Child kidney ct multi-structure segmentation and development feature vector generation method and system

Abstract

The application provides a child kidney CT multi-structure segmentation and development feature vector generation method and system, and relates to the technical field of medical image processing. The method comprises the following steps: pre-processing a child kidney CT image; constructing a multi-structure collaborative segmentation neural network to output a plurality of probability segmentation atlases of anatomical substructures; extracting a basic feature set of each substructure based on the segmentation result; constructing a dynamic multi-index correction and weighted fusion mechanism to obtain a final fusion index; and generating a high-dimensional development feature vector based on the final fusion index and standardizing the output. The application realizes fine collaborative segmentation of multiple anatomical substructures of a child kidney and standardized generation of a high-dimensional development feature vector, and provides a comprehensive data basis for child kidney development evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for multi-structure segmentation and developmental feature vector generation in pediatric kidney CT. Background Technology

[0002] Childhood is a critical stage for kidney development, during which the morphology, volume, density, and spatial location of various anatomical substructures of the kidney undergo dynamic changes with age. Accurate assessment of a child's kidney development is of significant clinical importance for the early diagnosis, treatment planning, and prognostic evaluation of diseases such as congenital kidney dysplasia and nephroblastoma. Currently, clinical practice of assessing pediatric kidney development based on CT imaging primarily relies on manually measuring a few parameters such as the kidney's long diameter, width, thickness, and cortical thickness, and making judgments based on normal reference ranges set at discrete age groups. However, manual measurement is inefficient and highly subjective; discrete reference standards cannot reflect the continuous changes in kidney development, and the limited number of geometric parameters is insufficient to comprehensively characterize the coordinated developmental information of multiple kidney structures.

[0003] In the field of automated medical image segmentation, deep learning technology has been applied to the processing of pediatric renal CT images. Most existing segmentation methods target a single object, such as the entire kidney, a renal tumor, or pyelonephritis, using three-dimensional convolutional neural networks to automatically extract the renal region or lesion region. Meanwhile, some studies have constructed renal age prediction models based on parameters such as the segmented kidney volume, achieving a certain level of prediction accuracy in the normal population. However, these methods are limited to the entire kidney or a few structures such as the cortex and medulla, failing to achieve fine and coordinated segmentation of multiple anatomical substructures such as the renal parenchyma, renal cortex, renal medulla, renal pelvis, renal artery, renal vein, and ureter. The extracted features have low dimensionality and cannot provide sufficient information for comprehensive developmental assessment.

[0004] Furthermore, research specifically targeting pediatric patients for multi-structural segmentation and developmental assessment of the kidneys is currently scarce. Due to significant differences between children and adults in organ size, morphology, and developmental stage, directly transferring segmentation models and assessment methods trained on adult CT data to pediatric data makes it difficult to guarantee segmentation accuracy and assessment precision. Tubular structures such as the renal pelvis, ureter, renal artery, and renal vein are more delicate in children, posing a greater challenge to automatic segmentation. These structures are precisely the indispensable components for constructing complete anatomical models and comprehensively assessing kidney development. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT, so as to overcome the defects of existing technologies such as single segmentation object, low feature dimension, and lack of multi-structure collaborative representation, and realize fine collaborative segmentation of multiple anatomical substructures of pediatric kidney and standardized generation of high-dimensional developmental feature vectors.

[0006] To achieve the above objectives, this invention provides a method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT scans, comprising the following steps: Step S1: Acquire CT images of the child's kidneys and preprocess the CT images to obtain a standardized CT image volume; Step S2: Construct a multi-structure collaborative segmentation neural network, using the standardized CT image volume as input, and outputting a probability segmentation map of multiple anatomical substructures of the kidney; Step S3: Based on the probability segmentation map, perform three-dimensional surface reconstruction on each anatomical substructure and extract the basic feature set of each anatomical substructure; Step S4: Based on the basic feature set, construct a dynamic multi-index correction and weighted fusion mechanism to obtain the final fusion index; Step S5: Generate a high-dimensional developmental feature vector based on the final fusion index, standardize the high-dimensional developmental feature vector, and output a standardized developmental feature vector.

[0007] This invention also provides a system for multi-structure segmentation and developmental feature vector generation in pediatric kidney CT scans, including: The image preprocessing module is used to acquire CT images of children's kidneys and preprocess them to obtain a standardized CT image volume. The multi-structure collaborative segmentation module is used to construct a multi-structure collaborative segmentation neural network, which takes the standardized CT image volume as input and outputs a probability segmentation map of multiple anatomical substructures of the kidney. The 3D surface reconstruction and feature extraction module is used to perform 3D surface reconstruction of each anatomical substructure based on the probability segmentation map and extract the basic feature set of each anatomical substructure. The dynamic multi-indicator correction and weighted fusion module is used to construct a dynamic multi-indicator correction and weighted fusion mechanism based on the basic feature set to obtain the final fused index. The high-dimensional developmental feature vector generation module is used to generate high-dimensional developmental feature vectors based on the final fusion index, standardize the high-dimensional developmental feature vectors, and output standardized developmental feature vectors.

[0008] Therefore, the present invention employs the above-mentioned method and system for multi-structure segmentation and developmental feature vector generation in pediatric renal CT scans, and the beneficial technical effects are as follows: (1) This invention constructs a multi-structure collaborative segmentation neural network to achieve fine collaborative segmentation of multiple anatomical substructures such as renal parenchyma, renal cortex, renal medulla, renal pelvis, renal artery, renal vein and ureter in pediatric kidney CT images. By introducing the adjacency and inclusion relationships between each anatomical substructure as prior knowledge in the multi-structure collaborative attention module, and using the weighted sum of the Dice loss of each structure and the boundary constraint loss between structures as the training loss function, the anatomical rationality of the segmentation results of each substructure is effectively constrained, overcoming the defects of the prior art that the segmentation object is single and only targets the whole kidney or a single lesion.

[0009] (2) This invention extracts the morphological parameters, density parameters, position parameters and symmetry parameters of each substructure based on segmentation to form a basic feature set, and realizes a comprehensive quantitative representation of multi-dimensional developmental information by comprehensively calculating the developmental maturity index within the structure, the coordination degree index of inter-structure development and the age-conditional global deviation index. Then, a high-dimensional developmental feature vector is generated by feature splicing and nonlinear transformation of the multilayer perceptron encoder. This overcomes the defects of low feature dimension and lack of multi-structure collaborative representation in the prior art, and provides a rich information basis for the assessment of children's kidney development.

[0010] (3) This invention constructs a multi-structure collaborative segmentation neural network specifically for children's data and an age-conditional feature calibration mechanism. It uses children's kidney CT image dataset for training and introduces calendar age as conditional information for feature calibration, which eliminates the interference of normal age variation on developmental assessment and avoids the problem of insufficient accuracy when directly transferring adult models to children's data. At the same time, the entire segmentation and feature generation process does not require manual intervention and realizes end-to-end automated processing, overcoming the defects of low efficiency and strong subjectivity of manual measurement. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT in this invention; Figure 2 Line graph showing the predicted age of kidney development for different age groups. Detailed Implementation

[0012] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0013] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0014] Example 1 like Figure 1 As shown, the method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT includes the following steps: Step S1: Acquire CT images of the child's kidneys, preprocess the CT images of the child's kidneys, and obtain the standardized CT image volume.

[0015] In this embodiment, the pediatric renal CT imaging data were obtained from 1000 enhanced abdominal CT scans of children aged 0-18 years collected from multiple centers, including 800 children with normal development and 200 children with clinically diagnosed renal developmental abnormalities. All CT images were acquired using a 64-slice or higher multi-slice spiral CT scanner with a tube voltage of 120kV, automatic tube current adjustment, a scan slice thickness of 0.625mm, a reconstruction slice thickness of 1.25mm, and a matrix... , vision .

[0016] The specific process for preprocessing pediatric kidney CT images is as follows: CT value normalization: The CT values ​​of the original CT images are truncated to the range of [-1024, 3071] HU to remove extreme value interference such as bone and metal artifacts. Then, the truncated CT values ​​are normalized to the interval of [0, 1] through linear mapping to facilitate neural network training.

[0017] Isotropic resampling: Because CT images have different spatial resolutions in different directions (inter-slice resolution is usually lower than intra-slice resolution), trilinear interpolation is used to resample all images to isotropic voxel size. This enables 3D convolutional neural networks to effectively extract spatial features in various directions.

[0018] Denoising: Non-local mean filtering is used for denoising, with a search window size of [size missing]. pixels, similarity window size is Pixels, filtering and smoothing parameters Set as ,in This represents the estimated standard deviation of image noise. This filtering method effectively suppresses quantum and thermal noise in CT images while preserving the edge details of each substructure of the kidney.

[0019] The volumetric size of the preprocessed standardized CT images is uniformly set to... Voxels facilitate the batch input of data into neural networks for training and inference.

[0020] Step S2: Construct a multi-structure collaborative segmentation neural network, using the standardized CT image volume as input, and outputting a probability segmentation map of multiple anatomical substructures of the kidney.

[0021] In this embodiment, the specific structure of the multi-structure collaborative segmentation neural network is as follows: (a) Encoder module.

[0022] The encoder module uses a 3D U-Net architecture as its backbone network, comprising five coding layers. Each coding layer contains two... Convolutional layers (stride 1, padding 1), each followed by a batch normalization layer and a ReLU activation function, then a... Max pooling layers (with a stride of 2) are used for downsampling. The number of feature channels in each coding layer is 32, 64, 128, 256, and 512, respectively. The encoder module performs multi-scale semantic feature extraction on the input normalized CT image volume and outputs multi-layer coded feature maps. The spatial resolution of each feature map is as follows: , , , , .

[0023] (ii) Multi-structure collaborative attention module.

[0024] The multi-structure collaborative attention module receives the deepest encoded feature map output by the encoder module. The structural relationship graph is constructed using the adjacency and inclusion relationships between the various anatomical substructures as prior knowledge. Specifically, the structural relationship graph uses seven anatomical substructures (renal parenchyma, renal cortex, renal medulla, renal pelvis, renal artery, renal vein, and ureter) as nodes, and the spatial adjacency and inclusion relationships between the substructures as edges. The weights of the edges are set according to prior anatomical knowledge: the weight of inclusion relationship is 0.8 (e.g., the renal cortex is included in the renal parenchyma), the weight of adjacency relationship is 0.5 (e.g., the renal artery and renal vein run together), and the weight of no direct relationship is 0.

[0025] Spatial context information is passed between anatomical substructures through a graph attention network. The graph attention network consists of two graph attention layers, each with four attention heads, and each attention head outputs a feature dimension of 64. After modeling the spatial dependencies between anatomical substructures, enhanced multi-scale structural association features are output, and the feature maps are restored to the spatial resolution corresponding to each layer of the encoder through deconvolution.

[0026] (iii) Decoder module.

[0027] The decoder module receives enhanced multi-scale structural correlation features and fuses them step-by-step with the encoded features of the corresponding layers of the encoder module via skip connections. The decoder contains five decoding layers, each of which first... Deconvolution (with a stride of 2) is used for upsampling, then concatenated with the feature map of the corresponding layer of the encoder along the channel dimension, and then passed through two... Convolutional layers (stride 1, padding 1), each followed by a batch normalization layer and a ReLU activation function. The number of feature channels in each decoding layer are 512, 256, 128, 64, and 32, respectively. Finally, the original input resolution is gradually restored, and the decoded feature map is output.

[0028] (iv) Multi-task output module.

[0029] The multi-task output module receives the decoded feature map and outputs the probability segmentation map of each anatomical substructure through seven parallel segmentation heads. Each segmentation head contains a... The convolutional layer maps the 32-channel feature map into a 2-channel output (background and target), and then uses the Softmax activation function to obtain the probability value of each voxel belonging to the substructure.

[0030] (v) Training of multi-structure collaborative segmentation neural networks.

[0031] In this embodiment, the multi-structure collaborative segmentation neural network adopts the following training configuration: Training dataset: 1000 labeled data points were divided into a training set (700 cases), a validation set (150 cases), and a test set (150 cases) in a ratio of 7:1.5:1.5. Annotations were manually performed layer by layer by three senior radiologists using ITK-SNAP software, and the consistency of the annotations was verified by the Kappa coefficient test (Kappa>0.85).

[0032] Optimizer: Adam optimizer, initial learning rate 1×10⁻⁶ -4 Weight decay 1×10 -5 Batch size: 2 (due to the large size of 3D images). Number of training epochs: 200 epochs. Early stopping strategy: Training stops when the validation set loss does not decrease for 20 consecutive epochs. Data augmentation: Random rotation (±10°), random scaling (±5%), random elastic deformation, and random brightness / contrast adjustment are used to enhance the model's generalization ability.

[0033] The training loss function of a multi-structure collaborative segmentation neural network is a weighted sum of the Dice loss of each anatomical substructure and the boundary constraint loss between structures. The total loss function of the multi-structure collaborative segmentation neural network is... The calculation formula is: ; in, This represents the number of anatomical substructures (7 in this embodiment). For the first A true label for an anatomical substructure. For the first Predicted probability segmentation map of each anatomical substructure. The Dice similarity loss for each anatomical substructure, The loss is the boundary constraint loss between structures. and These are the weighting coefficients.

[0034] Dice loss weighting coefficients for each anatomical substructure The values ​​of each substructure are set according to the voxel proportion and clinical importance, as shown in Table 1.

[0035] Table 1

[0036] inter-structural boundary constraint loss weighting coefficient The value is set to 0.3 to penalize predictions that overlap or have excessive gaps between adjacent substructures. The boundary constraint loss is calculated based on the gradient magnitude differences of each substructure boundary, encouraging adjacent substructure boundaries to maintain reasonable spatial topological relationships.

[0037] (vi) Example of segmentation results.

[0038] The segmentation accuracy (Dice similarity coefficient) of each anatomical substructure on the test set of 150 cases is shown in Table 2.

[0039] Table 2

[0040] Although the segmentation accuracy of tubular structures such as the renal pelvis, renal artery, renal vein, and ureter is lower than that of large-volume structures such as renal parenchyma, it still reaches a clinically acceptable level, verifying the effectiveness of the multi-structure collaborative attention module in segmenting small structures.

[0041] Step S3: Based on the probability segmentation map, perform three-dimensional surface reconstruction of each anatomical substructure and extract the basic feature set of each anatomical substructure.

[0042] In this embodiment, based on the probability segmentation map of each anatomical substructure output in step S2, a threshold of 0.5 is first used to segment the probability... Figure 2 Values ​​are converted to obtain voxel labels for each substructure. Then, the Marching Cubes algorithm is used to reconstruct the 3D surface of each substructure, with the isosurface threshold set to 0.5, to reconstruct the triangular mesh surface model of each substructure for calculating morphological parameters.

[0043] The basic feature set includes morphological parameters, density parameters, positional parameters, and symmetry parameters, which are extracted in the following ways: Morphological parameters: The volume of each anatomical substructure (in mL) was calculated based on voxel count multiplied by voxel volume; the surface area of ​​each anatomical substructure (in mm) was calculated based on the reconstructed surface triangular mesh. 2Principal component analysis was performed on each substructure voxel to calculate the maximum, median, and minimum diameters (in mm) of each anatomical substructure along the three principal axes, which were denoted as major diameter, width diameter, and thickness, respectively. Sphericity was determined using the formula... Calculation, where For volume, The surface area is sphericity, which ranges between 0 and 1; the eccentricity is calculated based on the eigenvalues ​​of principal component analysis and reflects the degree to which the substructure deviates from the spherical shape.

[0044] Density parameters: Calculate the CT values ​​of all voxels within each anatomical substructure, and calculate the mean CT value (unit: HU), standard deviation of CT value (unit: HU), skewness of CT value distribution (reflecting the symmetry of density distribution), and kurtosis of CT value distribution (reflecting the sharpness of density distribution).

[0045] Position parameters: The three-dimensional centroid coordinates (unit: mm) are calculated based on the average coordinates of all voxels of each anatomical substructure; the centroid offset is defined as the magnitude of the offset vector of each substructure centroid relative to the overall centroid of the kidney (unit: mm); the relative position vector takes the center of the renal hilum as the origin and records the three-dimensional relative coordinates of the centroids of each substructure.

[0046] Symmetry parameters: Calculate the volume ratio of corresponding anatomical substructures of the left and right kidneys respectively. and morphological similarity ,in and These are the volumes of the corresponding anatomical substructures of the left and right kidneys, respectively. and These are the three-dimensional shape descriptor vectors for the corresponding anatomical substructures of the left and right kidneys, respectively.

[0047] The following is an example of the basic characteristic set of a normal 8-year-old male child (Table 3).

[0048] Table 3

[0049] Step S4: Based on the basic feature set, construct a dynamic multi-index correction and weighted fusion mechanism to obtain the final fusion index.

[0050] The dynamic multi-index correction and weighted fusion mechanism in step S4 specifically includes: calculating the developmental maturity index within the structure. Interstructural Coordination Development Indicators and age-conditional global deviation index Global deviation index based on age condition The numerical values ​​determine the correction strategy, and the corrected intra-structural developmental maturity index and the corrected inter-structural coordinated developmental coordination index are weighted and fused to obtain the final fused index. ; Intrastructural developmental maturity index The calculation formula is: ; in, =7 represents the number of anatomical substructures. =8 represents the feature dimension of each anatomical substructure. For the first The first anatomical substructure One morphological parameter, For the corresponding weight coefficients, and .

[0051] Interstructural Coordination Development Indicators The calculation formula is: ; in, =12 represents the number of structural relationship features. For the first The inter-structural relationship features include the relative volume ratio between structures, the distance between the centroids of structures, the surface contact area between structures, and the CT value gradient features between structures. For the corresponding weight coefficients, and ; Age-conditional global deviation index The calculation formula is: ; in, For distance measurement function, The basic feature vector (56 dimensions, 7 substructures) is composed of a basic feature set. 8 features) The calendar age of the child being tested. For calendar age The expected basic feature vector on the corresponding normal developmental trajectory.

[0052] Normal developmental trajectory The model was determined using an age-feature regression model pre-built on a dataset of 800 normal children. Specifically, for each baseline feature (e.g., left renal parenchymal volume), a cubic spline regression model was built using calendrical age as the independent variable and the feature value as the dependent variable (K=5 nodes, with node positions uniformly distributed according to age quantiles). For calendrical age... , This is a vector of regression predictions for all 56 features.

[0053] In this embodiment, the basic feature vector of a normally developing boy who is 8 years old (calendar age) is... Euclidean distance between =0.342. Normal development in children of different ages. The reference range for the values ​​is shown in Table 4.

[0054] Table 4

[0055] In this embodiment, the first threshold =0.55, second threshold =0.85, third threshold =1.20. The above threshold is based on 800 normal children. Determining the percentile of a value: Corresponding to the 95th percentile, Corresponding to the 99th percentile, It corresponds to the 99.9th percentile.

[0056] Global deviation index based on age condition The correction strategy is determined by the preset discrimination interval into which the value falls: When age is conditionalized, the global deviation index When the value is less than or equal to the first threshold, it is determined to be a normal developmental pattern. No correction is made to the intrastructural developmental maturity index or the interstructural collaborative development coordination index, and weighted fusion is directly performed. When age is conditionalized, the global deviation index When the value is greater than the first threshold and less than or equal to the second threshold, it is determined to be a dominant developmental deviation pattern within the structure, and the global deviation index is conditionalized using age. Intrastructural developmental maturity indicators Make individual corrections; When age is conditionalized, the global deviation index When the value is greater than the second threshold and less than or equal to the third threshold, it is determined to be a dominant interstructural developmental deviation pattern, and the global deviation index is conditionalized using age. Indicators of interstructural synergistic development Make individual corrections; When age is conditionalized, the global deviation index When the value exceeds the third threshold, it is determined to be a global developmental deviation pattern, and the global deviation index is conditionally categorized based on age. Simultaneously, the developmental maturity index within the structure was analyzed. and interstructural coordination development index Make corrections.

[0057] Without making any corrections to the intrastructural developmental maturity index and the interstructural co-developmental coordination index, let , ; Using age-conditional global deviation indicators Intrastructural developmental maturity indicators The specific method for making individual corrections is as follows: ; in, This is the revised intrastructural developmental maturity index. A global deviation index based on age conditions The correction amount for the calculated developmental maturity index within the structure. The first correction factor is set to 0.5 in this embodiment; Using age-conditional global deviation indicators Indicators of interstructural synergistic development The specific method for making individual corrections is as follows: ; in, This is the revised index of interstructural co-development coordination. A global deviation index based on age conditions The correction amount for the calculated inter-structural co-development coordination index. The second correction factor is set to 0.5 in this embodiment; Using age-conditional global deviation indicators Simultaneously, the developmental maturity index within the structure was analyzed. and interstructural coordination development index The specific method for making the correction is as follows: ; ; in, Add a correction to the intrastructural developmental maturity index under the global developmental deviation pattern. Add a correction factor to the inter-structural coordinating developmental coordination index under the global developmental deviation mode. (Value is 0.8) and (A value of 0.8) is the corresponding additional correction factor; , , and These are age-conditional global deviation indicators. The calculated correction function for the corresponding indicator, and the conditional global deviation of each correction function from the indicator with age. It increases rather than linearly.

[0058] The specific method for weighted fusion of the revised intra-structural developmental maturity index and the revised inter-structural collaborative developmental coordination index is as follows: ; in, Weighting of developmental maturity indices within the structure. The fusion weights are used for the coordination degree index of inter-structural development, and ; Intrastructural developmental maturity index fusion weighting The fusion weight of the inter-structural synergistic development coordination index The determination method is as follows: ; ; in, To conditionalize global deviation indicators by age It is a monotonically decreasing function of the independent variable; When it is determined to be a dominant developmental deviation pattern within the structure ,make The coordination degree index of inter-structural synergistic development has a higher weight in the fusion process; When it is determined to be a dominant developmental deviation pattern between structures ,make The developmental maturity index within the structure has a higher weight in the fusion process; When it is determined to be a global developmental deviation pattern ,make ; When it is determined to be a normal developmental pattern , and The difference between them does not exceed 0.2.

[0059] Step S5: Generate a high-dimensional developmental feature vector based on the final fusion index, standardize the high-dimensional developmental feature vector, and output a standardized developmental feature vector.

[0060] In step S5, based on the final fusion index The specific process of generating high-dimensional developmental feature vectors involves: merging the final fusion index... The structural feature encoding vectors of each anatomical substructure Inter-structural relationship vectors and age-calibrated residual characteristics The data is then stitched together and subjected to a nonlinear transformation using a multilayer perceptron encoder. ; in, It is a multilayer perceptron encoder; High-dimensional developmental feature vectors; age-calibrated residual features The residual vector between the basic features of each anatomical substructure and the expected age feature benchmark is composed of the basic feature set and the expected basic feature vector. The high-dimensional developmental feature vector is obtained by element-wise subtraction; the high-dimensional developmental feature vector is then standardized. Perform Z-score standardization.

[0061] Multilayer perceptron encoder The structure is shown in Table 5.

[0062] Table 5

[0063] The multilayer perceptron encoder is trained using an autoencoder architecture, with unsupervised pre-training on a normal children's dataset. The reconstruction loss is the mean squared error, the training epochs are 100, and the learning rate is 1×10⁻⁶. -3 In this embodiment, after inputting a 97-dimensional concatenated feature vector, a 32-dimensional high-dimensional developmental feature vector is output.

[0064] This embodiment visualizes and verifies the 32-dimensional standardized developmental feature vector output by the above-mentioned 8-year-old normally developing boy. t-SNE dimensionality reduction (30 perplexity, 1000 iterations) was performed on 150 samples in the test set. The results show that: normally developing samples of the same age group form tight clusters in the feature space, and the cluster centers of each age group are smoothly and continuously arranged along the age direction, verifying that the generated high-dimensional developmental feature vector has good age discrimination ability and continuous developmental representation ability.

[0065] The following are the developmental age prediction results of this embodiment on 150 samples in the test set (Table 6).

[0066] Table 6

[0067] like Figure 2 The linear regression results of predicted developmental age and calendar age for each age group showed a high linear correlation between the two. The slope and intercept of the regression line were close to the ideal values, indicating that the standardized developmental feature vector generated by this method can accurately reflect the true developmental status of children's kidneys at different age groups.

[0068] The above results show that the method of the present invention can stably generate standardized developmental feature vectors with continuous developmental characterization capabilities, with low prediction errors in all age groups, and the error distribution is uniform with no systematic bias.

[0069] Example 2 A system for multi-structure segmentation and developmental feature vector generation in pediatric renal CT scans, including: The image preprocessing module is used to acquire CT images of children's kidneys and preprocess them to obtain a standardized CT image volume. The multi-structure collaborative segmentation module is used to construct a multi-structure collaborative segmentation neural network, which takes the standardized CT image volume as input and outputs a probability segmentation map of multiple anatomical substructures of the kidney. The 3D surface reconstruction and feature extraction module is used to perform 3D surface reconstruction of each anatomical substructure based on the probability segmentation map and extract the basic feature set of each anatomical substructure. The dynamic multi-indicator correction and weighted fusion module is used to construct a dynamic multi-indicator correction and weighted fusion mechanism based on the basic feature set to obtain the final fused index. The high-dimensional developmental feature vector generation module is used to generate high-dimensional developmental feature vectors based on the final fusion index, standardize the high-dimensional developmental feature vectors, and output standardized developmental feature vectors.

[0070] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0071] Therefore, the present invention employs the above-mentioned method and system for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT, which can achieve fine collaborative segmentation of multiple anatomical substructures, automatic extraction of multi-dimensional quantitative features, and efficient generation of standardized developmental feature vectors, providing a comprehensive data foundation for the assessment of pediatric kidney development.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT, characterized in that, Includes the following steps: Step S1: Acquire CT images of the child's kidneys and preprocess the CT images to obtain a standardized CT image volume; Step S2: Construct a multi-structure collaborative segmentation neural network, using the standardized CT image volume as input, and outputting a probability segmentation map of multiple anatomical substructures of the kidney; Step S3: Based on the probability segmentation map, perform three-dimensional surface reconstruction on each anatomical substructure and extract the basic feature set of each anatomical substructure; Step S4: Based on the basic feature set, construct a dynamic multi-index correction and weighted fusion mechanism to obtain the final fusion index; Step S5: Generate a high-dimensional developmental feature vector based on the final fusion index, standardize the high-dimensional developmental feature vector, and output a standardized developmental feature vector.

2. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 1, characterized in that, Step S1 involves preprocessing the pediatric kidney CT images, including: CT value normalization, isotropic resampling to a uniform voxel size, and denoising using nonlocal mean filtering.

3. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 1, characterized in that, The multi-structure collaborative segmentation neural network in step S2 includes an encoder module, a multi-structure collaborative attention module, a decoder module, and a multi-task output module; The encoder module uses a three-dimensional convolutional neural network to extract multi-scale semantic features from the input standardized CT image volume and outputs a multi-layer coded feature map. The multi-structure collaborative attention module receives the multi-layer encoded feature map output by the encoder module, uses the adjacency and inclusion relationships between each anatomical substructure as prior knowledge to construct a structural relationship graph, transmits spatial context information between each anatomical substructure through the graph attention network, models the spatial dependency relationship between each anatomical substructure, and outputs enhanced multi-scale structural association features. The decoder module receives enhanced multi-scale structural correlation features, and gradually restores the spatial resolution by fusing the multi-layer encoded feature maps of the corresponding layers of the encoder module through skip connections, and outputs the decoded feature map. The multi-task output module receives the decoded feature map and outputs the probability segmentation map of each anatomical substructure through multiple segmentation heads; Multiple anatomical substructures include the renal parenchyma, renal cortex, renal medulla, renal pelvis, renal artery, renal vein, and ureter; The training loss function of a multi-structure collaborative segmentation neural network is a weighted sum of the Dice loss of each anatomical substructure and the boundary constraint loss between structures.

4. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 1, characterized in that, The basic feature set includes morphological parameters, density parameters, positional parameters, and symmetry parameters; Morphological parameters include the volume of each anatomical substructure, the surface area of ​​each anatomical substructure, the maximum diameter of each anatomical substructure, the minimum diameter of each anatomical substructure, the sphericity of each anatomical substructure, and the eccentricity of each anatomical substructure. Density parameters include the average CT value within each anatomical substructure, the standard deviation of the CT value within each anatomical substructure, the skewness of the CT value distribution within each anatomical substructure, and the kurtosis of the CT value distribution within each anatomical substructure. The positional parameters include the three-dimensional centroid coordinates of each anatomical substructure, the centroid offset of each anatomical substructure, and the relative position vector of each anatomical substructure. Symmetry parameters include the volume ratio of the corresponding anatomical substructures of the left and right kidneys and the morphological similarity of the corresponding anatomical substructures of the left and right kidneys.

5. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 1, characterized in that, The dynamic multi-index correction and weighted fusion mechanism in step S4 specifically includes: calculating the developmental maturity index within the structure. Interstructural coordination development index and age-conditional global deviation index Global deviation index based on age condition The numerical values ​​determine the correction strategy, and the corrected intra-structural developmental maturity index and the corrected inter-structural coordinated developmental coordination index are weighted and fused to obtain the final fused index. ; Intrastructural developmental maturity index The calculation formula is: ； in, This represents the number of anatomical substructures. These are the characteristic dimensions of each anatomical substructure. For the first The first anatomical substructure One morphological parameter, For the corresponding weight coefficients, and ; Interstructural Coordination Development Indicators The calculation formula is: ； in, This represents the number of features relating the structures. For the first The inter-structural relationship features include the relative volume ratio between structures, the distance between the centroids of structures, the surface contact area between structures, and the CT value gradient features between structures. For the corresponding weight coefficients, and ; Age-conditional global deviation index The calculation formula is: ； in, For distance measurement function, The basic feature vector is composed of the basic feature set. The calendar age of the child being tested. For calendar age The expected basic feature vector on the corresponding normal developmental trajectory.

6. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 5, characterized in that, Global deviation index based on age condition The correction strategy is determined by the preset discrimination interval into which the value falls: When age is conditionalized, the global deviation index When the value is less than or equal to the first threshold, it is determined to be a normal developmental pattern. No correction is made to the intrastructural developmental maturity index or the interstructural collaborative development coordination index, and weighted fusion is directly performed. When age is conditionalized, the global deviation index When the value is greater than the first threshold and less than or equal to the second threshold, it is determined to be a dominant developmental deviation pattern within the structure, and the global deviation index is conditionalized using age. Indicators of developmental maturity within the structure Make individual corrections; When age is conditionalized, the global deviation index When the value is greater than the second threshold and less than or equal to the third threshold, it is determined to be a dominant interstructural developmental deviation pattern, and the global deviation index is conditionalized using age. Indicators of interstructural synergistic development Make individual corrections; When age is conditionalized, the global deviation index When the value exceeds the third threshold, it is determined to be a global developmental deviation pattern, and the global deviation index is conditionally categorized based on age. Simultaneously, the developmental maturity index within the structure was analyzed. and interstructural coordination development index Make corrections; Among them, the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

7. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 6, characterized in that, Without making any corrections to the intrastructural developmental maturity index and the interstructural co-developmental coordination index, let , ; Using age-conditional global deviation index Indicators of developmental maturity within the structure The specific method for making individual corrections is as follows: ； in, This is the revised intrastructural developmental maturity index. A global deviation index based on age conditions The correction amount for the calculated intrastructural developmental maturity index. This is the first correction factor; Using age-conditional global deviation index Indicators of interstructural synergistic development The specific method for making individual corrections is as follows: ； in, This is the revised index of interstructural co-development coordination. A global deviation index based on age conditions The correction amount for the calculated inter-structural co-development coordination index. This is the second correction factor; Using age-conditional global deviation index Simultaneously, the developmental maturity index within the structure was analyzed. and interstructural coordination development index The specific method for making the correction is as follows: in, Add a correction to the intrastructural developmental maturity index under the global developmental deviation pattern. Add a correction factor to the inter-structural coordinating developmental coordination index under the global developmental deviation mode. and This is the corresponding additional correction factor.

8. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 6, characterized in that, The specific method for weighted fusion of the revised intra-structural developmental maturity index and the revised inter-structural collaborative developmental coordination index is as follows: ； in, Weighting of developmental maturity indices within the structure. The fusion weights are used to determine the coordination degree index of inter-structural synergistic development. Intrastructural developmental maturity index fusion weighting The fusion weight of the inter-structural synergistic development coordination index The determination method is as follows: ， ； in, To conditionalize global deviation indicators by age It is a monotonically decreasing function of the independent variable; when it is determined to be a dominant developmental deviation pattern within the structure, ,make When determined to be a dominant developmental deviation pattern between structures, ,make When it is determined to be a global developmental deviation pattern, ,make When it is determined to be a normal developmental pattern, , and The difference between them does not exceed 0.

2.

9. The method for multi-structure segmentation and developmental feature vector generation of pediatric kidney CT according to claim 1, characterized in that, In step S5, based on the final fusion index The specific process of generating high-dimensional developmental feature vectors involves: merging the final fusion index... The structural feature encoding vectors of each anatomical substructure Inter-structural relationship vectors and age-calibrated residual characteristics The data is then stitched together and subjected to a nonlinear transformation using a multilayer perceptron encoder. ； in, It is a multilayer perceptron encoder. It is a high-dimensional developmental feature vector; Standardizing high-dimensional developmental feature vectors results in high-dimensional developmental feature vectors. Perform Z-score standardization.

10. A system for multi-structure segmentation and developmental feature vector generation in pediatric kidney CT scans, characterized in that, include: The image preprocessing module is used to acquire CT images of children's kidneys and preprocess them to obtain a standardized CT image volume. The multi-structure collaborative segmentation module is used to construct a multi-structure collaborative segmentation neural network, which takes the standardized CT image volume as input and outputs a probability segmentation map of multiple anatomical substructures of the kidney. The 3D surface reconstruction and feature extraction module is used to perform 3D surface reconstruction of each anatomical substructure based on the probability segmentation map and extract the basic feature set of each anatomical substructure. The dynamic multi-indicator correction and weighted fusion module is used to construct a dynamic multi-indicator correction and weighted fusion mechanism based on the basic feature set to obtain the final fused index. The high-dimensional developmental feature vector generation module is used to generate high-dimensional developmental feature vectors based on the final fusion index, standardize the high-dimensional developmental feature vectors, and output standardized developmental feature vectors.

Images