data to obtain CT scan data.

[0114] Specifically, construct the structure-preserving diffusion partial differential equation: ,in This represents the grayscale intensity of the current CT image at spatial location (x, y, z). For partial derivative operators, For the time iteration step variable, For spatial gradient operators, This is the diffusion weight map from the previous step. A multi-step numerical iteration is performed using the finite difference method, with each step executed only within the fidelity region. The system performs direction-guided diffusion processing (e.g., principal axis guidance from the structure tensor): ,in The structure guides the diffusion tensor field. The main direction of diffusion characteristic value (maximum). The first principal direction feature vector, The symbol for vector transpose. These are the diffusion characteristic values ​​in the secondary principal direction. The second principal direction feature vector, The minimum directional diffusion characteristic value, This is the eigenvector of the third principal direction. Output the harmonicized CT data. , which is the harmonic result of the CT image after this iteration.

[0115] This invention provides a structure-preserving diffusion partial differential equation, which is a structure-specialized form of the anisotropic diffusion equation, wherein the diffusion rate... The system adaptively adjusts to gradient changes, protecting structural edges and smoothing unstructured regions. Through a structure tensor guidance mechanism, a structure-guided diffusion tensor field is constructed at each voxel. This tensor field allows the system to control diffusion, preferentially propagating along the principal structural directions while suppressing diffusion perpendicular to boundaries, effectively avoiding edge blurring and texture stretching. Numerically, the system employs the finite difference method for discrete iteration of the diffusion equation, with each step performed only within the region defined by the structure-fidelity mask, ensuring computational efficiency and regional accuracy. After multiple iterations, the system outputs a harmonicized CT image. This image, while preserving the original anatomical structure morphology, significantly reduces the influence of noise and artifacts, and improves the continuity and spatial stability of tissue boundaries. By guiding the diffusion direction through the structural tensor, information diffusion in the edge direction is significantly reduced, protecting tissue boundaries; in structurally consistent regions, the diffusion weight is higher, and the diffusion process can effectively smooth ray artifacts and scanning noise; from a mathematical derivation perspective, this method combines the form of anisotropic diffusion PDE with principal axis analysis of the graph structural tensor, possessing a clear physical and structural interpretation, and can be stably implemented through standard numerical calculation methods (such as finite difference and tensor convolution).

[0116] Optionally, the anatomical mask generation includes:

[0117] Multi-directional structural flow field maps are constructed based on CT scan data to obtain structural flow field map data;

[0118] Specifically, matrix analysis is performed on the three-dimensional CT image I(x,y,z) to construct the structure tensor:

[0119] ;

[0120] in The structure tensor matrix, The sign of the partial derivative. The intensity value of a 3D CT image. The x-coordinate of the image. The y-coordinate of the image. Let z be the z-axis coordinate of the image. For each voxel, calculate the eigenvectors and eigenvalues ​​of the matrix to obtain the principal axis direction vector field (i.e., the principal directions of the structure); construct a multi-directional structural flow field map. ,in The principal eigenvector is obtained by selecting the principal axis direction corresponding to the largest eigenvalue. The largest eigenvalue serves as a structural continuity guide graph.

[0121] This invention provides a structure tensor constructor, where the diagonal of the function represents the second-order brightness variation of a 3D image along each axis, and the remaining equal-intersection terms describe the joint variation in each direction; the matrix... This is the Hessian structure tensor, used to characterize the curvature, extremum direction, and texture orientation of local regions in an image. Eigenvalue decomposition is performed on the structure tensor corresponding to each voxel, yielding three eigenvalues ​​and their corresponding eigenvectors. The system selects the eigenvector corresponding to the largest eigenvalue as the principal structural direction at that location, i.e., the direction of extension of the strongest response. The principal eigenvector provides a sense of direction for the local structure of each voxel, revealing the spatial orientation and connectivity of structures such as organs, blood vessels, and boundaries. Guided by the largest eigenvalue, the system can more accurately track continuous paths along the principal structural direction, avoiding breaks or deviations. The Hessian tensor can simultaneously perceive curvature, edges, and linear structures, exhibiting extremely high sensitivity for identifying anatomical boundaries and tubular structures (such as hepatic vessels and bronchi). From a mathematical derivation perspective, the structure tensor... The construction is based on the covariance relationship of the second-order partial derivatives. Its eigenvalues ​​and eigenvectors reflect the curvature magnitude and trend of the local brightness variation surface in different directions. The direction of the largest eigenvalue represents the local principal axis direction, corresponding to the image texture direction, edge channel, or blood vessel extension direction.

[0122] directional consistency path extraction is performed on the structural flow field map data to obtain directional consistency path data;

[0123] Specifically, the system performs a direction consistency path extraction operation based on the principal direction vector information contained in the structural flow field map to identify regions with continuous structural directions in the image. This process starts from multiple voxel positions as initial seed points and gradually advances along the principal structural direction vector, forming a path tracing line. The direction of advancement at each step is guided by the principal structural direction vector of the current voxel, constituting a spatial structural trajectory. During path tracing, the system performs direction consistency judgment. If the angle between the current advancement direction and the direction of the previous tracing step is less than a preset direction offset threshold (e.g., 15 degrees), the direction is considered continuous, and the path can continue to extend; otherwise, tracing is terminated to prevent the path from mistakenly entering regions of structural abrupt changes. Path construction employs graph search strategies such as depth-first search (DFS) or breadth-first search (BFS), and path energy is calculated for each path. ,in Path energy represents the sum of directional consistency-weighted structural variation intensities along the entire structural path. For path node indexing, This represents the total number of nodes in the path. This is the cosine of the angle between the current point's structural direction and the overall path direction. The angle between the path direction and the main direction of the structure. The structural principal direction feature value at the current path point represents the maximum structural change rate at that point. The energy function combines the characteristics of directional stability and structural strength. At each voxel point on the path, the system records the directional angle between that point and the previous node, and calculates the overall directional consistency score of the path by combining this with the structural strength score corresponding to the local principal direction. The smaller the angle and the higher the structural strength, the greater the path energy, indicating stronger structural continuity and higher stability. The system outputs a set of structural paths consisting of multiple highly consistent paths, where each path represents a structural region in the image that maintains a high degree of directional continuity.

[0124] This invention provides a path energy calculation formula that incorporates directional stability (through... ) and structural strength (through Two key factors effectively filter out path regions with strong structural continuity and high directional consistency. After path extraction, the system filters all paths by energy, retaining high-energy paths as the set of directionally consistent paths, and outputting the path set as a highly stable subgraph in the structural flow field graph. By driving with the main direction and introducing directional consistency control, the system effectively avoids paths from entering high-noise regions or regions with abrupt directional changes, improving the structural interpretability of the paths. The path energy formula combines directional stability with structural strength, providing objective quantitative indicators for path ranking and filtering. Using only paths with high directional consistency as the source of structural templates effectively improves the boundary closure and structural rationality of the anatomical region segmentation. It can be expanded into a high-energy subgraph extraction module in the structural graph, facilitating integration with graph neural networks or path optimization algorithms to achieve multi-scale structural modeling. From a mathematical modeling perspective, this method uses the main direction of the structural tensor as a priori to guide path generation, introduces angle constraints to control spatial consistency during path growth, uses the cosine square function to retain the contribution of directional consistency, and strengthens the structural expressiveness by multiplying with the main eigenvalue, exhibiting good stability, interpretability, and robustness.

[0125] Based on the directional consistency path data, a path aggregation domain graph is constructed to obtain the anatomical mask data.

[0126] Specifically, all directional consistent paths are mapped back to voxel space, and the number of times each voxel is covered by a path is counted to form a path density map: ,in The path density map value represents the number of times a 3D voxel point (x, y, z) is covered by a path with consistent orientation. For the i-th direction consistent path, A set of paths with consistent direction. For indicator functions, These are the coordinates of three-dimensional voxel points. For the path density map... Perform threshold segmentation (such as the Otsu method or local maximum inter-class variance method) to generate a binary mask: ,in The segmentation threshold for the path density map. For other conditions.

[0127] This invention provides a formula for calculating the number of times each voxel is covered by a path. The density map obtained by this formula is essentially a three-dimensional spatial projection map, reflecting the participation of voxels in the structural path within the anatomical structure. Higher participation indicates greater importance in structural extensibility and directional consistency. The system processes the path density map. A threshold segmentation operation is performed to identify regions with structural consensus characteristics. Preferably, the system uses the Otsu automatic thresholding method or the local maximum inter-class variance method to determine the threshold. The density map is segmented into binary masks. It does not rely on manually set boundaries or shape templates, but identifies structural regions solely based on the natural aggregation behavior of paths in space, exhibiting strong adaptability. By aggregating the spatial projections of multiple highly consistent paths, it effectively eliminates interference from isolated paths or paths deviating from their intended direction, extracting core regions with higher structural stability. From a mathematical derivation perspective, the path density map... This is equivalent to performing a spatial coverage integral on the path set in three-dimensional space, forming a directional path clustering intensity map of the local region. Regions with higher density have a higher "frequency of occurrence" in the structural space, representing stronger consistency and repeatability of the anatomical structure. Furthermore, image thresholding strategies such as the Otsu method can automatically obtain the optimal segmentation boundary in a statistical sense, ensuring that the segmentation results maximize inter-class variance and minimize intra-class consistency.

[0128] Optionally, the multi-scale decomposition feature extraction includes:

[0129] S21. Perform multi-resolution spherical harmonic transformation and three-dimensional non-equidistant wavelet decomposition on the three-dimensional impedance data to obtain the first impedance decomposition data and the second impedance decomposition data, respectively.

[0130] Specifically, the three-dimensional impedance data Mapping to a spherical coordinate system: ,in Let be the radius of the sphere, and let represent the distance from the origin to that point. The polar angle represents the angle between the point and the z-axis. The azimuth angle represents the angle of a point in the xy-plane. The x-coordinate in three-dimensional space The y-coordinate in three-dimensional space. Let z be the z-coordinate in three-dimensional space. Use the standard coordinate transformation formula to transform rectangular coordinates to spherical coordinates. Perform a spherical harmonic expansion on the transformed spherical impedance function: ,in The three-dimensional impedance function in spherical coordinates. Let be the order of the spherical harmonic function. Let be the highest order of the spherical harmonic expansion. Let be the order of the spherical harmonic function. The amplitude coefficient at each frequency order. Use spherical harmonic basis functions. Set multi-resolution levels. The spherical harmonic components of different frequency bands are extracted to construct a spherical multi-scale texture intensity map (i.e., the first impedance decomposition data).

[0131] The system performs non-equidistant wavelet decomposition on the original voxel data. It reconstructs and normalizes the grid mapping of the data using non-equidistant interpolation algorithms (such as B-spline interpolation or local radial basis functions). A three-dimensional asymmetric wavelet filter bank (using wavelet bases such as Daubechies or Coiflet) is constructed, and multi-level wavelet convolution is performed on the interpolated impedance voxel data to extract the frequency domain response at each scale in wavelet space. The system outputs the low-frequency approximation components at each scale. and high-frequency detail components along the x, y, and z directions. , representing local structural edges, texture abrupt changes, and high-frequency perturbations, respectively. The system performs weighted fusion of high-frequency response components at all scales to construct a multi-scale sensitivity map, i.e., the second impedance decomposition data.

[0132] This invention provides a spherical harmonic expansion calculation formula. This formula sets up a multi-level spherical harmonic expansion order set L={2,4,6,8}, extracting spherical harmonic components from different frequency bands to construct a spherical texture intensity map. This map is used to express the anatomical structural stability and symmetry distribution characteristics in impedance distribution, forming the first impedance decomposition data. Lower-order l-terms describe the overall background and slow changing trends, while higher-order terms reveal texture perturbations, edge changes, and lesion details. Through expansion at different L-levels, the system can generate a series of spherical texture maps, providing layered information. Abnormal structures are usually characterized by abnormally increased amplitudes of higher-order components; their spatial frequency characteristics can be accurately located using spherical harmonic expansion. The spherical harmonic basis functions are strictly orthogonal bases, ensuring the uniqueness and reversibility of the decomposition results, while also possessing clear physical and geometric meanings. Spherical harmonic expansion originates from solving the spherical coordinate solution of the Laplace equation and is the most natural decomposition tool for spherical functions. The change of the impedance function on the sphere can be regarded as a projection onto a set of orthogonal bases, and its expansion coefficients... This describes the energy intensity at different spatial frequencies. By controlling the highest order... It can achieve continuous characterization from smooth backgrounds to high-frequency disturbances.

[0133] S22. Calculate the multi-scale contrast change rate based on the first impedance decomposition data to obtain the contrast change rate data;

[0134] Specifically, for each spherical harmonic frequency layer l, its local contrast change is calculated: ,in For scale layer The rate of change of local contrast at position (x, y, z) Spherical harmonic coefficients The gradient value at the spatial location (x, y, z). To prevent small constants from having a denominator of zero, Let be the spherical harmonic coefficients at position (x,y,z) at spherical harmonic frequency level l and order m. The system calculates the average contrast change of each voxel at different scales; and constructs a voxel-level contrast change trend map C(x,y,z), i.e., contrast change rate data.

[0135] This invention provides a formula for calculating local contrast changes, which enables the clear differentiation of high-frequency abrupt change regions under different response intensities. For example, at tissue boundaries or lesion edges, drastic changes in high-frequency components lead to… Larger, and It may be relatively small at this time. The value will increase significantly, indicating a high rate of contrast change at that point. Fusing local gradient information from multiple spherical harmonic frequency layers can keenly capture tissue details and boundary features at different scales; regions with high contrast change rates often represent areas with strong structural discontinuities or severe signal perturbations, which can be used to assist in the localization of anomalous regions in EIT images; normalization processing improves numerical stability and contrast separability, making the response curves between different tissue structures more distinguishable; it can be embedded in subsequent feature fusion or mask extraction models, participating as part of the representation map in image segmentation, registration, or anomaly analysis. Mathematically, the numerator part of the formula... Capturing spatial mutation features, denominator part It serves to normalize the response intensity, forming a ratio index in the form of "rate of change / intensity". The derivation idea comes from the analysis of the relative variability of contrast gradient, similar to the local response contrast estimation method in image enhancement or texture analysis, and has good physical interpretability and engineering adaptability.

[0136] S23. Calculate the maximum level of local response based on the second impedance decomposition data to obtain the maximum level data of local response;

[0137] Specifically, the system traverses each spatial voxel point. and collect it at all decomposition scales. The three-dimensional wavelet coefficients are denoted as follows: The wavelet coefficients represent the degree of abrupt changes in local structure or the intensity of texture change response of a voxel at that scale level. The system calculates the strongest scale level that satisfies the following conditions: ,in voxel point Maximum response scale level In order to make The largest scale level, In scale Lower body pigment The three-dimensional wavelet coefficient values. Output the maximum response scale plot. , representing the scale at which each voxel is most significant, serving as the data of the highest level of local response.

[0138] This invention provides a formula for calculating the strongest scale hierarchy. This formula iterates through all voxels and performs the above operations to construct the output graph. The most significant response scale for each voxel in the multi-scale wavelet domain is called the local response maximization level data. This method automatically selects the most expressive scale order for each voxel, avoiding overfitting or feature loss caused by a single scale. Due to the good time-frequency locality of wavelet coefficients, this method can more accurately identify the core structural abrupt change level under noise interference or boundary ambiguity. Scale discrimination can be completed solely through maximum response calculation, avoiding complex model training or statistical analysis processes. Mathematically, the core of this formula is the operation of finding the maximum absolute value of wavelet coefficients, which is essentially a "local maximization level retrieval" problem based on multi-scale response intensity. By comparing the structural response intensity at different scales, the system can effectively extract the frequency domain scale that best characterizes structural changes.

[0139] S24. Based on the contrast change rate data and the local response maximum level data, perform cross-scale trend calculation to obtain decomposed feature data.

[0140] Specifically, the system first constructs a multi-scale response stack, decomposing the contrast change rate data of the original image by scale to generate a set of contrast response layers at different spatial scales. Each scale level can be implemented using a spatial pyramid structure, for example, by using Difference of Gaussian (DoG) or Laplacian filtering (LoG). Typical scale level values ​​include, but are not limited to, those mentioned above. =1, 2, 4, and 8. Subsequently, based on the aforementioned multi-scale stack, the system analyzes the contrast response value variation trend at each voxel location across multiple scale levels. For each voxel, the system statistically analyzes its response intensity at different scales and locates the scale level with the highest response intensity in the multi-scale response sequence, defined as the local maximum response level. This level reflects the main response scale of the voxel's structural features in spatial scale, indicating whether the region belongs to coarse or fine structure. The system establishes a cross-scale contrast response curve for each voxel, with the scale level as the horizontal axis and the contrast change rate as the vertical axis, forming the voxel's scale response trajectory. By fitting this response trajectory, the system extracts trend indicators of the voxel's cross-scale response behavior, including but not limited to trend slope indicators, representing the rate of change of the response value with increasing scale level. That is, the system performs linear fitting on the contrast response value sequence of the voxel at multiple scale levels, assuming the scale level set is... The corresponding response value is The sequence is processed by the least squares method. Fitting a straight line, the slope of which is the trend slope index; the response peak count index represents the fluctuation frequency of the voxel's structural response at multiple scales, i.e., the system treats the voxel's multi-scale response curve as a one-dimensional sequence, uses the neighborhood extremum detection method (such as five-point local maximum determination), and counts the number of local maxima in the curve, which is recorded as the peak count; the maximum response deviation index is the difference between the voxel's response value at the principal response level and its overall response mean, i.e., the maximum response value that the system determines for that voxel in the multi-scale response. And its level, and then calculate the average of all response values. The degree of deviation is represented by the difference between the maximum response value and the mean, which is defined as: The principal scale stability index represents the significance of a voxel at its principal response level, equivalent to the standard deviation multiple of the response value at that scale relative to the overall response mean. In other words, it represents the system's response value at the principal response level (the scale where the maximum response value is located). Subtract the overall response mean and divide by the standard deviation of all scale response values. To obtain standardized indicators This indicates the significance of the principal scale response. The system combines the above trend indicators to form a high-dimensional structural trend vector, thus generating decomposed feature data.

[0141] Optionally, the resilient harmonic network mapping includes:

[0142] Functional-anatomical dual-image processing is performed based on decomposed feature data and anatomical mask data to obtain functional-anatomical dual-image data.

[0143] Specifically, the system constructs a functional map based on the generated decomposed feature data to characterize the spatial structural relationships of local impedance anomaly regions. Each node in the functional map is defined as a voxel location where the decomposed feature value reaches a local maximum or minimum, representing a potential functionally active point or tissue abnormality region. The system calculates the spatial distance and the angle between the feature directions among all nodes. If the Euclidean distance between two nodes is less than a set spatial threshold (e.g., 3 mm), and the angle between their feature gradient directions is less than a preset angle threshold (e.g., 15 degrees), a connection edge is established between the two nodes. The weight of the connection edge is jointly determined by feature similarity and spatial proximity. For example, the system performs a weighted fusion based on the consistency of the feature gradient directions between two points (e.g., cosine similarity) and spatial distance (e.g., Gaussian function mapping), with the weights determined through empirical data. The system constructs an anatomical map based on the generated anatomical mask data to represent the spatial skeletal relationships of organs or tissue structures. The nodes in the anatomical map are key structural points extracted from the mask map, such as representative geometric center locations like organ edge points, bony landmarks, or organ axes. The edge connection rules between nodes are set based on structural connectivity (e.g., whether they are continuous regions) and spatial consistency (e.g., relative positional stability). The edge weight assignment method combines density differences in CT images (e.g., changes in Hounsfield values) and structural orientation consistency indicators. After constructing the functional and anatomical maps, the system performs bi-graph alignment processing on both. For example, the system applies a graph neural network encoder (e.g., graph attention network GAT or graph convolutional network GCN) to the functional and anatomical maps respectively to extract the low-dimensional embedding vector of each node, thereby representing its contextual information in the graph structure. The system uses graph alignment algorithms (e.g., maximizing graph structural mutual information or minimizing graph edit distance) to evaluate the matching relationship between the two graphs at the structural level and establish a corresponding mapping relationship between functional nodes and anatomical nodes. This mapping result can be represented as a set of cross-graph mapping matrices, constituting functional-anatomical bi-graph data.

[0144] Elastic harmonic configuration mapping is performed based on functional-anatomical dual-map data to obtain configuration mapping data;

[0145] Specifically, the system initializes the configuration field, using the matching node pairs in the functional-anatomical dual-map as control points, and establishes an elastic mesh model with the following transformation rules: ,in For spatial points Coordinates after configuration transformation To input spatial coordinates, For control point index, The deformation intensity weight for control point k, The local deformation response defined for a kernel function (such as RBF or TPS). The system undergoes harmonic deformation through a harmonic deformation function: ,in Let be the total harmonic energy function. The weighting coefficient for the elastic deformation energy term. For elastic deformation energy, Weighting coefficients for the node registration error term. For the registration error term of the graph nodes, The weighting coefficients for the space regularization term are... This represents the smoothness regularization term of the deformation field. During the optimization process, the system employs common numerical optimization methods (such as the quasi-Newton method L-BFGS or the adaptive gradient descent algorithm Adam) to iteratively update the parameters of the elastic mesh model. In each iteration, the system adjusts the amplitude and direction of the deformation kernel function based on the gradient information of the current energy function, gradually bringing the overall configuration field closer to the optimal state in terms of control point registration, elastic rationality, and spatial smoothness. The system outputs a complete three-dimensional configuration transformation field, representing the deformation mapping relationship of each spatial location based on functional-anatomical alignment.

[0146] This invention provides a formula for calculating the configuration transformation of an elastic mesh model. This formula initializes the configuration field model in three-dimensional space using matching node pairs extracted from a function-anatomy dual-map as control points. Utilizing the function-anatomy pairings between graph nodes as control points, it drives the entire graph to flexibly adjust the configuration field while maintaining spatial continuity and controllable deformation. Optimization is achieved through harmonic deformation functions, such as using numerical optimization methods (e.g., the quasi-Newton method L-BFGS or the Adam adaptive gradient descent algorithm) to adjust the deformation weights of the control points. The solution is obtained iteratively. In each iteration, the system dynamically adjusts the response amplitude and direction of each kernel function based on the gradient information of the current deformation field, so that the overall configuration field maintains global elastic rationality and spatial continuity as much as possible while ensuring the registration accuracy of control points. The system applies all control point response kernel functions. The linear combination forms the overall configuration mapping function. And by minimizing the harmonic energy function This achieves harmonized deformation alignment between the functional and anatomical diagrams.

[0147] Based on the configuration mapping data, a topology-conserving configuration transformation is performed to obtain the configuration calibration impedance map data.

[0148] Specifically, the system extracts physiologically topologically significant structural paths from anatomical mask data as the main structural skeleton. For example, in a thoracic impedance map, the main blood vessel course or airway midline curve can be extracted as the core topological structural path. During the elastic configuration mapping process, the system applies a strong topological preservation constraint to this main structural skeleton. This constraint, by constructing a structural guiding field and combining it with a Laplacian coordinate preservation strategy, ensures that these critical paths maintain their connectivity and topological invariance after overall deformation. After the topological constraint is constructed, the system formally performs the configuration transformation operation. The aforementioned optimized elastic configuration transformation field is applied to the three-dimensional voxel coordinates of the original impedance map data. For each target voxel point during the deformation process, the system traces its corresponding position in the original impedance map in reverse, and performs numerical resampling and compensation on the original image data using interpolation algorithms (such as trilinear interpolation or spline interpolation) to obtain a spatially continuous, grayscale-fidelity impedance value distribution. Because the deformation field application process introduces a certain degree of blurring effect on the original image boundaries, the system performs boundary recalibration, including edge-guided enhancement based on the structural boundary map. For example, it extracts high-frequency details of the boundaries using Gabor filters or sharpens structural edges by combining anatomical boundary gradient fields. The system outputs configuration calibration impedance map data that has completed spatial reshaping and preserves the original topological logic.

[0149] Optionally, the regional impedance feature extraction includes:

[0150] S31. Based on the anatomical mask data, perform impedance voxel clustering on the configuration calibration impedance map data to obtain impedance voxel clustering data.

[0151] Specifically, the system extracts all voxel points from the configuration calibration impedance map and performs spatial filtering based on the anatomical mask data, retaining only voxel points located within the anatomical mask. The system constructs a feature vector for each retained voxel, containing an impedance value component (representing the calibrated impedance value at that voxel), a gradient component (a vector representation of the rate of change of impedance around the voxel), and a boundary distance component (defined as the distance from the voxel point to the nearest anatomical boundary). The system then constructs a joint feature vector for each retained voxel, containing the impedance value, gradient vector, and boundary distance. Next, the system calculates the feature similarity between all voxel pairs based on a Gaussian kernel function, constructing a similarity matrix. ,in voxels eigenvectors, It is a natural exponential function. voxels eigenvectors, This is the bandwidth factor. The system calculates the standard Laplace matrix. Where L is the standard graph Laplacian matrix, D is the degree matrix, and W is the similarity weight matrix, i.e. The system performs eigenvalue decomposition on the voxels, obtaining the eigenvectors corresponding to the k smallest eigenvalues ​​to construct a low-dimensional embedding representation. It identifies cluster peaks based on local density functions (such as the number of voxels in a preset neighborhood for each point), which are used to estimate the number of cluster centers, k. The system performs K-means or density clustering on the voxel points in the embedding space to obtain a cluster label for each voxel, outputting impedance voxel clustering data. The system outputs a voxel-level cluster label map, where each voxel is assigned a corresponding cluster identifier code, representing its impedance response cluster. This cluster label map is the impedance voxel clustering data.

[0152] This invention provides a formula for constructing a similarity matrix. Through Gaussian kernel similarity, the system can better capture the nearest neighbor relationships between local features. Spectral embedding preserves the mainstream direction of data variation, revealing the intrinsic structure of the data without relying on linear assumptions. Data points in the low-dimensional embedding space are easier to cluster and visualize, facilitating the integrated identification of multimodal impedance features. Bandwidth coefficient... The settings allow control over the distribution of the similarity function, thus flexibly adapting to different data densities and structural change rates. By characterizing the "vibrational modes" or optimal partitioning directions in the graph structure using the minimum eigenvector of the Laplacian matrix L, spectral embedding maps the high-dimensional structure to a low-dimensional space, enabling high-quality clustering results to be obtained using simple clustering methods (such as K-means) in this space. This approach is more stable and interpretable than directly partitioning in the high-dimensional feature space.

[0153] S32. Based on the impedance voxel clustering data, perform topological feature extraction and local density feature extraction to obtain topological feature data and local density feature data, respectively.

[0154] Specifically, topological feature extraction involves the system treating each impedance cluster in the aforementioned clustering label graph as a topologically connected subgraph. This means that all voxels within a cluster are treated as graph nodes, and edge connections are constructed based on their three-dimensional adjacency relationships to form an undirected adjacency graph structure. For each subgraph, the system calculates the following topological structure indices: betweenness centrality, representing the frequency with which a voxel node acts as a shortest path transit point in the graph (a higher value indicates a stronger bridging role for the node within the cluster); proximity centrality, representing the reciprocal of the sum of the shortest path distances from a voxel to all other voxel nodes within the cluster (a higher value indicates the node is closer to the geometric center of the cluster structure); and local clustering coefficient, representing the probability that a voxel forms a triangular closed loop structure in its neighborhood. The system records these indices in units of voxels, forming a topologically connected subgraph with a dimension of [missing information]. The topological feature tensor, where For all valid voxel points, The number of topological indicators (e.g., 3).

[0155] The specific process of local density feature extraction is as follows: for each voxel point located within a cluster, the system defines its local spatial neighborhood as a three-dimensional spherical region with that point as its center and a preset radius (e.g., 2 mm), denoted as... For each voxel i, the system performs the following density and rate of change feature extraction process, including calculating the number of voxels in the neighborhood. ; Calculate the standard deviation or variance of the impedance gradient values ​​within the neighborhood; Perform calculations based on a weighted kernel function of impedance value differences to obtain local density data, i.e. ,in This represents the local density data for voxel i. A larger value indicates the presence of numerous tissue units in its neighborhood with impedance values ​​similar to its own. voxels The number of neighboring voxels, For any voxel in the neighborhood, It is a natural exponential function. This represents the impedance value of the current voxel. The impedance value of the neighboring voxels. The bandwidth coefficient for response sensitivity (e.g., between 0.1 and 0.5). The system outputs a three-dimensional local density feature map, recording the density response value of each voxel.

[0156] This invention provides a weighted kernel function based on impedance difference, which calculates... That is, within the specified radius of the sphere and the voxel The number of spatially adjacent voxels reflects their physical contact density; the standard deviation or variance of the impedance gradient values ​​within the neighborhood is statistically analyzed to quantify the degree of tissue texture variation in a local region; based on the impedance values ​​of each voxel within the neighborhood... impedance value of the central voxel The differences between the components are analyzed, and a Gaussian kernel function is introduced for weighted calculation to obtain local density response values. The Gaussian kernel is used to measure neighborhood impedance similarity, effectively highlighting homogeneous structural regions and suppressing abrupt texture changes or boundary noise. Through a dual factor of local statistics and response similarity, density assessment achieves both spatiality and attribute consistency. The output can form a three-dimensional local density map. This density calculation method combines the ideas of "neighborhood kernel weighting" and "attribute similarity," quantifying the self-consistency of impedance distribution within a local neighborhood to measure the stability of tissue structures. Compared to traditional voxel frequency statistics, this method incorporates electrical impedance information while preserving spatial semantics, improving the accuracy and robustness of local region identification.

[0157] S33. Construct an impedance-structure joint graph based on topological feature data and local density feature data to obtain impedance-structure joint graph data;

[0158] Specifically, each voxel in the configuration calibration impedance map is considered a node in the graph. The system associates impedance clustering labels, topological indices, and density values ​​with each node. In the three-dimensional voxel space, spatial adjacency is used to establish edge connections between nodes, such as 3D 6-neighborhoods or 8-neighborhoods. Edge weight calculation: ,in For edge weights, This is the topological similarity weighting coefficient, with a value of 0.4. For nodes and The topological similarity index can be calculated based on Euclidean distance or cosine similarity between their topological feature vectors. This is the density similarity weighting coefficient, with a value of 0.4. The degree of local density similarity between two voxels can also be calculated using the Gaussian kernel function of the density difference. This is the structural edge cue weighting coefficient, with a value of 0.2. It is a binary function that indicates whether two voxels are simultaneously located in the edge region of a CT structural image. If they are, it takes the value 1; otherwise, it takes the value 0.

[0159] This invention provides an edge weight calculation formula that fuses node attributes with impedance clustering, topological attributes, and local density information to achieve a graph structure coupling representation of multi-source features. The graph edge weights combine spatial adjacency, topological similarity, and structural edge cues, enhancing the representational capabilities of subsequent processing. It effectively preserves the continuity of anatomical edges and the structural consistency of functional response regions, resulting in higher accuracy and stability for subsequent tasks. The weight design follows the principle of "local consistency + structural constraints," avoiding information confusion at structural abrupt changes and improving the interpretability and controllability of the graph structure. This graph model, by explicitly introducing physical connectivity (spatial adjacency) and implicit attribute similarity (topological and density similarity) to form a structure-aware graph representation, enables various spatial statistical and graph neural network models to perform joint modeling and analysis based on this.

[0160] S34. Perform structural principal mode impedance processing on the impedance-structure joint diagram data to obtain regional impedance characteristic data.

[0161] Specifically, the impedance value of each voxel in the configuration calibration impedance diagram is mapped to the corresponding node in the impedance-structure joint diagram, forming a graph signal sequence, denoted as the impedance signal. The impedance signal is defined as the scalar value of the impedance corresponding to each node in the graph. The system constructs a graph frequency domain transform based on the graph Laplace operator, performs graph Laplace matrix calculation, and constructs the standard Laplace matrix L based on the graph's adjacency matrix and degree matrix; eigenvalue decomposition is performed on L to obtain the eigenvector matrix U, where each column is a graph frequency basis; the node impedance signal is then processed. Mapping to the frequency domain, calculate the graph frequency coefficients. That is, through matrix multiplication Obtain; Graphical frequency coefficients This represents the projection of the signal onto different graph frequencies. Low frequencies correspond to the dominant structural mode or global trend, while high frequencies reflect local disturbances or anomalies. The system selects the frequency components corresponding to the first few lowest eigenvalues ​​(e.g., the first 10) in the graph spectrum to represent the dominant structural mode. These frequency components are then subjected to a weighted inverse transform in the eigenvector matrix to reconstruct the dominant mode impedance map signal. This signal reflects the dominant spatial impedance variation trend in the tissue region. The local offset between the reconstructed signal and the original graph signal is calculated, characterizing the consistency or deviation of the local tissue response in the dominant mode direction. Combining the spatial distribution of nodes and the signal gradient direction, the dominant directional impedance response mode of the tissue structure is extracted. This involves calculating the gradient direction vector of each node based on the difference in dominant mode impedance values ​​between adjacent nodes in the graph structure, formally approximated by a first-order graph difference operator or Laplace gradient. The gradient direction vectors within the local neighborhood (e.g., the k-th order adjacency domain) are statistically aggregated, and the local principal direction is extracted using maximum direction consistency or principal component analysis (PCA). This principal direction represents the principal axis direction of the tissue response under the dominant mode impedance variation within the anatomical region. By using the principal direction vector as the directional feature of the structural response mode, and combining its intensity (such as the mean local gradient magnitude) and directional consistency (such as the Rayleigh statistic), a quantifiable directional impedance response index is constructed, namely the principal directional impedance response mode. Based on the above, regional impedance characteristic data are obtained.

[0162] This invention provides a formula for calculating graph frequency coefficients. Through transformation using this formula, the system can resolve the spatial variation patterns in the original impedance signal into a set of frequency components. Low-frequency components correspond to the dominant modes or global trends on the graph structure (such as uniform tissue impedance distribution), while high-frequency components reflect local perturbations, boundary abrupt changes, or heterogeneous changes in potential lesion areas. Frequency domain mapping decomposes the impedance signal into multi-scale modes, which helps identify structural perturbations that are difficult to manifest in the original space. High-frequency coefficients often contain noise or lesion-related local perturbations, which the system can use to analyze their anatomical distribution characteristics. Signal compression is achieved by retaining the main low-frequency components, and the original impedance signal can also be restored through inverse graph Fourier transform. The system constructs the graph Laplace matrix L of the structural graph and then performs eigenvalue decomposition to obtain the orthogonal eigenvector matrix U, which serves as the graph frequency basis. Essentially, it is a function defined on a graph structure, and its frequency domain behavior can be modulated by the geometric properties of the graph structure itself. This is calculated through projection. The system can then obtain the energy distribution of the signal at different structural scales, forming the core basis for frequency analysis.

[0163] Optionally, the calculation of the regional anomaly probability value includes:

[0164] S41. Calculate the regional feature offset based on the regional impedance characteristic data and the preset state template builder to obtain the regional feature offset data. The state template builder is a regional reference template generated by multimodal clustering and structural consistency modeling based on the impedance characteristic distribution of prior normal samples.

[0165] Specifically, the system provides regional impedance feature data, including topological indices, local density, and impedance-structure graph embedding vectors. The state template builder refers to a regional reference model built based on prior normal samples; its construction methods may include multimodal Gaussian mixture models, kernel density estimation models, or latent variable distribution models generated by autoencoders. The system trains on different anatomical regions to obtain corresponding normal distribution templates. The system traverses each anatomical region, summarizes the impedance features corresponding to that region, and constructs feature vectors containing the aforementioned topological indices, density values, and graph embedding vectors. The system compares the feature vectors with the normal state models of their respective anatomical regions; it calculates the offset using Mahalanobis distance, KL divergence, or contrastive loss. ,in Let be the feature offset of the i-th anatomical region. For Kullback–Leibler divergence, For the first Impedance characteristic vector of each anatomical region Anatomical region The corresponding normal state feature template, a statistical model trained from normal samples, can be a Gaussian Mixture Model (GMM), Kernel Density Estimation (KDE), or the latent variable distribution of an autoencoder. This yields the feature offset for each region, i.e., the region feature offset data.

[0166] This invention provides an offset calculation formula, when For multi-Gaussian models, this distance can be approximated using Mahalanobis distance; for autoencoder models, reconstruction error or contrastive loss function can be used for measurement. The system traverses each anatomical region and outputs its corresponding feature offset. This data constitutes regional feature offset data. By comparing with a normal regional template, it effectively measures the degree of abnormality in each region, possessing strong individualized discrimination capabilities. Statistical indicators such as KL divergence or Mahalanobis distance can be used to measure differences between nonlinear distributions, avoiding errors caused by using only low-order statistics such as mean or variance. The introduction of multi-feature joint modeling, including topology, density, and embedding, enhances the adaptability of identification to different lesion manifestations. It provides an efficient, stable, and adjustable numerical framework, facilitating subsequent collaborative processing with modules such as regional structural function map fusion and probability mapping. Voxel-level impedance features are elevated to regional-level structural modeling, and offset estimation is performed using statistical distributions under normal conditions. Since the differences between real individuals and standard templates often lack uniformity and linear structure, using asymmetric distribution measurement methods such as KL divergence can more realistically characterize the abnormal distribution risk of different anatomical regions.

[0167] S42. Perform structural and functional coordination weighted calculations on the regional feature offset data to obtain offset weighted data;

[0168] Specifically, the input region feature offset data represents the impedance feature offset of each anatomical region relative to the normal state template; the structural-functional pairing information of each region in the anatomical mask includes local CT density maps, structural connectivity maps (such as the Laplacian structural map), and functional pathways (such as the adjacent pathway map); for each anatomical region The system identifies all its structural adjacent regions. And calculate the structural-functional coordination factor of the region. :set up Indicates the region with his neighbors The structural connectivity edge strength between them (e.g., calculated from the edge gradient of a CT density map or anatomical image); let and They are respectively regions The impedance eigenvectors of its neighbors; the structural functional compatibility factor is calculated as follows: ,in As a structural-functional coordination factor, For adjacent region indexes, To be related to the anatomical area Adjacent structural contiguous regions set For the region its neighboring areas The strength of the structural connection edges between them It is a natural exponential function. Anatomical region The impedance eigenvector, Adjacent area The impedance eigenvector, This is the bandwidth parameter of the Gaussian kernel function. The system is based on the coordination factor. For the original offset value Weighting is applied to obtain the offset value after structural function adjustment (i.e., offset weighted data). ,Right now .

[0169] This invention provides a formula for calculating a structural-functional coordination factor. This formula emphasizes feature similarity, giving higher contribution weights to adjacent regions with similar features. By leveraging the anatomical connectivity and functional similarity between regions, it ensures that the interpretation of offsets is constrained by the consistency of spatial anatomical structures. The weighted offset values ​​more accurately reflect the pathological trends within the current structural environment, reducing the false positive rate. The coordination factor possesses cross-regional modeling potential and can be used to construct larger-scale multi-regional anomaly detection models. This method uses structural diagrams... Based on this, a harmonic factor for regional hierarchy is constructed by fusing anatomical connectivity strength (structural topology) and functional consistency (impedance characteristic distance) to reweight the original regional offset, so that the offset interpretation results have tissue structure consistency and regional functional continuity, thereby improving the clinical usability of EIT-CT fusion assessment.

[0170] S43. Based on the offset weighted data, perform regional anomaly probability mapping on the anatomical mask data to obtain regional anomaly probability value data.

[0171] Specifically, the offset values ​​are processed into a probability distribution range (e.g., 0–1): ,in This represents the probability value of abnormalities in the anatomical region. For the natural index term, The parameter used to adjust the slope of the probability response ranges from [1.0, 10.0]. For offset-weighted data, For all weighted offset values The system calculates the probability value for each anatomical region. The diffusion mapping is applied to its corresponding spatial mask region. Specifically, this involves assigning values ​​to all voxels within each region, with initial values ​​set to the values ​​for the corresponding region. If there are blurred boundaries or structural intersections, a spatial neighborhood diffusion strategy (such as a 3D Gaussian smoothing kernel or region mask convolution) can be used to expand and fuse the probability values. The system outputs regional anomaly probability value data, that is, the anomaly probability value map of each anatomical region in the three-dimensional structural map.

[0172] This invention provides a formula for calculating the probability mapping of regional anomalies. Mathematically, this formula ensures that the output value always remains within the (0,1) interval and exhibits monotonically increasing properties, meeting the requirements for probabilistic interpretation. The higher the weighted offset value of a region and the greater its difference from the average state, the higher the probability of it being mapped as an anomaly. The system maps the anomaly probability value of each region back to its corresponding spatial mask region in the 3D CT image, uniformly assigning values ​​to each voxel within the mask to achieve region-level voxel-level probability synchronization. In areas with blurred boundaries or overlapping structures, the system can employ a 3D Gaussian smoothing kernel function or a convolution strategy based on mask weights to perform spatial diffusion and edge fusion processing on the probability values, ensuring image continuity and natural boundary transitions. The structural offset results are transformed into interpretable anomaly probability indicators, facilitating spatial localization and risk assessment of EIT changes by users; it supports probability comparison between different structural regions, establishing a quantitative standard for anomaly levels between different anatomical regions through a unified Sigmoid normalization model; and it integrates structural masks to achieve spatial weighting and anatomical fusion, improving morphological consistency and image visibility through spatial mapping and diffusion.

[0173] Optionally, this application also provides a system for assessing the electrical impedance status of fused CT features, used to perform the method for assessing the electrical impedance status of fused CT features as described above, wherein the system for assessing the electrical impedance status of fused CT features includes:

[0174] The CT structure perception module is used to perform CT scans on the subject to obtain CT scan data; and to generate anatomical mask data based on the CT scan data.

[0175] Impedance feature alignment module is used to acquire three-dimensional impedance data through an impedance scanning device; perform multi-scale decomposition feature extraction based on the three-dimensional impedance data to obtain decomposed feature data; and perform elastic harmonic network mapping based on the decomposed feature data and anatomical mask data to obtain configuration calibration impedance map data.

[0176] The structural guided impedance feature extraction module is used to extract regional impedance features from the configuration calibration impedance map data based on the anatomical mask data, and obtain regional impedance feature data.

[0177] The regional anomaly probability value calculation module is used to calculate the regional anomaly probability value based on the regional impedance characteristic data, and obtain the regional anomaly probability value data.

[0178] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0179] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method of fusion of CT features with electrical impedance state assessment, characterized by, The method comprises: performing CT scanning on a subject to obtain CT scanning data; and generating an anatomical mask based on the CT scanning data to obtain anatomical mask data; acquiring three-dimensional impedance data by means of an electrical impedance scanning device; extracting multi-scale decomposition features based on the three-dimensional impedance data to obtain decomposition feature data; and mapping the decomposition feature data and the anatomical mask data by means of an elastic harmonics network to obtain a shape-corrected impedance map data; extracting regional impedance feature data from the shape-corrected impedance map data based on the anatomical mask data; performing regional feature offset calculation based on the regional impedance feature data and a preset state template constructor to obtain regional feature offset data, wherein the state template constructor is a regional reference template generated by means of a multi-modal clustering and structural consistency modeling method based on an impedance feature distribution of prior normal samples; performing structural functional coordination weighted calculation on the regional feature offset data to obtain offset weighted data; and performing regional anomaly probability mapping on the anatomical mask data based on the offset weighted data to obtain regional anomaly probability value data.

2. The method of claim 1, wherein, The CT scanning comprises: performing CT scanning on a subject under a preset first radiation intensity condition and a preset second radiation intensity condition to obtain first CT scanning data and second CT scanning data, respectively; constructing a response difference map based on the first CT scanning data and the second CT scanning data to obtain response difference map data; performing structural fidelity region estimation on the response difference map data to obtain structural fidelity region data; performing structural gradient harmonization based on the structural fidelity region data to obtain the CT scanning data.

3. The method of claim 2, wherein, The structural fidelity region estimation comprises: performing dose intensity change rate calculation on the response difference map data to obtain response intensity change rate data; performing image gradient flow field calculation on the response difference map data based on the response intensity change rate data to obtain image gradient flow field data; performing phase disturbance entropy calculation based on the image gradient flow field data to obtain phase disturbance entropy data; performing non-convex consistency feature space projection based on the response intensity change rate data, the image gradient flow field data, and the phase disturbance entropy data to obtain the structural fidelity region data.

4. The method of claim 2, wherein, The structural gradient harmonization comprises: generating a structural gradient field map based on the structural fidelity region data to obtain structural gradient field map data; performing fidelity region structure diffusion weight calculation on the structural gradient field map data to obtain a fidelity region structure map data; performing structural gradient harmonization diffusion on the fidelity region structure map data to obtain the CT scanning data.

5. The method of claim 1, wherein, The anatomical mask generation comprises: constructing a multi-directional structural flow field map based on the CT scanning data to obtain structural flow field map data; performing directional consistency path extraction on the structural flow field map data to obtain directional consistency path data; constructing a path aggregation domain map based on the directional consistency path data to obtain the anatomical mask data.

6. The method of claim 1, wherein, The multi-scale decomposition feature extraction comprises: performing multi-resolution spherical harmonic transformation and three-dimensional non-uniform wavelet decomposition based on the three-dimensional impedance data to obtain first impedance decomposition data and second impedance decomposition data, respectively; According to the first impedance decomposition data, multi-scale contrast change rate calculation is performed to obtain contrast change rate data; According to the second impedance decomposition data, local response maximum level calculation is performed to obtain local response maximum level data; According to the contrast change rate data and the local response maximum level data, cross-scale trend calculation is performed to obtain decomposition feature data.

7. The method of claim 1, wherein, The elastic harmonic network mapping includes: According to the decomposition feature data and the anatomical mask data, functional-anatomical dual-map processing is performed to obtain functional-anatomical dual-map data; According to the functional-anatomical dual-map data, elastic harmonic configuration mapping is performed to obtain configuration mapping data; According to the configuration mapping data, topological conservation configuration transformation is performed to obtain configuration calibration impedance map data.

8. The method of claim 1, wherein, The regional impedance feature extraction includes: According to the anatomical mask data, impedance voxel clustering is performed on the configuration calibration impedance map data to obtain impedance voxel clustering data; According to the impedance voxel clustering data, topological feature extraction and local density feature extraction are performed to obtain topological feature data and local density feature data, respectively; According to the topological feature data and the local density feature data, impedance-structure joint graph construction is performed to obtain impedance-structure joint graph data; The structure principal modal impedance processing is performed on the impedance-structure joint graph data to obtain regional impedance feature data.

9. A fusion CT feature's electrical impedance state assessment system, characterized by, The system for performing the fusion CT feature electrical impedance state evaluation method as claimed in claim 1 includes: A CT structure perception module is configured to perform CT scanning on a subject to obtain CT scanning data, and to generate anatomical mask data according to the CT scanning data; An impedance feature alignment module is configured to acquire three-dimensional impedance data by an electrical impedance scanning device, to extract multi-scale decomposition features according to the three-dimensional impedance data to obtain decomposition feature data, and to perform elastic harmonic network mapping according to the decomposition feature data and the anatomical mask data to obtain configuration calibration impedance map data; A structure-guided impedance feature extraction module is configured to extract regional impedance features according to the anatomical mask data and the configuration calibration impedance map data to obtain regional impedance feature data; A regional anomaly probability value calculation module is configured to calculate regional anomaly probability values according to the regional impedance feature data to obtain regional anomaly probability value data.

Images