T12 thoracic vertebra positioning and muscle segmentation method and skeletal muscle index fitting system

Abstract

This invention relates to a method for T12 thoracic vertebra localization and muscle segmentation, and a skeletal muscle index fitting system, belonging to the field of medical image processing technology. It solves the problem that existing technologies cannot automatically locate the T12 thoracic vertebra and segment the thoracic and abdominal wall muscles from chest CT plain images to obtain reliable quantitative indicators for skeletal muscle index. The method includes acquiring standardized CT volume data, extracting a set of candidate vertebral center points, and generating a distance feature matrix; constructing a weighted directed graph, and using a path relaxation-based iterative global optimal search strategy to solve for the optimal vertebral sequence path in the weighted directed graph, thereby obtaining the T12 thoracic vertebra level index and corresponding level image; constructing a multi-channel input, using a muscle segmentation model to obtain a segmentation probability mask for the thoracic and abdominal wall muscles at the T12 level, and then fitting the skeletal muscle index through a regression model. This achieves automatic T12 thoracic vertebra localization, thoracic and abdominal wall muscle segmentation, and integrated skeletal muscle index prediction.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for locating the T12 thoracic vertebra and segmenting muscles, and a skeletal muscle index fitting system. Background Technology

[0002] Sarcopenia, a common comorbidity in patients with chronic obstructive pulmonary disease (COPD), has a prevalence ranging from 21.6% to over 56% in the COPD population and is significantly associated with decreased quality of life, poor prognosis, and increased mortality. The skeletal muscle index (SMI) is an internationally recognized core indicator for the quantitative assessment of sarcopenia, traditionally obtained by measuring the cross-sectional area of ​​skeletal muscle using L3 lumbar CT images and correcting for height. However, L3 abdominal CT scans are not routine examinations for COPD patients; chest CT scans are the standard imaging examination for diagnosis and follow-up, creating an irreconcilable mismatch in imaging modality. In recent years, the T12 thoracic spine level has been confirmed by multiple clinical studies as an effective alternative level for assessing sarcopenia in COPD patients. However, current technology lacks a systematic solution capable of fully automating T12 localization, muscle segmentation, and SMI calculation using chest CT, making it difficult to translate these clinical research findings into a large-scale, automated screening tool.

[0003] In spinal segmentation and vertebral localization, existing technologies primarily employ deep learning models for vertebral probability prediction, combined with graph optimization algorithms or dynamic programming methods to solve for the optimal vertebral sequence, thereby achieving automatic labeling of each vertebra. Among these methods, Dijkstra's algorithm and its variants are the most commonly used shortest path solution strategies. These approaches typically aim for complete identification of all vertebrae rather than specific optimization for any particular vertebra. In muscle segmentation, U-Net and its variants are widely used for automatic delineation of skeletal muscle tissue in CT images. Existing research has achieved high segmentation accuracy at both the L3 and T12 levels. However, these segmentation tools often rely on manual specification of the target level or are embedded in commercial systems that only support the L3 level. In SMI calculation, all known automatic assessment systems use a direct ratio method—muscle area divided by the square of height—to calculate SMI. This method fixes the relationship between muscle and height as a strict square relationship, failing to incorporate equally available demographic features related to muscle mass, such as weight, age, and gender.

[0004] In summary, current technologies lack fully automated localization capabilities at the T12 level. While spinal segmentation technology can output complete vertebral sequences, robust identification of the T12 level under anatomical variations (such as transitional vertebrae and compression fractures) remains a gap. Furthermore, existing T12 muscle segmentation tools all require manual input or correction of the level position, severely limiting automated processing capabilities in large-scale screening scenarios. Current technologies also lack end-to-end integrated system solutions. Spinal segmentation and muscle segmentation are developing independently, and there is still no integrated process that seamlessly connects precise T12 localization, automatic segmentation of thoracic and abdominal wall muscles, and SMI intelligent calculation. In addition, the SMI calculation methodology is simplistic and inherently biased. All existing systems use the direct ratio method, which assumes a strict square relationship between muscle mass and height. This simplified model lacks sufficient anatomical basis and fails to utilize multidimensional information clearly related to muscle mass, such as weight, sex, and age. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a method for T12 thoracic vertebra localization and muscle segmentation, as well as a skeletal muscle index fitting system, to solve the problem in the prior art that it is impossible to automatically and accurately complete T12 thoracic vertebra localization and thoracic and abdominal wall muscle segmentation from existing chest CT plain scan images of COPD patients in order to obtain reliable quantitative indicators for opportunistic screening of sarcopenia.

[0006] On one hand, embodiments of the present invention provide a method for locating and segmenting the muscles of the T12 thoracic vertebra, comprising the following steps: Acquire chest CT plain scan images and preprocess them to generate standardized CT volume data; Based on the standardized CT volume data, vertebral probabilistic volume data is generated, and then a set of candidate vertebral center points is extracted; based on the spatial distance distribution between adjacent candidate center points in the set of candidate vertebral center points, a distance feature matrix is ​​generated. A weighted directed graph is constructed based on the set of candidate vertebral center points and the distance feature matrix. The optimal vertebral sequence path is solved in the weighted directed graph using a global optimal search strategy based on path relaxation iteration. Then, the T12 thoracic vertebral level index and corresponding level image are obtained based on the optimal vertebral sequence path. A multi-channel input is constructed based on the T12 thoracic vertebral level index and corresponding level images, and then input into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. After post-processing optimization, the final muscle segmentation mask is obtained.

[0007] Furthermore, each candidate node in the set of candidate vertebrae center points includes a vector of three-dimensional spatial coordinates and the probability of each vertebrae category; the distance feature matrix is ​​a two-dimensional square matrix, the number of rows and columns of which are equal to the total number of candidate nodes in the candidate node set; the distance feature matrix is ​​generated in the following manner: Based on the vector of the probability of each vertebra category for each candidate node in the set of candidate vertebrae center points, the preliminary hierarchical label of each candidate node is determined, and the hierarchical label is the vertebra category. For candidate node pairs whose hierarchical labels are anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with the three-dimensional Euclidean distance between the candidate nodes; for candidate node pairs whose hierarchical labels are not anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with a preset invalid label value.

[0008] Furthermore, the construction of the weighted directed graph based on the set of candidate vertebral center points and the distance feature matrix includes: Each candidate node is used as a vertex in a weighted directed graph. Each vertex includes three-dimensional spatial coordinates, a vector of probabilities for each vertebra category, and a preliminary hierarchical label. For any two candidate nodes in the candidate node set, one is taken as the source node and the other as the target node for judgment. If they are simultaneously satisfied with anatomical order adjacency, consistent upstream and downstream spatial position, and reasonable spatial distance, then a directed edge from the source node to the target node is constructed. The weight of each directed edge is determined based on the confidence of the initial hierarchical labels of the source node and the target node, as well as the reasonableness of the spatial distance between the two nodes; where the confidence of the initial hierarchical labels of the source node and the target node is the maximum probability value in the vector of probabilities of each vertebra category.

[0009] Furthermore, the weight of each directed edge is obtained by the following formula: ; in, ; In the formula, Indicates from candidate nodes Pointing to candidate nodes The weight of the directed edge. Indicates candidate nodes Initial hierarchical tags Confidence level, Indicates candidate nodes Initial hierarchical tags The probability, Indicates candidate nodes With candidate nodes The three-dimensional Euclidean distance between them , Representing candidate nodes respectively Initial hierarchical tags With candidate nodes Initial hierarchical tags The mean and standard deviation of spatial distance for the corresponding vertebral categories in the normal population. Represents the Euclidean distance in three-dimensional space In For the mean, The probability density value in the Gaussian probability density function is the standard deviation. , These represent the preset first and second weight adjustment coefficients, respectively.

[0010] Furthermore, the global optimal search strategy based on path relaxation iteration is the Bellman-Ford algorithm, in which, when it is determined that there is a negative weight cycle in the weighted directed graph, a hierarchical path completion strategy is triggered to obtain the available vertebral sequence path, and this path is taken as the optimal vertebral sequence path.

[0011] Furthermore, the hierarchical path completion strategy includes: Identify the vertebral-level regions involved in the negative weight cycle, mark all candidate nodes within these regions as anomalous nodes, temporarily remove the anomalous nodes and their associated edges from the weighted directed graph, and re-execute the Bellman-Ford algorithm in the remaining graph; if a complete path from the virtual source to the virtual endpoint can be obtained after re-execution, then this complete path is considered a usable vertebral sequence path; otherwise... Restore the removed abnormal nodes, remove the core node of the negative weight cycle from the restored weighted directed graph, and execute the Bellman-Ford algorithm again; where the core node of the negative weight cycle is the two endpoint nodes of the edge with the largest weight value in the negative weight cycle; if a complete path can be obtained after the second execution, then the complete path is used as the usable vertebral sequence path; otherwise... Abandon auto-completion and terminate the automatic positioning process at the T12 level.

[0012] Furthermore, the muscle segmentation model adopts the Inf-Net deep neural network architecture, including an input transformation and fusion module, an encoder, a parallel partial decoder, an implicit reverse attention module, and an explicit edge attention module; the input of the muscle segmentation model is a multi-channel input, and the output is a preliminary segmentation probability mask for muscle segmentation at the T12 level; wherein, the input transformation and fusion module is set before the encoder, and is used to convert the multi-channel input into a fused feature map before feeding it into the encoder; The construction of a multi-channel input based on the T12 thoracic vertebral level index and corresponding level images includes: Based on the T12 thoracic vertebral level index and corresponding level images, as well as standardized CT volume data, a multi-channel input tensor is constructed. The central channel of the multi-channel input tensor is the T12 level image, and the other channels are the images of the adjacent K layers above and below the T12 level, respectively. The channels are stacked in spatial order. The multi-channel input tensor is mapped to a muscle window and then scaled to obtain the multi-channel input.

[0013] Furthermore, the loss function used during the training of the muscle segmentation model is... Represented as: ; In the formula, , , , , These represent the category-focused loss, boundary loss, anatomical symmetry loss, auxiliary loss, and lateral output consistency loss, respectively. , , , , These represent the weight coefficients of category-focused loss, boundary loss, anatomical symmetry loss, auxiliary loss, and side-output consistency loss, respectively.

[0014] Furthermore, the loss of anatomical symmetry Represented as: ; In the formula, Indicates the total number of muscle categories. , These represent the image height and image width, respectively. Indicates the first Line 1 Column pixels belong to the The predicted probability of muscle-like structures This indicates a mirror transformation of the predicted probability map along the sagittal axis of the human body. Represents the L2 norm. Indicates the first Line 1 Adaptive confidence mask for column pixels.

[0015] On the other hand, embodiments of the present invention provide a skeletal muscle index fitting system for T12 thoracic vertebra localization and muscle segmentation, comprising: The CT image input module is used to acquire chest CT plain scan sequence images and perform preprocessing to generate standardized CT volume data; The spinal segmentation module is used to generate vertebral probabilistic volume data based on the standardized CT volume data, and then extract a set of candidate center points for the vertebrae; based on the spatial distance distribution between adjacent candidate center points in the set of candidate center points for the vertebrae, a distance feature matrix is ​​generated. The T12 localization module is used to construct a weighted directed graph based on the set of candidate vertebral center points and the distance feature matrix. It employs a global optimal search strategy based on path relaxation iteration to solve for the optimal vertebral sequence path in the weighted directed graph, and then obtains the T12 thoracic vertebral level index and corresponding level image based on the optimal vertebral sequence path. The muscle segmentation module is used to construct a multi-channel input based on the T12 thoracic vertebra level index and the corresponding level image, and input it into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. After post-processing optimization, the final muscle segmentation mask is obtained. The SMI fitting module is used to extract the muscle volume parameters at the T12 level from the final muscle segmentation mask, and, combined with patient information, output the predicted value of skeletal muscle index through a regression fitting model.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: This invention provides a method for T12 thoracic vertebra localization and muscle segmentation, along with a skeletal muscle index fitting system. By employing the Bellman-Ford algorithm based on path relaxation iteration as the global optimal search strategy, and leveraging its negative edge handling capability, it can automatically identify abnormal connections in cases of anatomical anomalies. Furthermore, a hierarchical path completion strategy is used to progressively degrade the path when anomalies occur, ensuring that the system can still obtain usable vertebral sequence paths even under abnormal anatomical structures, effectively improving the robustness and success rate of T12 thoracic vertebra localization. By constructing a multi-channel input centered on the T12 level, the network simultaneously perceives the texture features and inter-layer morphological continuity features of the current level and adjacent levels. Utilizing the Inf-Net architecture, combined with an input transformation and fusion module before the encoder, the anatomical continuity of the segmentation results is effectively improved. Simultaneously, an anatomical symmetry loss function is introduced into the loss function. This method utilizes the anatomical prior constraint of bilateral symmetry of the chest and abdominal wall muscles to constrain the segmentation results, leveraging the enhancement effect of the symmetry prior and avoiding the forced erasure of pathological features such as unilateral muscle atrophy, thus ensuring the clinical authenticity of the segmentation results. By extracting muscle volume from the segmentation mask and combining it with patient information to construct multidimensional input features, a regression fitting model is used to predict the skeletal muscle index, realizing a shift from the traditional geometric measurement paradigm to a data-driven prediction paradigm, improving the accuracy of SMI prediction. It achieves full automation from chest CT plain scan input to skeletal muscle index output, without manual intervention. It can directly use the existing chest CT data of COPD patients for opportunistic screening of sarcopenia without the need for additional abdominal CT scans or body composition analysis equipment, eliminating radiation exposure caused by additional scans, and providing clinicians with an efficient, objective, and accessible tool for quantifying imaging biomarkers.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0019] Figure 1 This is a flowchart illustrating the T12 thoracic vertebra localization and muscle segmentation method provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the skeletal muscle index fitting system provided in Embodiment 2 of the present invention. Detailed Implementation

[0020] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0021] Example 1 A specific embodiment of the present invention discloses a method for locating the T12 thoracic vertebra and segmenting muscles, such as... Figure 1 As shown, it includes the following steps: S1. Acquire chest CT plain scan sequence images and perform preprocessing to generate standardized CT volume data.

[0022] Specifically, the standardized CT volume data is an isotropic three-dimensional voxel grid, with voxel intensity values ​​being truncated and Z-score standardized normalized CT values. The isotropic three-dimensional voxel grid is obtained by trilinear interpolation resampling of the original chest CT plain scan sequence images, ensuring that the inter-slice resolution is consistent with the intra-slice planar resolution. The normalized CT values ​​are obtained by truncating the resampled original voxel CT values ​​within the range of [-1024, +1500] Hounsfield Units, and then using a global Z-score standardization method to convert the truncated values ​​into a distribution with a mean of 0 and a variance of 1. The upper and lower limits of the truncation process, as well as the global mean and global standard deviation used in Z-score standardization, are recorded to support the reversible mapping of normalized voxel intensity values ​​to the original Hounsfield Unit values.

[0023] Preferably, the target voxel side length for trilinear interpolation resampling is 1mm × 1mm × 1mm.

[0024] S2. Based on the standardized CT volume data, generate vertebral probability volume data, and then extract a set of candidate vertebral center points; based on the spatial distance distribution between adjacent candidate center points in the set of candidate vertebral center points, generate a distance feature matrix.

[0025] In practice, each voxel position in the vertebral probabilistic volume data has a vector representing the probability of it belonging to each vertebral category; the vertebral probabilistic volume data is a three-dimensional probability map, and its spatial dimension is consistent with that of the standardized CT volume data.

[0026] In practice, standardized CT volume data is input into a trained 3D convolutional neural network to generate 24-channel vertebral probabilistic volume data; the 24 channels correspond to 24 vertebral categories from C1 to L5, and the sum of the 24-dimensional probability vectors at each voxel position is 1.

[0027] Specifically, the 3D convolutional neural network adopts the 3D U-Net architecture, which uses an encoder-decoder structure to perform pixel-level semantic classification on each voxel.

[0028] Specifically, the 3D convolutional neural network training uses a training dataset constructed from spinal CT images with expert annotations. The annotations include vertebral category labels for each voxel, covering 24 vertebral categories from C1 to L5, and a background class. Before training, the samples in the training dataset are standardized to ensure consistency with the generation method of the standardized CT data in step S1. During training, a supervised learning approach is used, trained end-to-end. The loss value between the predicted result and the ground truth label is calculated through forward propagation, and the network parameters are iteratively updated through backpropagation until the loss function value gradually converges to a preset range. A mini-batch stochastic gradient descent optimization strategy is employed during training, and model performance is monitored on both the training and validation sets. The model parameters with the best performance on the validation set are selected as the final training result. The loss function for network training uses a combination of a region overlap-based loss function and a pixel classification-based loss function. Preferably, a weighted combination of the Dice loss function and a multi-class cross-entropy loss function is used, where the Dice loss optimizes the overlap between the predicted segmentation region and the ground truth region, and the cross-entropy loss optimizes the classification accuracy of each voxel. The total loss is a weighted sum of the two values, and the network achieves a balance between region integrity and pixel accuracy by adjusting the weight coefficients.

[0029] During implementation, the set of candidate vertebral center points is extracted using the following method: Based on the vertebral probabilistic data, probability maps corresponding to each vertebral category are obtained, and the following operations are performed on the probability maps corresponding to each vertebral category: Voxels with probability values ​​greater than or equal to a preset probability threshold in the probability graph of the current vertebral category are selected as candidate voxels for the current vertebral category. Three-dimensional connected component analysis is performed on each candidate voxel to merge spatially adjacent candidate voxels into the same connected region; and connected regions with a voxel count less than a preset volume threshold are removed. Within each remaining connected region, one or more local peak points for that vertebra category are determined by local peak detection and subjected to consistency verification. If the verification passes, the local peak point is added to the candidate center point set as a candidate center point.

[0030] Specifically, the consistency check is as follows: For each local peak point, obtain the 24-dimensional probability vector at the location of the local peak point, determine the vertebral category corresponding to the maximum probability value in the 24-dimensional probability vector, and compare it with the vertebral category used when extracting the local peak point. If the vertebral category corresponding to the maximum probability value at the location is inconsistent with the vertebral category used when extracting, the verification fails; if the vertebral category corresponding to the maximum probability value at the location is consistent with the vertebral category used when extracting, the verification passes.

[0031] Understandably, the above verification ensures that the hierarchical label of the candidate node (i.e., the vertebral category corresponding to its highest probability) is consistent with the vertebral category on which it was extracted. This avoids category conflicts caused by the same anatomical location generating local peaks on the multi-category probability map, thereby ensuring the uniqueness and reliability of the candidate node hierarchical label and providing clean input data for subsequent graph optimization.

[0032] Specifically, the probability threshold is set based on a balance between retaining as many real vertebral regions as possible and suppressing falsely identified non-vertebral regions. For example, a value of 0.5 is used to ensure that only voxels with a network prediction confidence of 50% or higher are retained as candidates, thereby effectively filtering out false positive noise with low confidence without excessively omitting real vertebrae.

[0033] Specifically, the volume threshold is set based on the minimum identifiable scale of the vertebral anatomical structure. For example, it is set to 50 voxels to eliminate isolated false positive connected regions caused by image noise or blurred segmentation boundaries, thereby reducing noise interference in the candidate center point set.

[0034] Specifically, candidate center points are determined through local peak detection, employing a mean-shift clustering algorithm, including: Using each candidate voxel in the connected region as an initial seed point, it iteratively moves along the probability gradient direction to a local probability maximum point; Voxels that converge to the same local probability maximum point are grouped into one class; The spatial coordinates of the local probability maximum point are taken as a local peak point.

[0035] Specifically, each candidate node in the set of candidate vertebrae center points includes three-dimensional spatial coordinates and a vector of probabilities for each vertebrae category; the vector of probabilities for each vertebrae category is a 24-dimensional probability vector, which is the probability value of the candidate center point belonging to each vertebrae category from C1 to L5.

[0036] Preferably, after a candidate center point is added to the candidate center point set, if there are multiple candidate nodes in the set whose spatial coordinates are less than a preset spatial threshold, only the candidate node with the highest probability value in the 24-dimensional probability vector is retained, and the remaining duplicate candidate nodes are removed, so as to avoid redundant nodes with overlapping spatial positions in the candidate node set.

[0037] Preferably, after extracting the set of candidate vertebral center points, a candidate node confidence screening step is also included: For each candidate node, check whether the maximum probability value in the vector of probabilities of each vertebra category is less than the preset confidence threshold. If so, remove the candidate node from the set.

[0038] Specifically, the confidence threshold is limited based on the balance between the graph optimization algorithm's capacity to handle the number of candidate nodes and the reliability of the nodes. For example, it is set to 0.1 to eliminate unreliable candidate nodes with low probability confidence, thereby reducing the node size and computational complexity of the weighted directed graph.

[0039] In practice, the distance feature matrix is ​​a two-dimensional square matrix, with the number of rows and columns equal to the total number of candidate nodes in the candidate node set. The row index and column index of the matrix correspond to different candidate nodes, and the element in the i-th row and j-th column of the matrix represents the spatial relationship between the i-th candidate node and the j-th candidate node.

[0040] In practice, the distance feature matrix is ​​generated in the following way: Based on the vector of the probability of each vertebra category for each candidate node in the set of candidate vertebrae center points, the preliminary hierarchical label of each candidate node is determined, and the hierarchical label is the vertebra category. For candidate node pairs whose hierarchical labels are anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with the three-dimensional Euclidean distance between the candidate nodes. For candidate node pairs whose hierarchical labels are not anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with a preset invalid flag value, which indicates that there is no valid connection between the candidate nodes.

[0041] Specifically, the initial hierarchical label for each candidate node is determined, including: The vertebral category corresponding to the maximum probability value in the vector of vertebral category probabilities of each candidate node is taken as the preliminary hierarchical label of the candidate node; at the same time, the maximum probability value is recorded as the confidence level of the hierarchical label.

[0042] Specifically, the default invalidation flag value is positive infinity, so that the invalid connection will not participate in the shortest path search in subsequent graph optimization.

[0043] It should be noted that, in this embodiment, the anatomical order of C1 to L5 refers to the sequence of vertebrae along the cephalopod direction (from cranial to caudal), namely: C1, C2, C3, C4, C5, C6, C7, T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, L1, L2, L3, L4, L5, totaling 24 vertebrae. "Adjacent" in this order refers to two vertebrae that are consecutively positioned in the above sequence, such as T11 and T12 being adjacent, or T12 and L1 being adjacent.

[0044] S3. Construct a weighted directed graph based on the set of candidate vertebral center points and the distance feature matrix. Use a global optimal search strategy based on path relaxation iteration to solve for the optimal vertebral sequence path in the weighted directed graph. Then, obtain the T12 thoracic vertebral level index and corresponding level image based on the optimal vertebral sequence path.

[0045] During implementation, a weighted directed graph is constructed based on the set of candidate vertebral centroids and the distance feature matrix, including: Each candidate node is used as a vertex in a weighted directed graph. Each vertex includes three-dimensional spatial coordinates, a vector of probabilities for each vertebra category, and a preliminary hierarchical label. For any two candidate nodes in the candidate node set, one is taken as the source node and the other as the target node for judgment. If they are simultaneously satisfied with anatomical order adjacency, consistent upstream and downstream spatial position, and reasonable spatial distance, then a directed edge from the source node to the target node is constructed. The weight of each directed edge is determined based on the confidence levels of the initial hierarchical labels of the source and target nodes and the reasonableness of the spatial distance between the two nodes. The confidence levels of the initial hierarchical labels of the source and target nodes are the maximum probability values ​​in the vector of probabilities for each vertebra category, and the reasonableness of the spatial distance between the two nodes is determined using a Gaussian distribution parameter table of the distance between adjacent vertebrae.

[0046] In practice, the anatomical order is as follows: following the anatomical order from C1 to L5, the preliminary hierarchical label of the target node is the next level below the preliminary hierarchical label of the source node; that is, the preliminary hierarchical label of the target node is the next vertebral category below the preliminary hierarchical label of the source node.

[0047] Consistent spatial location upstream and downstream means that the target node is located on the tail side of the source node along the anatomical cephalopod direction; that is, in the spatial coordinate system of standardized CT volume data, the Z-axis coordinate value of the target node is greater than the Z-axis coordinate value of the source node, where the positive direction of the Z-axis corresponds to the anatomical direction from the head end to the foot end of the patient; if the Z-axis direction of the current CT image is opposite to the anatomical cephalopod direction, the Z-axis value in this coordinate system is inverted before comparison to ensure the consistency of spatial order judgment. A reasonable spatial distance is defined as the spatial distance between the source node and the target node falling within the anatomically reasonable range.

[0048] Specifically, the spatial distance between candidate nodes is evaluated for anatomical rationality using a pre-statistically calculated Gaussian distribution parameter table of adjacent vertebral distances: when the actual distance between a candidate node pair is greater than the sum of the mean distance of the pair of adjacent vertebrae and three times the standard deviation, the candidate node pair is judged to be anatomically unreasonable; the Gaussian distribution parameter table of adjacent vertebral distances contains the mean and standard deviation of the spatial distances of 23 pairs of adjacent vertebrae in the normal population, which are composed of 24 vertebral categories from C1 to L5.

[0049] More specifically, the Gaussian distribution parameter table of the distance between adjacent vertebrae was obtained by statistically analyzing the vertebral center points annotated by experts in a large number of normal spinal CT images.

[0050] In practice, the weight of each directed edge can be obtained using the following formula: ; In the formula, Indicates from candidate nodes Pointing to candidate nodes The weight of the directed edge. Indicates candidate nodes Initial hierarchical tags confidence level Indicates candidate nodes Initial hierarchical tags The probability, Indicates candidate nodes With candidate nodes The three-dimensional Euclidean distance between them , Representing candidate nodes respectively Initial hierarchical tags With candidate nodes Initial hierarchical tags The mean and standard deviation of spatial distance for the corresponding vertebral categories in the normal population. Represents the Euclidean distance in three-dimensional space In For the mean, The probability density value in the Gaussian probability density function is the standard deviation. This represents the Gaussian probability density function.

[0051] In practice, the weight of each directed edge can also be obtained using the following formula: ; in, ; In the formula, , These represent the preset first and second weight adjustment coefficients, respectively.

[0052] Specifically, by introducing first and second adjustable weight coefficients, the contributions of node discrimination information and spatial distance information to path selection are balanced, preventing any one factor from dominating the edge weights due to an excessively large numerical range. This improves the robustness and localization accuracy of graph search under different anatomical variations. Simultaneously, grid search on the validation set ensures the generalization ability of the coefficients. For example, , Set them to 1 and 0.5 respectively.

[0053] It is understandable that when the actual distance Exceeding When the distance falls within a certain range, the standard Gaussian probability density value is used. When the actual distance exceeds this range, the probability density value at the boundary is used as a constant extrapolation to ensure the continuity of the weight function at the boundary. By performing a soft truncation on the distance probability density through constant extrapolation, the discontinuity of the weight function at the boundary caused by hard truncation is eliminated, allowing the distance rationality evaluation to smoothly transition at the boundary of the normal physiological range, thus enhancing the numerical stability of path search.

[0054] In practice, the global optimal search strategy based on path relaxation iteration is the Bellman-Ford algorithm. When it is determined that there is a negative weight cycle in the weighted directed graph, a hierarchical path completion strategy is triggered to obtain the available vertebral sequence path, which is then used as the optimal vertebral sequence path.

[0055] In specific implementation, the global optimal search strategy based on path relaxation iteration includes: solving the shortest path distance from the virtual source point to each vertex and the predecessor vertex of each vertex through at most V-1 rounds of full edge relaxation iteration, where V represents the total number of vertices in the weighted directed graph including the virtual source point and the virtual endpoint; after completing V-1 rounds of relaxation iteration, the Vth round of full edge traversal is performed. If there is a relaxable edge in this round, it is determined that there is a negative weight cycle in the weighted directed graph, triggering the hierarchical path completion strategy to obtain a usable vertebral sequence path, which is then used as the optimal vertebral sequence path; otherwise, backtracking from the virtual endpoint along the predecessor vertices of each vertex after V-1 rounds of relaxation iteration back to the virtual source point, and using the backtracked path as the optimal vertebral sequence path.

[0056] In practice, a virtual source and a virtual endpoint are set in the weighted directed graph. The virtual source is an additional vertex in the weighted directed graph, which points to all candidate nodes with an initial level label of C1, or the first candidate node with an identifiable vertebral category at the head end within the CT scan range, with zero weight. The virtual endpoint is an additional vertex in the weighted directed graph, which points to all candidate nodes with an initial level label of L5, or the last candidate node with an identifiable vertebral category at the tail end within the CT scan range, with zero weight.

[0057] Specifically, the first identifiable vertebral category at the head end within the CT scan range is determined as follows: starting from C1, traversing downwards to find the highest vertebral category whose maximum probability value in the probability vector exceeds a preset confidence threshold; the last identifiable vertebral category at the tail end within the CT scan range is determined as follows: starting from L5, traversing upwards to find the lowest vertebral category whose maximum probability value in the probability vector exceeds a preset confidence threshold; wherein, when the CT scan range does not completely cover the entire sequence from C1 to L5, the virtual source point and virtual endpoint are automatically adjusted according to the actual scan range, and the starting and ending vertebral categories of the scan range are recorded to assist in verifying the anatomical rationality of the T12 localization results.

[0058] In practice, the distance from the virtual source point to itself is set to 0, and the distance to all other vertices is set to positive infinity. Each relaxation iteration involves performing a relaxation operation on each directed edge in the weighted directed graph: if the sum of the currently recorded distance from the virtual source point to the edge's starting point and the edge weight is less than the currently recorded distance to the edge's ending point, then the distance to the edge's ending point is updated to the shortest path distance, and its predecessor vertex is recorded. One relaxation iteration allows the shortest path information to be propagated one layer forward along one edge. Since the shortest path from the virtual source point to any vertex passes through at most V-1 edges, after V-1 relaxation iterations, the shortest path information from the virtual source point to all vertices has been propagated, and the shortest path distances to each vertex have converged. If a path repeatedly passes through a vertex, a cycle exists; removing the cycle yields a shorter path. Therefore, the shortest path must be an acyclic path, with no more than V-1 edges.

[0059] Preferably, after determining that a negative weight cycle exists in the weighted directed graph and before executing the hierarchical path completion strategy, the method further includes: determining whether the negative weight cycle is located on a reachable path from the virtual source point to the virtual endpoint; if the negative weight cycle is not on a reachable path, it does not affect the solution of the optimal vertebral sequence path, and the predecessor vertices of each vertex after V-1 rounds of relaxation iteration from the virtual endpoint are directly backtracked to the virtual source point, and the backtracked path is used as the optimal vertebral sequence path; if the negative weight cycle is on a reachable path, the hierarchical path completion strategy is triggered.

[0060] In practice, the hierarchical path completion strategy includes: Identify the vertebral-level regions involved in the negative weight cycle, mark all candidate nodes within these regions as anomalous nodes, temporarily remove the anomalous nodes and their associated edges from the weighted directed graph, and re-execute the Bellman-Ford algorithm in the remaining graph; if a complete path from the virtual source to the virtual endpoint can be obtained after re-execution, then this complete path is considered a usable vertebral sequence path; otherwise... Restore the removed abnormal nodes, remove the core node of the negative weight cycle from the restored weighted directed graph, and execute the Bellman-Ford algorithm again; where the core node of the negative weight cycle is the two endpoint nodes of the edge with the largest weight value in the negative weight cycle; if a complete path can be obtained after the second execution, then the complete path is used as the usable vertebral sequence path; otherwise... Abandon auto-completion, terminate the automatic localization process at the T12 level, mark the anomalies in the output, and add the message "T12 level localization requires manual review" to the results.

[0061] Specifically, after marking the abnormal nodes and temporarily removing them from the weighted directed graph, the process also includes: If a certain vertebral category has no alternative node in the graph, meaning all candidate nodes for that vertebral category are marked as abnormal, then during path backtracking, the missing vertebral category will be recorded, and the final report will include a message stating, "The vertebral body corresponding to this vertebral category is missing during automatic localization, possibly due to anatomical abnormalities or image quality."

[0062] Specifically, the vertebral hierarchical regions involved in the negative weight ring are identified, including: Extract the preliminary level labels corresponding to all vertices from the negative weighted ring to form a level label set; take the closed interval between the minimum level label and the maximum level label in this level label set as the vertebral level region involved in the negative weighted ring; mark all candidate nodes whose preliminary level labels fall into this closed interval as abnormal nodes.

[0063] Specifically, the initial hierarchical labels are mapped to incrementally increasing numerical labels according to the anatomical order from C1 to L5, with smaller numbers indicating closer to the head and larger numbers indicating closer to the foot; the minimum hierarchical label is the hierarchical label with the smallest number in the hierarchical label set, and the maximum hierarchical label is the hierarchical label with the largest number in the hierarchical label set.

[0064] During implementation, the node with the layer label T12 is located in the optimal vertebral sequence path, the three-dimensional spatial coordinates corresponding to the node are read, its Z-axis coordinate is mapped to the layer index of the original CT sequence, and the corresponding two-dimensional layer image is extracted from the standardized CT volume data according to the layer index, which is the layer image corresponding to the T12 thoracic vertebra.

[0065] Specifically, when mapping the Z-axis coordinates to the layer index, a linear mapping is performed based on the spatial position coordinates of the first layer in the standardized CT volume data and the uniform interlayer spacing, and the calculation result is rounded to the nearest integer.

[0066] Preferably, if no node with the T12 level label is located in the optimal vertebral sequence path, an abnormal path analysis is performed in combination with the negative weighted loop detection results, and the maximum likelihood sub-path containing the T12 label is searched in the candidate path; if it still cannot be located, a location failure mark is triggered, prompting manual review.

[0067] Understandably, the appearance of negative weight cycles in vertebral localization indicates a connection between candidate nodes that severely violates normal anatomical structures. For example, the distance between candidate points of two adjacent vertebrae may be abnormally short (e.g., compression fractures causing a significant reduction in vertebral height), or multiple candidate nodes may form a cyclic connection pattern that contradicts the normal sequence. Upon detecting a negative weight cycle, the algorithm should not be abandoned immediately, but rather a hierarchical completion strategy should be initiated to ensure that the algorithm can still provide usable outputs under abnormal anatomical structures.

[0068] S4. Construct a multi-channel input based on the T12 thoracic vertebral level index and corresponding level images, and input it into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. Then, after post-processing optimization, obtain the final muscle segmentation mask.

[0069] During implementation, a multi-channel input is constructed based on the T12 thoracic vertebral level index and corresponding level images, including: Based on the T12 thoracic vertebral level index and corresponding level images, as well as standardized CT volume data, a multi-channel input tensor is constructed. The central channel of the multi-channel input tensor is the T12 level image, and the other channels are the images of the adjacent K layers above and below the T12 level, respectively. The channels are stacked in spatial order. The multi-channel input tensor is mapped to a muscle window and then scaled to obtain the multi-channel input.

[0070] Specifically, taking the layer image corresponding to the T12 thoracic vertebra as the center, K adjacent layers are selected upwards and downwards along the Z-axis direction of the standardized CT volume data, for a total of 2K+1 layers. The selected 2K+1 layers are then stacked in spatial order to construct a multi-channel input tensor.

[0071] Preferably, K=3, that is, the input tensor contains 7 layers of image. Under this setting, the segmentation accuracy and computational efficiency are optimally balanced. If the value of K is too small, the inter-layer context information is insufficient, and if the value of K is too large, too much irrelevant information is introduced and the computational burden is increased.

[0072] Understandably, by constructing multi-channel inputs from adjacent layers with the same number of layers above and below the T12 layer, the network can simultaneously perceive the texture features of the current layer and the morphological continuity features of adjacent layers during inference, effectively reducing inter-layer breaks and local omissions common in single-layer segmentation, and improving the anatomical continuity of the segmentation results.

[0073] Specifically, muscle window mapping of multi-channel input tensors refers to adjusting the window width and window level of the grayscale display range of each channel of the multi-channel input tensor to optimize the contrast of muscle tissue in the image, making it easier for the network to extract muscle features.

[0074] More specifically, the preset muscle window has a width of 350 Hounsfield Units and a window level of 40 Hounsfield Units to maximize the grayscale contrast between muscle tissue and surrounding fat and internal organs.

[0075] Specifically, the CT images of the T12 level and adjacent levels are subjected to muscle window adjustment. Pixel values ​​within the CT range of [-135, +215] Hounsfield Units are linearly mapped to the display grayscale range, and values ​​exceeding this range are truncated to the lower or upper grayscale limit, respectively. The muscle window has a width of 350 Hounsfield Units and a window level of 40 Hounsfield Units. After muscle window adjustment, the grayscale contrast between muscle tissue and surrounding adipose tissue and bone is maximized, and display interference from adipose tissue and bone is suppressed.

[0076] Specifically, the multi-channel input tensor mapped from the muscle window is scaled to normalize the pixel values ​​of each channel to the [0, 1] interval, ensuring that the pixel value distribution is consistent with the input data during model training. The scaling is linear, resulting in similar grayscale ranges for the muscle tissue regions in each channel image. This scaled multi-channel input tensor is then used as the input to the trained muscle segmentation model.

[0077] Specifically, the segmentation target areas of the muscle segmentation model include: erector spinae, latissimus dorsi, external oblique, rectus abdominis, and intercostal muscles. The segmentation target area is all skeletal muscle tissue on the horizontal section of the T12 vertebral body, from the subcutaneous fat layer to the posterior edge of the vertebral body.

[0078] In implementation, the muscle segmentation model adopts the Inf-Net deep neural network architecture, including an input transformation and fusion module, an encoder, a parallel partial decoder, an implicit reverse attention module, and an explicit edge attention module. The input of the muscle segmentation model is a multi-channel input, and the output is a preliminary segmentation probability mask for muscle segmentation at the T12 level. The input transformation and fusion module is set before the encoder to convert the multi-channel input into a fused feature map before feeding it into the encoder.

[0079] In practice, the input transformation and fusion module extracts the planar texture features of each layer of the multi-channel input tensor through 2D convolution channel by channel, and then uses 3×1×1 3D convolution along the channel dimension to capture the continuous morphological features between layers. The weights of each channel are automatically adjusted through the channel attention mechanism to obtain the fused features that fuse 2D spatial features and 3D inter-layer features.

[0080] The encoder performs multi-layer convolutional encoding on the fused features to extract feature maps at different scales, including low-level and high-level features; The parallel partial decoder is used to aggregate high-level features from the encoder to generate a global feature map; The implicit inverse attention module is used to perform inverse attention operations step by step from deep to shallow based on high-level features of the encoder to enhance the discrimination ability of boundary regions; The explicit edge attention module is used to extract muscle boundary information from the encoder's low-level features for explicit edge supervision.

[0081] During implementation, the loss function used when training the muscle segmentation model is... Represented as: ; In the formula, , , , , These represent the category-focused loss, boundary loss, anatomical symmetry loss, auxiliary loss, and lateral output consistency loss, respectively. , , , , represents the weight coefficients of the category focus loss, boundary loss, anatomical symmetry loss, auxiliary loss, and lateral output consistency loss, respectively. Specifically, the category focus loss handles class imbalance and controls the penalty bias for false positives and false negatives; the boundary loss enhances the supervision weights of boundary pixels; the anatomical symmetry loss constrains the left-right symmetry of the segmentation results; the auxiliary loss supervises multi-scale lateral outputs; and the lateral output consistency loss constrains the consistency of predictions across different scales of lateral outputs.

[0082] In practice, the category focuses on losses. Represented as: ; In the formula, Indicates the total number of muscle categories. This represents the total number of pixels in a single image. Indicates the first The pixel belongs to the first The predicted probability of muscle-like structures Indicates the first The pixel belongs to the first The true probability of muscle-like structures , These represent the first and second balancing factors of the category-focused loss, respectively. Indicates the focus parameter, Represents the numerical smoothing term. Indicates the first Class weights for muscle-like structures.

[0083] Specifically, These correspond to the erector spinae, latissimus dorsi, external oblique, rectus abdominis, and intercostal muscles, respectively.

[0084] Specifically, the first and second balancing factors of the category-focused loss are used to control over-segmentation and under-segmentation penalties. Their specific values ​​are determined by performing a grid search on the validation set with a step size of 0.1, selecting the combination that maximizes the F1 score on the validation set. The search range is [0.1, 0.5]. The search range is [0.5, 0.9], and satisfies... + =1. This indicates that the penalty for missed partitions is higher than that for over-partitions. For example, , Set them to 0.3 and 0.7 respectively.

[0085] Specifically, focusing on parameters To reduce the loss contribution of easily classifiable pixels, a search is performed on the validation set in the range [1,3] with a step size of 0.5, selecting the highest value of Dice on the validation set. For example, Set to 2.

[0086] Specifically, numerical smoothing term To prevent the denominator from being zero, an example setting is 10. -5 .

[0087] Specifically, category weights This is used to assign higher weights to categories with low voxel percentages, and is represented as: ; In the formula, Indicates the first The percentage of muscle-like structures in the training set. Indicates the first The percentage of muscle-like structures in the training set. Take the maximum value for all muscle categories.

[0088] In practical implementation, boundary loss Represented as: ; In the formula, Indicates the first Boundary weights for each pixel.

[0089] Specifically, no. Boundary weights of each pixel Represented as: ; In the formula, Indicates the first The Euclidean distance from each pixel to the true segmentation boundary. This represents the boundary weight magnitude control coefficient. This represents the boundary influence range control coefficient.

[0090] Specifically, the boundary weight magnitude control coefficient The value range is [2, 10]. Within this range, a search is performed with a step size of 1 to select the value that has the lowest Hausdorff distance to the 95th percentile of the validation set. For example, Set it to 5.

[0091] Specifically, the boundary influence range control coefficient is set based on anatomical priors, for example, Set it to 5.

[0092] Specifically, no. Boundary weights of each pixel The range is [1, 1+ The weight of the background region (far from the boundary) approaches 1; the weight of the boundary reaches 1+. The gradient contribution is approximately six times that of the background pixels. This means that the loss weights for boundary pixels are increased to six times that of the background pixels, allowing the network to receive stronger boundary localization gradient signals during backpropagation, effectively improving the segmentation accuracy of low-contrast boundaries.

[0093] In practice, anatomical symmetry loss Represented as: ; In the formula, , These represent the image height and image width, respectively. Indicates the first Line 1 Column pixels belong to the The predicted probability of muscle-like structures This indicates a mirror transformation of the predicted probability map along the sagittal axis of the human body. Represents the L2 norm. Indicates the first Line 1 Adaptive confidence mask for column pixels.

[0094] Specifically, no. Line 1 Adaptive confidence mask for column pixels Represented as: ; In the formula, This represents the threshold for the difference between the left and right predictions. An adaptive confidence mask controls the strength of the symmetry constraint, automatically reducing the constraint weight in regions with large prediction differences to prevent unilateral muscle atrophy in pathological states from being forcibly constrained to symmetry, leading to feature loss. When there is a significant difference between the left and right predictions, the symmetry constraint weight in that region is automatically reduced to preserve individualized pathological features. Specifically, the left and right prediction difference threshold The highest value of Dice in the validation set is selected by searching the range [0.1, 0.5] with a step size of 0.05. For example, Set it to 0.3.

[0095] Specifically, the human sagittal axis is set as follows: Or it can be dynamically determined based on the results of spinal region detection. Specifically, if the human sagittal axis is set to... ,but The mirror position is .

[0096] In practice, auxiliary losses Represented as: ; In the formula, Indicates the first The prediction results of each side output branch after upsampling to the original image size Indicates the true label, Indicates standard Dice loss. This indicates the output weights for each side.

[0097] Preferably, , , Set them to 0.2, 0.3, and 0.5 respectively.

[0098] In practice, this results in side output consistency loss. Represented as: .

[0099] In implementation, a sample set is constructed for training the muscle segmentation model. Each training sample contains an input image and corresponding labels. The input images are derived from chest or abdominal CT scans. After standardization preprocessing in step S1 and muscle window adjustment in step S4, images of K adjacent layers are taken upwards and downwards along the Z-axis, centered at the T12 vertebral level, and stacked in spatial order to form a multi-channel input. The labels are pixel-level binary masks corresponding to the T12 layer images in the input images. All labeled samples are proportionally divided into training, validation, and test sets for training and validation.

[0100] Specifically, the image size is 512×512 pixels. Pixels belonging to muscle tissue in the mask are marked as 1, and pixels not belonging to muscle tissue are marked as 0. The labels were generated by radiologists who delineated the boundaries of the muscle regions pixel by pixel on the original T12-level CT image.

[0101] Specifically, the training data covers images from CT scanners of different brands and models, under different scanning conditions, and patients with different disease states, including images from conventional-dose CT and low-dose CT scans.

[0102] Preferably, the total number of training samples is no less than 2,000 T12 layer images, of which COPD patient samples account for no less than 50%.

[0103] In practice, the training samples are preprocessed, including: The pixel spacing of each CT image layer is resampled to a uniform size to ensure that all samples have consistent spatial resolution; the CT values ​​are truncated to the range of [-135, +215] HU (corresponding to the muscle window display range with a window width of 350 HU and a window level of 40 HU) and linearly normalized to the interval of [0,1]; the labels are interpolated by nearest neighbor during resampling to maintain the integrity of the binary mask.

[0104] Specifically, during the training process, the same random transformation is applied to the input image and its corresponding label to generate diverse training samples.

[0105] Preferably, the training samples are further augmented by geometric transformation, intensity transformation, and elastic deformation. The data augmentation is performed randomly before each round of training, and different random seeds are used in different rounds. The augmented samples are not stored persistently and are generated online during the training process.

[0106] During implementation, the muscle segmentation model was trained using a four-round iterative training strategy: In the first round, the network parameters are randomly initialized, and training is performed using a fully labeled training dataset. The loss weights in the total loss function are configured with the first preset value. An early stopping strategy is adopted during training: training is terminated when the Dice coefficient on the validation set no longer increases for 20 consecutive rounds, and the network parameters W1 that achieve the best performance on the validation set are saved.

[0107] In the second round, the network parameters W1 from the first round are used as initial values, and fine-tuning is performed using the same training dataset as the first round. Through backpropagation, while maintaining region segmentation capabilities, the network parameters are pushed towards the backward gradient of the optimization boundary, making the model more focused on boundary localization accuracy. Training terminates when the Hausdorff distance to the 95th quantile on the validation set no longer decreases for 10 consecutive rounds, and the optimized network parameters W2 are saved.

[0108] In the third round, the network parameters W1 from the first round and W2 from the second round are used to perform inference predictions on the entire training set. The Dice coefficient difference between the prediction results of the two rounds is calculated for each sample. The samples are sorted in descending order of Dice difference, and the top 30% are selected as the hard sample set. For the hard sample set, data augmentation is applied, and the loss function weights are restored to the configuration of the first round for fine-tuning training. The optimized network parameters W3 are then saved.

[0109] In the fourth round, for each training sample, predictions are made using the network parameters W1 from the first round, W2 from the second round, and W from the third round, resulting in three probability maps. The probability entropy of the three predictions is calculated pixel-by-pixel. For pixels with an entropy value exceeding 0.3, a majority voting strategy is used to generate refined pseudo-labels. For pixels with an entropy value not exceeding 0.3, the prediction result of model W1 from the first round is directly used as the supervision signal. Under the supervision of the refined pseudo-labels, the network parameters W3 from the third round are loaded as initial values, and final fine-tuning is performed on the full training dataset. The weights of the loss function hyperparameters are adjusted to the third preset value. Training terminates when the segmentation metric on the full validation set saturates, and the final network parameters W are saved. final .

[0110] Specifically, , , , , The first preset values ​​are 1.0, 0.5, 0.4, 0.5, and 0.2. , , , , The second preset values ​​are 1.0, 0.8, 0.4, 0.5, and 0.2. , , , , The third preset value is 1.0, 0.4, 0.3, 0.3, 0.1.

[0111] During implementation, post-processing of the initial segmentation probability mask includes: The segmentation probability mask is binarized using a preset segmentation threshold to obtain an initial binary mask; Perform connected component analysis on the binary mask, remove isolated connected regions with an area smaller than a preset area threshold, and fill in the internal holes; Perform morphological closing operations on the binary mask after connected component filtering to smooth the segmentation boundaries; The morphologically processed segmentation mask is mapped back to the original HU space. Pixels in the segmentation mask coverage area whose HU values ​​exceed the standard CT value range of skeletal muscle tissue are corrected to background pixels to obtain the final muscle segmentation mask.

[0112] Specifically, the standard CT value range for skeletal muscle tissue is the overflow pixels in the range of [-29, +150].

[0113] Preferably, the segmentation threshold is set to 0.5, the area threshold is set to 50 pixels, and the kernel size of the morphological closing operation is 5×5 pixels.

[0114] Compared with existing technologies, this embodiment provides a T12 thoracic vertebra localization and muscle segmentation method. By employing the Bellman-Ford algorithm based on path relaxation iteration as the global optimal search strategy, and utilizing its negative edge processing capability, it can automatically identify abnormal connections in cases of anatomical anomalies. Furthermore, a hierarchical path completion strategy is used to progressively degrade the processing when anomalies occur, ensuring that the system can still obtain usable vertebral sequence paths even under abnormal anatomical structures, effectively improving the robustness and success rate of T12 thoracic vertebra localization. By constructing a multi-channel input centered on the T12 level, the network simultaneously perceives the texture features and inter-layer morphological continuity features of the current level and adjacent levels. Utilizing the Inf-Net architecture, combined with the input transformation and fusion module before the encoder, the anatomical continuity of the segmentation results is effectively improved. Simultaneously, an anatomical symmetry loss is introduced into the loss function, using the anatomical prior of left-right symmetry of the thoracic and abdominal wall muscles to constrain the segmentation results. This leverages the enhancement effect of the symmetry prior, avoiding the forced removal of pathological features such as unilateral muscle atrophy, and ensuring the clinical authenticity of the segmentation results.

[0115] Example 2 Based on the T12 thoracic vertebra localization and muscle segmentation method of Embodiment 1, a skeletal muscle index fitting system for T12 thoracic vertebra localization and muscle segmentation is also provided, such as... Figure 2 As shown, it includes: The CT image input module is used to acquire chest CT plain scan sequence images and perform preprocessing to generate standardized CT volume data; The spinal segmentation module is used to generate vertebral probabilistic volume data based on the standardized CT volume data, and then extract a set of candidate center points for the vertebrae; based on the spatial distance distribution between adjacent candidate center points in the set of candidate center points for the vertebrae, a distance feature matrix is ​​generated. The T12 localization module is used to construct a weighted directed graph based on the set of candidate vertebral center points and the distance feature matrix. It employs a global optimal search strategy based on path relaxation iteration to solve for the optimal vertebral sequence path in the weighted directed graph, and then obtains the T12 thoracic vertebral level index and corresponding level image based on the optimal vertebral sequence path. The muscle segmentation module is used to construct a multi-channel input based on the T12 thoracic vertebra level index and the corresponding level image, and input it into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. After post-processing optimization, the final muscle segmentation mask is obtained. The SMI fitting module is used to extract the muscle volume parameters at the T12 level from the final muscle segmentation mask, and, combined with patient information, output the predicted value of skeletal muscle index through a regression fitting model.

[0116] Specifically, the total number of foreground pixels in the final muscle segmentation mask is counted, and the muscle volume at the T12 level is obtained by combining this with the physical spatial resolution of the CT image. The physical spatial resolution of the CT image includes pixel pitch and slice thickness.

[0117] Specifically, patient information includes patient height, weight, age, and gender. This patient information is obtained from DICOM metadata or electronic medical record systems.

[0118] During implementation, the input features of the regression fitting model include: equivalent volume of muscle at the T12 level, patient height, patient weight, patient age, and patient gender; among which gender is encoded as a binary variable.

[0119] In practice, the regression fitting model is a multiple linear regression model, with input features as independent variables and skeletal muscle index as dependent variable, and the regression coefficients are fitted using the least squares method.

[0120] In practice, the regression fitting model adopts a stacking ensemble regression strategy. The first layer of base learners includes multiple linear regression, random forest regression, and extreme gradient boosting regression, while the second layer of meta-learners uses ridge regression. During training, five-fold cross-validation is used to generate the training set predictions of the base learners to prevent overfitting.

[0121] Understandably, regression fitting models can implicitly learn the complex nonlinear mapping relationship between muscle mass distribution, morphological features and skeletal muscle index, overcoming the anatomical limitation of the direct ratio method which simplifies the muscle-height relationship to a fixed power. Using muscle volume and height, weight, age and gender as inputs, it implicitly incorporates feature dimensions such as muscle mass distribution that cannot be captured by direct calculation methods.

[0122] Specifically, while the regression fitting model outputs the predicted value of the skeletal muscle index, it also outputs the prediction confidence label: calculate the standard deviation of the predicted values ​​of the three base learners, and if the standard deviation is less than the preset threshold, output a high confidence label; otherwise, output a low confidence label and prompt manual review.

[0123] In practice, the regression fitting model is trained and obtained in the following way: The training set is composed of samples labeled with real skeletal muscle index values. The muscle volume, height, weight, age, and gender of the training set samples are used as input features, and the real skeletal muscle index values ​​are used as supervision signals. The regression model is trained using supervised learning. During the training process, cross-validation is used to fine-tune hyperparameters, and the model performance is evaluated on an independent test set.

[0124] Preferably, before inputting the features into the regression model, the method further includes: RobustScaler is used to standardize the input features, which involves subtracting the feature median and dividing by the interquartile range to eliminate the influence of dimensions and improve robustness to outliers. The standardization parameters of RobustScaler are obtained by prefitting the training set.

[0125] Furthermore, the skeletal muscle index fitting system also includes clinical decision support for the prediction results: The predicted skeletal muscle index (SMI) value output by the regression model is compared with the preset diagnostic threshold. When the predicted SMI value for men is less than 52.4 cm² / m² or the predicted SMI value for women is less than 38.5 cm² / m², a screening prompt of "high risk of sarcopenia" is automatically output; otherwise, a screening prompt of "low risk of sarcopenia" is output.

[0126] It should be noted that the performance metrics of the regression model on the test set are: R² = 0.856, MAE = 0.249, and RMSE = 0.315. The R² value indicates that the model can explain approximately 85.6% of the SMI variance, demonstrating high predictive accuracy. The training set R² = 0.756, and the difference between the training set R² and the test set R² indicates that the model did not exhibit significant overfitting. Bland-Altman consistency analysis results show that the mean deviation between the regression model's predicted values ​​and the reference standard is approximately +0.13, with all data points falling within the 95% consistency bound, no extreme outliers, and overall consistency being acceptable.

[0127] Compared with existing technologies, this embodiment provides a skeletal muscle index fitting system for T12 thoracic vertebra localization and muscle segmentation. By extracting muscle volume from the segmentation mask and combining it with patient information to construct multidimensional input features, a regression fitting model is used to predict the skeletal muscle index. This realizes a shift from the traditional geometric measurement paradigm to a data-driven prediction paradigm, improving the accuracy of SMI prediction. It achieves full automation from chest CT plain scan input to skeletal muscle index output, requiring no manual intervention. It can directly utilize existing chest CT data of COPD patients for opportunistic screening of sarcopenia without the need for additional abdominal CT scans or body composition analysis equipment. It eliminates radiation exposure caused by additional scans and provides clinicians with an efficient, objective, and accessible tool for quantifying biomarkers in imaging. The overall processing time from input to SMI output for a single CT image is reduced to the second level, T12 localization of the Bellman-Ford algorithm can be completed in sub-seconds, the inference time of Inf-Net muscle segmentation meets the requirements of real-time processing, and the SMI prediction of the regression model is calculated in milliseconds. The overall processing speed has been improved by two orders of magnitude compared with the traditional manual method, making it technically possible to conduct large-scale sarcopenia screening for COPD patients and significantly improving the efficiency of respiratory clinical work.

[0128] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for locating the T12 thoracic vertebra and segmenting muscles, characterized in that, Includes the following steps: Acquire chest CT plain scan images and preprocess them to generate standardized CT volume data; Based on the standardized CT volume data, vertebral probabilistic volume data is generated, and then a set of candidate vertebral center points is extracted; based on the spatial distance distribution between adjacent candidate center points in the set of candidate vertebral center points, a distance feature matrix is ​​generated. A weighted directed graph is constructed based on the set of candidate vertebral center points and the distance feature matrix. The optimal vertebral sequence path is solved in the weighted directed graph using a global optimal search strategy based on path relaxation iteration. Then, the T12 thoracic vertebral level index and corresponding level image are obtained based on the optimal vertebral sequence path. A multi-channel input is constructed based on the T12 thoracic vertebral level index and corresponding level images, and then input into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. After post-processing optimization, the final muscle segmentation mask is obtained.

2. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 1, characterized in that, Each candidate node in the set of candidate vertebrae center points includes a vector of three-dimensional spatial coordinates and the probability of each vertebrae category; the distance feature matrix is ​​a two-dimensional square matrix, the number of rows and columns of which are equal to the total number of candidate nodes in the candidate node set; The distance feature matrix is ​​generated in the following manner: Based on the vector of the probability of each vertebra category for each candidate node in the set of candidate vertebrae center points, the preliminary hierarchical label of each candidate node is determined, and the hierarchical label is the vertebra category. For candidate node pairs whose hierarchical labels are anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with the three-dimensional Euclidean distance between the candidate nodes; for candidate node pairs whose hierarchical labels are not anatomically adjacent, the corresponding matrix elements in the distance feature matrix are filled with a preset invalid label value.

3. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 2, characterized in that, The construction of a weighted directed graph based on the set of candidate vertebral centroids and the distance feature matrix includes: Each candidate node is used as a vertex in a weighted directed graph. Each vertex includes three-dimensional spatial coordinates, a vector of probabilities for each vertebra category, and a preliminary hierarchical label. For any two candidate nodes in the candidate node set, one is taken as the source node and the other as the target node for judgment. If they are simultaneously satisfied with anatomical order adjacency, consistent upstream and downstream spatial position, and reasonable spatial distance, then a directed edge from the source node to the target node is constructed. The weight of each directed edge is determined based on the confidence of the initial hierarchical labels of the source node and the target node, as well as the reasonableness of the spatial distance between the two nodes; where the confidence of the initial hierarchical labels of the source node and the target node is the maximum probability value in the vector of probabilities of each vertebra category.

4. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 3, characterized in that, The weight of each directed edge is obtained by the following formula: ； in, ； In the formula, Indicates from candidate nodes Pointing to candidate nodes The weight of the directed edge. Indicates candidate nodes Initial hierarchical tags confidence level Indicates candidate nodes Initial hierarchical tags The probability, Indicates candidate nodes With candidate nodes The three-dimensional Euclidean distance between them , Representing candidate nodes respectively Initial hierarchical tags With candidate nodes Initial hierarchical tags The mean and standard deviation of spatial distance for the corresponding vertebral categories in the normal population. Represents the Euclidean distance in three-dimensional space In For the mean, The probability density value in the Gaussian probability density function is the standard deviation. , These represent the preset first and second weight adjustment coefficients, respectively.

5. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 3, characterized in that, The global optimal search strategy based on path relaxation iteration is the Bellman-Ford algorithm. When it is determined that there is a negative weight cycle in the weighted directed graph, a hierarchical path completion strategy is triggered to obtain the available vertebral sequence path, and this path is taken as the optimal vertebral sequence path.

6. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 5, characterized in that, The hierarchical path completion strategy includes: Identify the vertebral-level regions involved in the negative weight cycle, mark all candidate nodes within these regions as anomalous nodes, temporarily remove the anomalous nodes and their associated edges from the weighted directed graph, and re-execute the Bellman-Ford algorithm in the remaining graph; if a complete path from the virtual source to the virtual endpoint can be obtained after re-execution, then this complete path is considered a usable vertebral sequence path; otherwise... Restore the removed abnormal nodes, remove the core node of the negative weight cycle from the restored weighted directed graph, and execute the Bellman-Ford algorithm again; where the core node of the negative weight cycle is the two endpoint nodes of the edge with the largest weight value in the negative weight cycle; if a complete path can be obtained after the second execution, then the complete path is used as the usable vertebral sequence path; otherwise... Abandon auto-completion and terminate the automatic positioning process at the T12 level.

7. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 1, characterized in that, The muscle segmentation model adopts the Inf-Net deep neural network architecture, including an input transformation and fusion module, an encoder, a parallel partial decoder, an implicit inverse attention module, and an explicit edge attention module. The input of the muscle segmentation model is a multi-channel input, and the output is a preliminary segmentation probability mask for muscle segmentation at the T12 level. The input transformation and fusion module is set before the encoder and is used to convert the multi-channel input into a fused feature map before feeding it into the encoder. The construction of a multi-channel input based on the T12 thoracic vertebral level index and corresponding level images includes: Based on the T12 thoracic vertebral level index and corresponding level images, as well as standardized CT volume data, a multi-channel input tensor is constructed. The central channel of the multi-channel input tensor is the T12 level image, and the other channels are the images of the adjacent K layers above and below the T12 level, respectively. The channels are stacked in spatial order. The multi-channel input tensor is mapped to a muscle window and then scaled to obtain the multi-channel input.

8. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 1, characterized in that, The loss function used during the training of the muscle segmentation model is... Represented as: ； In the formula, , , , , These represent the category-focused loss, boundary loss, anatomical symmetry loss, auxiliary loss, and lateral output consistency loss, respectively. , , , , These represent the weight coefficients of category-focused loss, boundary loss, anatomical symmetry loss, auxiliary loss, and side-output consistency loss, respectively.

9. The method for locating the T12 thoracic vertebra and segmenting muscles according to claim 8, characterized in that, The loss of anatomical symmetry Represented as: ； In the formula, Indicates the total number of muscle categories. , These represent the image height and image width, respectively. Indicates the first Line 1 Column pixels belong to the The predicted probability of muscle-like structures This indicates a mirror transformation of the predicted probability map along the sagittal axis of the human body. Describing the L2 norm, Indicates the first Line 1 Adaptive confidence mask for column pixels.

10. A skeletal muscle index fitting system for T12 thoracic vertebra localization and muscle segmentation, characterized in that, include: The CT image input module is used to acquire chest CT plain scan sequence images and perform preprocessing to generate standardized CT volume data; The spinal segmentation module is used to generate vertebral probabilistic volume data based on the standardized CT volume data, and then extract a set of candidate center points for the vertebrae; based on the spatial distance distribution between adjacent candidate center points in the set of candidate center points for the vertebrae, a distance feature matrix is ​​generated. The T12 localization module is used to construct a weighted directed graph based on the set of candidate vertebral center points and the distance feature matrix. It employs a global optimal search strategy based on path relaxation iteration to solve for the optimal vertebral sequence path in the weighted directed graph, and then obtains the T12 thoracic vertebral level index and corresponding level image based on the optimal vertebral sequence path. The muscle segmentation module is used to construct a multi-channel input based on the T12 thoracic vertebra level index and the corresponding level image, and input it into the trained muscle segmentation model to obtain a preliminary segmentation probability mask of the thoracic and abdominal wall muscles at the T12 level. After post-processing optimization, the final muscle segmentation mask is obtained. The SMI fitting module is used to extract the muscle volume parameters at the T12 level from the final muscle segmentation mask, and, combined with patient information, output the predicted value of skeletal muscle index through a regression fitting model.

Images