An intelligent segmentation system based on thoracic surgery CT image recognition

Abstract

This invention relates to the fields of medical image processing and artificial intelligence, specifically to an intelligent segmentation system based on thoracic CT image recognition. The system includes: a data acquisition step: acquiring target three-dimensional volumetric image data containing anatomical manifold prior constraints; a probability generation step: generating an initial segmentation probability field and a voxel-level uncertainty distribution map based on the original grayscale matrix; a topology calculation step: calculating the topological consistency entropy, representing the degree of deviation, using a topology data analysis algorithm; a strategy determination step: performing entropy value comparison and determining a dynamic adaptive repair strategy for high uncertainty regions; and a segmentation optimization step: reconstructing the initial segmentation probability field to output the target three-dimensional topological structure. This invention ensures the logical integrity of the vascular and bronchial tree structures in low signal-to-noise ratio environments, reducing surgical risks.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence, specifically to an intelligent segmentation system based on thoracic CT image recognition. Background Technology

[0002] In the application scenario of thoracic surgery CT image recognition, the intelligent segmentation system relies on high-precision three-dimensional volume data to ensure the logical integrity of anatomical structures such as bronchial trees and vascular trees. The system usually needs to process the original gray-level matrix to output a voxel-level segmentation probability field.

[0003] For segmentation of such complex anatomical structures, existing solutions generally adopt a direct pixel classification architecture. This involves acquiring CT image sequences through convolutional neural networks, performing end-to-end mapping using the grayscale features of local pixels, and directly using the model's output probability as the final segmentation result. Although this approach has a certain fitting ability under ideal image quality conditions, it relies excessively on the pixel intensity information of the data itself and lacks constraints from topological geometric rules. When encountering low signal-to-noise ratio scenarios such as interference from metal artifacts, respiratory motion deformation, or pathological structural distortions, the purely data-driven algorithm is prone to blindly fitting noisy pixel data, leading to misidentification of artifacts as tissue or physically impossible structures such as ruptured blood vessels or abnormal closure of the trachea due to signal attenuation.

[0004] Furthermore, traditional evaluation metrics based on pixel overlap rate are unable to detect such topological errors, and existing models lack quantitative perception and dynamic repair mechanisms for voxel-level uncertainties. This results in the inability to effectively retrieve anatomical manifold priors for completion in blurred image regions, making it difficult to support the stringent requirements of surgical navigation for structural connectivity. Therefore, how to establish a segmentation mechanism with topological self-correction capabilities, effectively identify high-uncertainty regions, introduce anatomical manifold prior constraints, and improve the logical correctness and robustness of anatomical structure segmentation under complex interference has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent segmentation system based on thoracic surgical CT image recognition. Specifically, the technical solution of this invention includes:

[0006] The data acquisition module is used to acquire the target three-dimensional volumetric image data to be processed. The target three-dimensional volumetric image data includes: the original grayscale matrix, the preset anatomical manifold prior constraints containing standard topological homology feature data, the preset topological feature benchmark value, and the preset entropy critical threshold.

[0007] The first processing module is used to generate an initial segmentation probability field and a voxel-level uncertainty distribution map based on the original gray-scale matrix and using a preset generative segmentation model.

[0008] The second processing module is used to calculate the topological consistency entropy, which characterizes the degree of deviation between the current segmentation result and the geometric rules of the manifold, based on the initial segmentation probability field and the preset anatomical manifold prior constraints, using a topological data analysis algorithm.

[0009] The third processing module is used to perform a numerical comparison between the topological consistency entropy and the preset entropy threshold, and to determine a dynamic adaptive repair strategy for the high uncertainty region in the voxel-level uncertainty distribution map based on the comparison result.

[0010] The segmentation optimization module is used to reconstruct the initial segmentation probability field based on the dynamic adaptive repair strategy and output a target three-dimensional topological structure that conforms to the preset topological feature benchmark value.

[0011] Preferably, based on the original grayscale matrix, an initial segmentation probability field and a voxel-level uncertainty distribution map are generated using a preset generative segmentation model, including: calling the original grayscale matrix; inputting the original grayscale matrix into a preset Bayesian neural network or Monte Carlo Dropout network; calculating the predicted mean of each voxel classification result through multiple forward propagation sampling to generate the initial segmentation probability field, and calculating the prediction variance to generate the voxel-level uncertainty distribution map.

[0012] Preferably, based on the initial segmentation probability field and the preset anatomical manifold prior constraints, a topological consistency entropy characterizing the degree of deviation between the current segmentation result and the geometric rules of the manifold is calculated using a topological data analysis algorithm. This includes: calling the initial segmentation probability field; calculating the persistent homology features corresponding to the initial segmentation probability field and extracting the current Betti number sequence; calculating the Bathystan distance between the current Betti number sequence and the standard Betti number sequence in the preset anatomical manifold prior constraints; and directly defining the value of the Bathystan distance as the topological consistency entropy, wherein the value of the topological consistency entropy is used to quantify the structural risk of connectivity breaks or anomalous closures.

[0013] Preferably, the third processing module determines a dynamic adaptive repair strategy for high uncertainty regions in the voxel-level uncertainty distribution map based on the comparison result between the topological consistency entropy and the preset entropy threshold. Specifically, it is configured to: determine that the system is in a high-risk state of topological failure when the topological consistency entropy is greater than the preset entropy threshold, and select a forced correction strategy based on prior shape completion as the dynamic adaptive repair strategy; and determine that the system is in a topologically stable state when the topological consistency entropy is less than or equal to the preset entropy threshold, and select a refined strategy based on pixel grayscale fitting as the dynamic adaptive repair strategy.

[0014] Preferably, based on the dynamic adaptive repair strategy, the initial segmentation probability field is reconstructed to output a target 3D topological structure that conforms to the preset topological feature benchmark value, including: when the dynamic adaptive repair strategy is the forced correction strategy, ignoring the original grayscale information of the high uncertainty region; using a preset manifold projection algorithm to forcibly map the initial segmentation probability field to the latent space where the preset anatomical manifold prior constraints are located; and using a generative shape completion network to generate a connected structure that fills the broken regions while maintaining the conservation of the Betti number, so as to output the target 3D topological structure.

[0015] Preferably, based on the dynamic adaptive repair strategy, reconstructing the initial segmentation probability field and outputting a target 3D topological structure that conforms to the preset topological feature benchmark value further includes: when the dynamic adaptive repair strategy is the refinement strategy, retaining the original grayscale information of the high uncertainty region; using a preset edge-weighted loss function to maximize the pixel-level intersection-union ratio between the initial segmentation probability field and the original grayscale matrix; and smoothing the local boundaries to output the target 3D topological structure.

[0016] Preferably, the target three-dimensional volumetric image data contains unstructured noise interference caused by metal artifacts or breathing motion; the first processing module is further configured to identify the spatial location of the unstructured noise interference in the original grayscale matrix and mark the spatial location as an extremely high uncertainty region in the voxel-level uncertainty distribution map.

[0017] Preferably, the preset anatomical manifold prior constraints include standard topological homology features of the target bronchial tree or vascular tree in the absence of pathological distortions; the standard topological homology features include at least zero-dimensional Betty numbers representing the number of connected components and one-dimensional Betty numbers representing the number of loops.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. This invention effectively solves the problem of blindly fitting noise in traditional pixel classification models under low signal-to-noise ratio environments by introducing topological consistency entropy as the core monitoring indicator and Bathystan distance calculation mechanism. Unlike conventional schemes that only focus on pixel overlap rate, this scheme uses the continuous cohomology algorithm to extract the Betty number sequence, quantifies the deviation of the current segmentation result from the prior constraints of the standard anatomical manifold, and can provide timely warnings when topological risks such as blood vessel rupture or abnormal adhesion are detected. This significantly improves the system's logical judgment ability in extreme scenarios such as metal artifacts or lung collapse, and avoids surgical risks caused by image misjudgment.

[0020] 2. This invention establishes a dynamic game-theoretic segmentation strategy based on uncertainty perception and topological risk assessment, breaking through the technical bottleneck that a single model cannot simultaneously consider overall connectivity and local accuracy. By generating a voxel-level uncertainty distribution map through Monte Carlo sampling, the system can automatically switch to a forced correction mode based on prior shape completion in a high-risk state of topological failure, or execute a refined mode based on pixel grayscale fitting in a stable state, based on the comparison results of topological consistency entropy and entropy threshold. This enables the algorithm to adaptively process different image quality regions, ensuring that optimal results can be output under different signal-to-noise ratio conditions.

[0021] 3. This invention employs generative shape completion and manifold projection reconstruction technology based on anatomical manifold prior constraints to achieve logical repair of image missing or severely artifact-prone areas. By mapping the initial probability field of high uncertainty regions to a preset low-dimensional latent space and using a generative network that maintains the conservation of the Betti number for decoding, this method can fill in connected tubular structures that conform to the logic of human anatomy while ignoring the original damaged grayscale information. This effectively prevents blood vessel rupture or the generation of unnatural structures due to signal attenuation and ensures the integrity of the surgical navigation path in extreme areas where data is completely invalid.

[0022] 4. This invention constructs an unstructured noise source localization mask and an edge-weighted fine-grained processing mechanism, endowing the system with robustness and high-precision depiction capabilities under complex interference. The system can identify metal artifacts or respiratory motion stripes and mark them as regions with extremely high uncertainty, explicitly preventing subsequent algorithms from incorrectly fitting artifact textures. At the same time, in topologically stable regions, the edge-weighted loss function and local boundary smoothing are used to maximize the pixel-level intersection-union ratio between the initial segmentation probability field and the original grayscale matrix, ensuring accurate segmentation of small bronchi and vascular terminals, meeting the needs of surgery for fine operation. Attached Figure Description

[0023] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0024] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0026] Example 1

[0027] Please see Figure 1 An intelligent segmentation system based on thoracic surgical CT image recognition includes:

[0028] The data acquisition module is used to acquire the target three-dimensional volumetric image data to be processed. The target three-dimensional volumetric image data includes: the original grayscale matrix, the preset anatomical manifold prior constraints containing standard topological homology feature data, the preset topological feature benchmark value, and the preset entropy critical threshold.

[0029] The first processing module is used to generate an initial segmentation probability field and a voxel-level uncertainty distribution map based on the original gray-level matrix and using a preset generative segmentation model.

[0030] The second processing module is used to calculate the topological consistency entropy, which characterizes the degree of deviation between the current segmentation result and the geometric rules of the manifold, based on the initial segmentation probability field and the preset anatomical manifold prior constraints, using a topological data analysis algorithm.

[0031] The third processing module is used to perform a numerical comparison between the topological consistency entropy and the preset entropy threshold, and to determine a dynamic adaptive repair strategy for high uncertainty regions in the voxel-level uncertainty distribution map based on the comparison results.

[0032] The segmentation optimization module is used to reconstruct the initial segmentation probability field based on a dynamic adaptive repair strategy, and output the target 3D topology that conforms to the preset topological feature benchmark values.

[0033] This embodiment details an intelligent segmentation system based on thoracic CT image recognition. Addressing the extreme challenges common in thoracic surgical navigation, such as metal artifacts, respiratory motion deformation, and pathological structural distortions, this system proposes a dynamic game-theoretic segmentation strategy combining uncertainty perception and topological data analysis. Its core lies in constructing a computational architecture with topological self-correction capabilities. The system acquires and preprocesses the underlying data required for surgical navigation through a data acquisition module. This module not only acquires the patient's chest CT scan sequence (i.e., the original grayscale matrix) but also simultaneously retrieves a set of mathematical constraints from the system memory, encapsulating both into composite target three-dimensional volumetric image data.

[0034] In the definition of this system, image data is constructed as a data object containing four core fields: the original grayscale matrix, which refers to the set of CT voxel values ​​without any smoothing or enhancement processing, retaining all details but also containing noise; the preset anatomical manifold prior constraint containing standard topological homology feature data, which refers to the low-dimensional manifold space obtained by training on a large number of normal and typical pathological lung structures, which defines the topologically allowed geometric forms of bronchial trees and vascular trees, i.e., the carrier of the set of standard topological homology features, such as tree-like fractal structures, excluding physically impossible structures such as suspended blood vessels or annular trachea;

[0035] The preset topological feature baseline value is a specific quantitative indicator extracted from the aforementioned prior constraints and used for final output verification; that is, a specific integer target value, such as... Although it originates from the standard topological homology feature in the prior constraints, it is stored independently in the data structure so that the partitioning optimization module can directly call it as the convergence target;

[0036] A preset entropy threshold is a warning line used to determine whether topological collapse has occurred in the segmentation result; this data encapsulation method ensures the integrated transmission of image data and verification standards; the first processing module performs preliminary perception and quantifies the credibility of the data, and uses a preset generative segmentation model, such as a variant of 3DU-Net, to output an initial segmentation probability field and a voxel-level uncertainty distribution map based on the original gray-scale matrix. The former represents the probability that each voxel belongs to a specific anatomical structure, and the latter quantifies the degree of hesitation of the model in classifying each voxel through Bayesian inference;

[0037] The second processing module evaluates the logical correctness of the segmentation results from a mathematical topological perspective, and calculates the topological consistency entropy using topological data analysis algorithms. This index is used to characterize the degree of deviation of the currently segmented 3D object from standard anatomical rules in terms of connectivity and the number of holes. Based on this, the third processing module dynamically switches the repair strategy according to the risk level and performs numerical comparison to determine whether to accept the data or the prior rules. The segmentation optimization module performs specific repair operations and outputs the final target 3D topological structure that conforms to the preset topological feature benchmark values.

[0038] By introducing topological consistency entropy as the core monitoring indicator, this system no longer blindly fits noisy pixel data in low signal-to-noise ratio environments, such as metal artifacts or lung collapse scenarios. Instead, when topological risks are detected, it automatically switches to a generative completion mode based on manifold priors. This ensures the logical integrity of the vascular and bronchial tree structures in high-risk scenarios of intraoperative navigation, effectively reducing the surgical risks caused by image misjudgment.

[0039] Example 2

[0040] Based on the original grayscale matrix, an initial segmentation probability field and a voxel-level uncertainty distribution map are generated using a pre-defined generative segmentation model, including:

[0041] Call the original grayscale matrix;

[0042] Input the original grayscale matrix into a preset Bayesian neural network or Monte Carlo Dropout network;

[0043] By sampling through multiple forward propagation, the predicted mean of each voxel classification result is calculated to generate an initial segmentation probability field, and the prediction variance is calculated to generate a voxel-level uncertainty distribution map.

[0044] This embodiment details the process by which the first processing module generates the initial segmentation probability field and the voxel-level uncertainty distribution map. To simultaneously obtain the segmentation result and its confidence level in a single inference iteration, this embodiment introduces a Monte Carlo Dropout sampling mechanism. The system calls the original grayscale matrix and performs normalization preprocessing on it: using a preset lung window width and level (e.g., window width 1500, window level -600), the grayscale values ​​are truncated, and the truncated values ​​are linearly mapped to... The range is adjusted to fit the dynamic range of the neural network's input.

[0045] The normalized input is then fed into a pre-defined Bayesian neural network or a regular network with Dropout layers enabled; while keeping the input unchanged, the network's random deactivation function is enabled, and... Second forward propagation; calculate the statistic using the following formula: ; ; in, Voxel representation The predicted mean, i.e. the probability value in the initial segmentation probability field; Voxel representation The prediction variance is derived from the statistical dispersion of multiple samplings; This represents the total number of forward propagation samples; This represents the original grayscale matrix as input; Indicates the first The network parameter state at the time of the next sampling; Representing neural networks Output probability;

[0046] Calculated mean The variance is calculated as the final segmentation probability. As a measure of uncertainty;

[0047] This method not only provides segmentation results, but also accurately locates regions that the model cannot accurately identify through variance, which usually correspond to artifacts or blurred boundaries. This provides pixel-level confidence for subsequent modules to determine whether forced correction needs to be initiated, avoiding blind trust in low-quality image data.

[0048] Example 3

[0049] Based on the initial segmentation probability field and pre-defined anatomical manifold prior constraints, a topological consistency entropy, representing the degree of deviation between the current segmentation result and the geometric rules of the manifold, is calculated using a topological data analysis algorithm. This includes:

[0050] Call the initial segmentation probability field; calculate the persistent homology feature corresponding to the initial segmentation probability field, and extract the current Betti number sequence;

[0051] Calculate the Bathurstein distance between the current Betty number sequence and the standard Betty number sequence in the preset anatomical manifold prior constraints;

[0052] The Basestein distance is directly defined as the topological consistency entropy, where the magnitude of the topological consistency entropy is used to quantify the structural risk of connectivity breaks or anomalous closures.

[0053] This embodiment details the specific process by which the second processing module calculates the topological consistency entropy, which is the core step in identifying structural risks in this system. To quantify the degree of anomaly in the topological structure of the current segmentation result, this embodiment introduces topological consistency entropy based on Bathystan distance. It is important to note that, regarding the calculation process, the Betti number sequence defined here is not simply a static list of integers, but rather refers to a sequence of key critical parameters where the Betti numbers of each dimension change during continuous homology filtering. In the mathematical essence of topological data analysis, the information carried by this sequence is mapped one-to-one with the point pairs between feature generation and extinction times. The resulting continuous homology graph;

[0054] Therefore, calculating the Bathysstein distance between sequences of Betti numbers, in a specific algorithmic implementation, involves calculating their corresponding persistent homology graphs, i.e., the Bathysstein distances between point sets, thereby resolving the mapping relationship between the terminology description and the algorithm's input data structure; the calculation method is as follows: ; in, Represents topological consistency entropy, used to quantify the risk of connectivity breaks or anomalous closures; This represents the persistent homology graph corresponding to the current Betty number sequence, containing features. This originates from the initial segmentation probability field; The reference persistent homology graph corresponding to the standard Betti number sequence is derived from the pre-defined anatomical manifold prior constraints. To address the data type matching problem, i.e., how to calculate the distance between the point set and the scalar, this embodiment clarifies the method based on the scalar Betti number benchmark value. To two-dimensional point set Construction mapping rules: for standard Betti numbers If its target value is an integer Then in The Construction in dimensional graphs The coordinates are The ideal feature point, where, This is the normalized maximum filter value, for example, 1.0, representing an infinitely surviving topological feature; if the target value is 0, then... It is an empty set; for example, for a simply connected structure. , This makes the two inputs in the Bathurst distance formula... and All contain specific The coordinates are a two-dimensional point set, thus eliminating type errors and making distance calculations mathematically feasible; express and The set of all possible joint distributions between them.

[0055] To eliminate mathematical ambiguity, the symbol here... Strictly defined as all those who... and Let be the joint probability measure space of the marginal distributions, i.e. This is known as the transmission plan set in optimal transmission theory; This represents the metric space in which the persistent homology feature resides, specifically referring to the extended two-dimensional upper half-plane. ,in diagonal In order to solve and Inconsistent base values ​​lead to transmission schedule The problem of being empty, for example when For the empty set When noisy points are present, this embodiment performs a point set augmentation step before computation to balance the quality and construct an augmented set: and ; in, This indicates projecting the point set onto the diagonal. The operation; this step ensures This allows the joint distribution to exist, thus enabling the elimination of redundant topological noise features at the cost of their distance to the diagonal, i.e., the noise is considered to be absorbed. This represents the integration region, i.e., the Cartesian product space containing the source feature distribution and the target feature distribution; Representing feature points and The Euclidean distance between them; The order of the distance is represented; in this embodiment, it is specifically set to 2, i.e., 2-Wasserstein distance, to apply a quadratic penalty to larger geometric deformations; if set... This corresponds to the bulldozer distance;

[0056] The system calls the initial segmentation probability field; in this step, to ensure that the algorithm captures the foreground target, i.e., the connectivity of blood vessels or bronchi, rather than the hollow structure of the background, the system must perform super-level set filtering or pre-invert the probability field. Afterwards, conventional lower level set filtering is applied; this embodiment preferably uses the probability field inversion method, which maps probability values ​​to pseudo-potential energy values, so that high-probability voxels correspond to low-potential energy points. Therefore, it is consistent with the benchmark value. middle The feature definitions are matched; the processed probability field is scanned using a continuous cohomology algorithm, and the process of the appearance and disappearance of topological features is recorded by changing the probability threshold, generating... Extract the current Betty number sequence information, i.e., the point set in the graph, and construct a reference point set according to the above rules. ;calculate with standard anatomical template The distance between them in Bathurst;

[0057] To translate the mathematical theory of the above definition of continuous integrals into computer-executable code, this embodiment specifically uses the discrete bipartite graph minimum weight perfect matching algorithm for calculation, or first takes the negative value of the cost matrix ( Then call the maximum weight matching algorithm: given and For a discrete set of points, the system constructs a cost matrix. , among which, element The Euclidean distance between point pairs The power, including the diagonal projection points; call the Hungarian algorithm or The algorithm solves for the optimal transfer matrix by transforming integration operations into solving... This is a combinatorial optimization problem; the discretization solution step ensures the accurate evaluation of the Bathystan distance in a digital computing environment; the distance value is directly defined as... ;

[0058] Traditional metrics based on pixel overlap rate cannot detect topological errors, such as a break in the middle of a blood vessel by a single pixel. The coefficients remain almost unchanged, but the clinical significance is completely different. This embodiment uses topological consistency entropy to directly measure the differences in topological structure, which can keenly capture minute breaks or abnormal adhesions, thus providing a mathematical basis for the system's topological self-correction capability.

[0059] Example 4

[0060] The third processing module determines a dynamic adaptive repair strategy for high-uncertainty regions in the voxel-level uncertainty distribution map based on the comparison between the topological consistency entropy and the preset entropy threshold. Specifically, this strategy is configured as follows:

[0061] In response to the topology consistency entropy being greater than the preset entropy threshold, the system is determined to be in a high-risk state of topology failure, and a forced correction strategy based on prior shape completion is selected as the dynamic adaptive repair strategy.

[0062] In response to the topological consistency entropy being less than or equal to a preset entropy threshold, the system is determined to be in a topologically stable state, and a refined strategy based on pixel grayscale fitting is selected as the dynamic adaptive repair strategy.

[0063] This embodiment details the logic of the third processing module determining the dynamic adaptive repair strategy based on topological consistency entropy; to achieve a balance between fidelity and reasonableness, this embodiment introduces a state determination function. : ; in, This indicates the current repair strategy selection result of the system; This represents the topological consistency entropy calculated in the preceding steps; This represents a preset entropy threshold value, derived from statistics of a large-scale pathological database; to clarify the statistical definition of this parameter, this embodiment specifies... The specific calculation formula is as follows: ; in, This is the set of all topological consistency entropy values ​​calculated from a database containing at least 1000 lung CT images without pathological abnormalities; or it can be defined using quantiles, i.e. This statistical approach ensures that the system is only considered to be topologically ineffective when the entropy value exceeds the normal range of anatomical variation, i.e., outside the 99% confidence interval, thus determining the inversion point of the control flow logic. This indicates a forced correction strategy based on prior shape completion; This represents a refined strategy based on pixel grayscale fitting;

[0064] The system performs numerical comparison; in response to In other words, if the topological differences are too large, such as detecting a ruptured main bronchus, the system determines that it is in a high-risk state of topological failure, considering the original image data to be severely damaged or subject to extremely strong interference, and selects... Prepare to restructure; respond to In other words, the topology is basically normal, the system determines that it is in a topologically stable state, and considers the original image data reliable. It emphasizes precise edge depiction;

[0065] This strategy enables dynamic game theory within the algorithm; under low-risk conditions, the system functions as a high-precision segmentation tool; under high-risk conditions, such as severe artifacts leading to blood vessel rupture, the system automatically transforms into a generative AI with associative capabilities, prioritizing the connectivity of anatomical structures and preventing misleading surgical navigation due to image quality issues.

[0066] Example 5

[0067] Based on a dynamic adaptive repair strategy, the initial segmentation probability field is reconstructed, outputting a target 3D topological structure that conforms to a preset topological feature benchmark value, including:

[0068] When the dynamic adaptive repair strategy is a forced correction strategy, the original grayscale information of the high uncertainty region is ignored.

[0069] Using a pre-defined manifold projection algorithm, the initial segmentation probability field is forcibly mapped to the latent space containing the pre-defined anatomical manifold prior constraints;

[0070] By using a generative shape completion network, while maintaining the conservation of the Betti number, a connected structure is generated to fill in the broken regions, so as to output the target 3D topology.

[0071] This embodiment details the workflow of the segmentation optimization module when a forced correction strategy is employed. To restore reasonable anatomical structures in the event of missing image information, this embodiment introduces a manifold projection reconstruction algorithm. The specific mapping model is as follows: ; ; in, Representing the potential space The optimal latent vector in the space; to ensure the feasibility of gradient calculation, this embodiment explicitly specifies the latent space variables. The vector dimension is Before executing the numerical optimization loop, the system utilizes the encoder in the pre-defined anatomical manifold prior constraint model. Region of interest Perform forward inference to obtain the initial latent vector. and will Initialize to This ensures that the search process begins with an effective neighborhood on the manifold; solving The numerical optimization process uses the Adam optimizer, with the initial learning rate set to 1. To prevent the program from entering an infinite loop, this embodiment strictly defines the iteration stopping condition: the optimization loop will terminate when any of the following conditions are met: the number of iterations reaches the upper limit. The change in the objective function value between two consecutive iterations. Less than the convergence threshold ;

[0072] It represents the low-dimensional latent space where the pre-defined anatomical manifold prior constraints reside, derived from a variational autoencoder trained on undistorted anatomical data; This represents a pre-defined generative shape completion network; this embodiment discloses this network. Specific architectural parameters to support computation graph construction: A 3DVAE decoder structure is adopted; to achieve the dimensionality transformation from one-dimensional latent vectors to three-dimensional feature maps, the network head contains a key connection layer: a fully connected layer that transforms the input 256-dimensional latent vectors... Mapped to A dimensional feature vector; a reshaping layer that transforms the feature vector into a shape of... The three-dimensional feature tensor, i.e., number of channels × depth × height × width;

[0073] This tensor is fed into 5 transposed convolutional layers, with each convolutional kernel having a size of [size missing]. The step size is 2, and the padding is 1; except for the output layer, each layer is followed by and Activation function; the output layer uses Activation function to constrain output values ​​to interval, thereby ensuring The gradient is numerically stable during backpropagation; this leads to The output size is fixed as ;

[0074] Represents the initial segmentation probability field; for the generator network The output dimension is fixed. The original image and initial segmentation probability field usually have the following characteristics: The large size leads to dimensional mismatch issues, which are explicitly addressed in this embodiment. Preprocessing steps: After performing the above subtraction operation Previously, the system located the center of the high-uncertainty region, cropped out the region of interest, and used a trilinear interpolation algorithm to resample the size of the ROI to [the desired size]. This spatial normalization step ensures and Strict alignment in the tensor dimension makes the subtraction operations and norm calculations in the formula physically feasible; The weighted norm is expressed by the following formula: ;

[0075] To ensure the completeness of the mathematical definition, the region of high uncertainty is explicitly defined here as a set. Variance in the voxel-level uncertainty distribution plot Exceeding the threshold The set of voxels; accordingly, the weights The assignment rule is: when voxel hour ,otherwise ;Preset threshold here It is not a fixed constant, but a dynamic threshold determined based on statistical distribution; specifically, the system calculates the first _th ... Percentiles, for example Set it to Or adopt This mechanism ensures that only statistically significant outliers with high uncertainty, such as centers of metal artifacts, are completely ignored. The topological constraint loss is represented by the following formula: ; Among them, function symbols Defined as a mapping function from a three-dimensional structure to its topological invariants, i.e. ,in Indicates the first Homophonic groups, For voxel-based structures The constructed cubic complex; it quantifies the pores in the structure, i.e. , and connected components, i.e. The quantity; it should be clarified here that the formula... This refers to the preset topological feature baseline value, for example... To address the gradient vanishing problem (i.e., non-differentiability) caused by the Betti number being a discrete integer, this embodiment employs a differentiable topological layer technique for specific implementation; in the above formula... In actual computation, it is replaced by the feature persistence loss of the persistent homology graph.

[0076] System Calculation The continuous homology graph is used to sort the lifecycles of feature points in the graph from largest to smallest, and the top ones are selected. One feature is treated as the signal, and the rest as noise; the loss function is defined as the sum of the squares of the lifetimes of the noise features plus the distance between the signal features and the ideal lifetime; due to the feature points and The value directly corresponds to the critical pixel value in the voxel grid, which affects the input. The derivative of is non-zero, thus ensuring the computability and convergence of the gradient in the backpropagation algorithm;

[0077] Regularization coefficient This embodiment specifies that the range of values ​​for this coefficient is as follows: The specific value depends on the signal-to-noise ratio of the input image; to achieve adaptive adjustment, this embodiment also discloses a dynamic calculation formula based on gradient balance: ; in, To prevent division by zero; in the above formula This indicates that the loss function is calculated with respect to the latent vector. gradient vector Norm; the dynamic weight This mechanism ensures that in each iteration, the gradient update step size generated by the topology constraint term and the gradient update step size generated by the data fitting term are numerically on the same order of magnitude, preventing optimization deviations caused by excessive differences in gradient values. This ensures that the gradient contributions of the topology constraint term and the data fitting term remain on the same order of magnitude during optimization, effectively preventing... Too large an image may cause the image details to be ignored, while too small an image may result in ineffective topology restoration. This represents the final output target 3D topology.

[0078] The system identifies high-uncertainty regions, such as areas covered by metal artifacts, and reduces the pixel weights of these regions to zero during the optimization process; it also utilizes a manifold projection algorithm to search for potential spaces. The latent vector that best matches the current remaining reliable structure Using generators decoding While maintaining the conservation of the Betty number, i.e. without generating new breakpoints or loops, a connected structure is generated to fill the broken regions. The repaired feature map output by the generator is inversely resampled back to the spatial dimensions of the original region of interest (ROI) using a trilinear interpolation algorithm. The inversely sampled repaired data is then backfilled to the corresponding coordinate positions of the original target 3D volumetric image data, covering the original high-uncertainty regions, thereby generating a complete target 3D topology through spatial fusion. To ensure the effectiveness and reproducibility of this generation process, this embodiment further discloses the training details of the preset anatomical manifold prior constraint model, i.e., the VAE.

[0079] In addition to the aforementioned decoder In addition to the architecture, the accompanying encoder adopts a mirror structure: it contains 5 three-dimensional convolutional layers with channel numbers of 32, 64, 128, 256, and 512 respectively, and convolutional kernels... Step size 2, padding 1; finally, the mean vector is output through a fully connected layer. Sum of logarithmic variance vector All dimensions are 256;

[0080] Loss function used in model training It consists of a weighted sum of reconstruction error and KL divergence: The training dataset contains 5000 expert-reviewed lung CT slices without pathological abnormalities; the training convergence criterion is set as follows: under the Adam optimizer, the validation set... 20 consecutive indicators The improvement was less than This training process ensures potential space. It can compactly encode normal anatomical morphology, thus providing an efficient search space for manifold projection algorithms;

[0081] This scheme uses a deep generative model to fill in the anatomical structures obscured by artifacts; because the generation process is strictly limited to the anatomical manifold space. Therefore, the generated structure must conform to the logic of human anatomy. For example, it must generate continuous tubular structures to connect breakpoints, thus achieving a survival-first navigation guarantee in extreme areas where data is completely invalid.

[0082] Example 6

[0083] Based on a dynamic adaptive repair strategy, the initial segmentation probability field is reconstructed, and the target 3D topological structure that conforms to the preset topological feature benchmark values ​​is output. This also includes:

[0084] When the dynamic adaptive repair strategy is a refined strategy, the original grayscale information of the high uncertainty area is preserved.

[0085] By using a preset edge-weighted loss function, the pixel-level intersection-union ratio between the initial segmentation probability field and the original grayscale matrix is ​​maximized; local boundaries are smoothed to output the target 3D topology.

[0086] This embodiment details the workflow of the segmentation optimization module when a refined strategy is employed. To improve boundary accuracy while maintaining topological correctness, this embodiment introduces an edge-weighted intersection-union ratio (IUU) loss function. : ; in, This represents the edge-weighted loss function; Voxel representation The initial segmentation probability field value; Voxel representation The corresponding reference value of the original grayscale matrix after thresholding; The spatial weight map is generated based on the gradient magnitude of the original grayscale matrix. The weighting rule is as follows: Extremely high values ​​are assigned in boundary regions with significant gradient magnitudes, specifically ranging from 10 to 50; a value of 1 is assigned in flat regions to enhance edge alignment. To clarify the variables... arrive The specific mapping relationship is disclosed in this embodiment. The specific function expression: ; Wherein, the constant 49 corresponds to Ensure the weight cap is 50; threshold Defined as the gradient magnitude matrix Adaptive threshold or average; The function continuously maps the gradient magnitude to The interval ensures the determinism and reproducibility of the calculation process;

[0087] The system retains the original grayscale information of high-uncertainty regions, assuming these fluctuations are genuine tissue textures rather than artifacts; it calculates the aforementioned loss function, focusing on maximizing the pixel-level intersection-over-union ratio (IoU) of the boundary regions; it smooths local boundaries, for example, using conditional random fields (CRFs) to remove glitches and output a high-precision 3D topological structure of the target; to make the CRF steps translatable into concrete code, this embodiment defines the energy function of the fully connected CRF. : ; Among them, the univariate potential function It is directly taken from the initial segmentation probability field; the binary potential function Using Gaussian kernel linear combination form: ;

[0088] To eliminate symbolic ambiguity, voxel space coordinates are specifically defined as follows: Unlike the aforementioned representation of probability, , This represents the voxel grayscale value. This is a tag compatibility function, specifically adopted in this embodiment. Model, i.e., when hour Otherwise, it is 0, used to penalize adjacent voxels with different labels; the parameter configuration is: appearance kernel weight. Smoothing kernel weights Scale parameters For cases where the original grayscale matrix consists of unenhanced CT voxel values, typically ranging from -1000 to +3000 HU, the above... If the parameter settings are applied directly to the raw data, it will cause the appearance kernel to saturate prematurely;

[0089] Therefore, this embodiment explicitly stipulates that before calculating the CRF energy function, the input grayscale matrix must be processed. Perform normalization preprocessing to map its values ​​to For example, a lung window setting of 1500 for width and -600 for truncation of linear mapping can be used; or, if the original HU value is used directly for calculation, then it is necessary to... The value was adjusted to 100 to match the dynamic range of the CT data and ensure the effectiveness of edge smoothing; this energy function clarifies the calculation method of the connection weights between nodes, allowing the smoothing process to be minimized. Solve this problem;

[0090] Under stable topological conditions, this strategy makes full use of the detailed information in the original images and achieves precise segmentation of tiny bronchi and vascular terminals through edge weighting technology, meeting the needs of surgery for delicate operations.

[0091] Example 7

[0092] The target's 3D volumetric image data contains unstructured noise interference caused by metal artifacts or breathing motion;

[0093] The first processing module is also used to identify the spatial location of unstructured noise interference in the original gray-scale matrix and mark the spatial location as an extremely high uncertainty region in the voxel-level uncertainty distribution map;

[0094] This embodiment details the processing mechanism for unstructured noise interference. To prevent the system from being misled by strong noise, this embodiment introduces a noise source localization mask. The first processing module uses a preset threshold segmentation method for high-density metal artifacts, or a frequency domain analysis method for breathing motion stripes, to identify unstructured noise interference. It determines the specific spatial location of these interferences in the original grayscale matrix and forcibly marks these locations as extremely high uncertainty regions in the voxel-level uncertainty distribution map, for example, by adjusting the variance... Set to the maximum value;

[0095] This step is equivalent to putting a filter on the system, explicitly telling subsequent algorithms which areas are absolutely unreliable; this can prevent subsequent repair strategies, especially refinement strategies, from incorrectly fitting artifact textures, ensuring the robustness of the system under harsh imaging conditions.

[0096] Example 8

[0097] The pre-defined anatomical manifold prior constraints include the standard topological homology features of the target bronchial tree or vascular tree in the absence of pathological distortions.

[0098] Standard topological homology features include at least zero-dimensional Betti numbers, which characterize the number of connected components, and one-dimensional Betti numbers, which characterize the number of loops.

[0099] This embodiment details the specific content of the preset anatomical manifold prior constraints; to enable the computer to understand what a normal lung is, this embodiment defines standard topological homology features. : ; in, The zero Wibbetti number is used to characterize the number of connected components. For a target bronchial tree without pathological abnormalities, it should theoretically be always equal to 1, i.e., a simply connected structure. The number represents a one-dimensional Betty number, used to characterize the number of loops. For a normal bronchial tree, it should theoretically be equal to 0, meaning there are no unnatural closed loops.

[0100] During system initialization, the system loads a configuration file containing the aforementioned Betti number constraints; when calculating topological consistency entropy, any factors that lead to... That is, breakage, or That is, the segmentation results of abnormal loops will all lead to a sharp increase in entropy;

[0101] By introducing the Betti number from algebraic topology as a hard constraint, this invention transforms complex anatomical knowledge into a mathematically invariant that can be computed by a computer; this ensures that no matter how the image is deformed, the final model generated by the system always remains a tree in mathematical properties, thus solving the problem of generating unnatural structures.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent segmentation system based on thoracic surgical CT image recognition, characterized in that, include: The data acquisition module is used to acquire the target three-dimensional volumetric image data to be processed. The target three-dimensional volumetric image data includes: the original grayscale matrix, the preset anatomical manifold prior constraints containing standard topological homology feature data, the preset topological feature benchmark value, and the preset entropy critical threshold. The first processing module is used to generate an initial segmentation probability field and a voxel-level uncertainty distribution map based on the original gray-scale matrix and using a preset generative segmentation model. The second processing module is used to calculate the topological consistency entropy, which characterizes the degree of deviation between the current segmentation result and the geometric rules of the manifold, based on the initial segmentation probability field and the preset anatomical manifold prior constraints, using a topological data analysis algorithm. The third processing module is used to perform a numerical comparison between the topological consistency entropy and the preset entropy threshold, and to determine a dynamic adaptive repair strategy for the high uncertainty region in the voxel-level uncertainty distribution map based on the comparison result. The segmentation optimization module is used to reconstruct the initial segmentation probability field based on the dynamic adaptive repair strategy and output a target three-dimensional topological structure that conforms to the preset topological feature benchmark value.

2. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 1, characterized in that, Based on the original grayscale matrix, an initial segmentation probability field and a voxel-level uncertainty distribution map are generated using a preset generative segmentation model, including: Call the original grayscale matrix; The original grayscale matrix is ​​input into a preset Bayesian neural network or Monte Carlo Dropout network; By sampling through multiple forward propagation steps, the predicted mean of each voxel classification result is calculated to generate the initial segmentation probability field, and the prediction variance is calculated to generate the voxel-level uncertainty distribution map.

3. The intelligent segmentation system based on thoracic CT image recognition according to claim 1, characterized in that, Based on the initial segmentation probability field and the preset anatomical manifold prior constraints, a topological consistency entropy, representing the degree of deviation between the current segmentation result and the geometric rules of the manifold, is calculated using a topological data analysis algorithm, including: Invoke the initial segmentation probability field; Calculate the persistent homology features corresponding to the initial segmentation probability field and extract the current Betty number sequence; Calculate the Bartherstein distance between the current Betty number sequence and the standard Betty number sequence in the preset anatomical manifold prior constraints; The value of the Basestein distance is directly defined as the topological consistency entropy, wherein the magnitude of the topological consistency entropy is used to quantify the structural risk of connectivity breaks or anomalous closures.

4. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 3, characterized in that, The third processing module determines a dynamic adaptive repair strategy for high-uncertainty regions in the voxel-level uncertainty distribution map based on the comparison between the topological consistency entropy and the preset entropy threshold. Specifically, this strategy is configured as follows: In response to the topology consistency entropy being greater than the preset entropy threshold, the system is determined to be in a high-risk state of topology failure, and a forced correction strategy based on prior shape completion is selected as the dynamic adaptive repair strategy. In response to the topological consistency entropy being less than or equal to the preset entropy threshold, the system is determined to be in a topologically stable state, and a refined strategy based on pixel grayscale fitting is selected as the dynamic adaptive repair strategy.

5. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 4, characterized in that, Based on the dynamic adaptive repair strategy, the initial segmentation probability field is reconstructed to output a target 3D topological structure that conforms to the preset topological feature benchmark value, including: When the dynamic adaptive repair strategy is the forced correction strategy, the original grayscale information of the high uncertainty region is ignored; Using a preset manifold projection algorithm, the initial segmentation probability field is forcibly mapped to the latent space where the preset anatomical manifold prior constraints are located; By using a generative shape completion network, while maintaining the conservation of the Betti number, a connected structure is generated to fill in the broken regions, so as to output the target three-dimensional topology.

6. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 4, characterized in that, Based on the dynamic adaptive repair strategy, the initial segmentation probability field is reconstructed to output a target 3D topological structure that conforms to the preset topological feature benchmark value, and the method further includes: When the dynamic adaptive repair strategy is the refined strategy, the original grayscale information of the high uncertainty region is preserved; By using a preset edge-weighted loss function, the pixel-level intersection-over-union ratio between the initial segmentation probability field and the original grayscale matrix is ​​maximized; The local boundaries are smoothed to output the target three-dimensional topology.

7. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 1, characterized in that, The target 3D volumetric image data contains unstructured noise interference caused by metal artifacts or breathing motion; The first processing module is also used to identify the spatial location of the unstructured noise interference in the original grayscale matrix, and to mark the spatial location as an extremely high uncertainty region in the voxel-level uncertainty distribution map.

8. The intelligent segmentation system based on thoracic surgical CT image recognition according to claim 5, characterized in that, The preset anatomical manifold prior constraints include the standard topological homology features of the target bronchial tree or vascular tree in the absence of pathological distortions. The standard topological homology features include at least zero-dimensional Betty numbers, which characterize the number of connected components, and one-dimensional Betty numbers, which characterize the number of loops.

Images