Cortical bone parameter quantification analysis method and system based on three-dimensional medical images

Abstract

The application discloses a femur cortical bone parameter quantitative analysis method and system based on a three-dimensional medical image, relates to the technical field of medical image processing and medical image intelligent analysis, and comprises the following steps: acquiring an initial femur three-dimensional medical image and performing pretreatment; segmenting the pretreated femur three-dimensional medical image to generate a bone tissue mask of a femur region and separate a cortical bone three-dimensional mask; generating a three-dimensional thickness distribution map of the femur cortical bone; extracting a global depth feature vector and a local statistical feature vector representing a three-dimensional space distribution mode of the cortical bone in the three-dimensional thickness distribution map, and combining the feature vectors to splice a comprehensive feature vector; and performing quantitative analysis on the comprehensive feature vector to generate a quantitative analysis result for representing a femur cortical bone thickness parameter space distribution mode. The application can automatically, accurately and comprehensively analyze the cortical bone thickness features in the femur three-dimensional CT image, and better realizes parameterization and quantization of the femur cortical bone thickness features.

Description

Technical Field

[0001] This application relates to the field of medical image processing and intelligent medical image analysis technology, specifically to a method and system for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images. Background Technology

[0002] Decreased bone density and degeneration of bone microstructure in weight-bearing bones such as the femur directly lead to a decline in skeletal structural integrity. Among these changes, deterioration of the bone structure in the hip region (such as the femoral neck and intertrochanteric region) is a core factor causing insufficient bone load-bearing capacity and reduced structural stability, significantly impacting both skeletal structural integrity and mechanical stability.

[0003] Currently, dual-energy X-ray absorptiometry (DXA) is a routine technique for bone mineral density testing, primarily used to obtain data related to area bone mineral density. However, this method has significant limitations: its results only reflect the overall bone density level and cannot accurately characterize the three-dimensional spatial distribution characteristics, geometric morphological parameters, and microstructural changes of cortical bone; relying solely on density test results makes it difficult to comprehensively characterize the true structural state of the bone, and mismatches between density test values ​​and actual bone structural conditions are prone to occur.

[0004] With the development of quantitative CT technology, it is now possible to accurately acquire three-dimensional structural information of bones, providing a reliable technical foundation for the quantitative analysis of bone structural parameters. As the outer dense support structure of weight-bearing bones such as the femur, subtle changes in the thickness parameters, porosity characteristics, and spatial distribution of cortical bone are core structural characterization indicators for quantifying bone structural integrity and analyzing trends in bone mechanical properties.

[0005] However, traditional methods for detecting and analyzing cortical bone thickness mostly rely on manual or semi-automatic operation, which has many technical shortcomings in practical applications:

[0006] 1) The detection process is cumbersome and inefficient: It requires professionals to manually draw and mark on multi-layer two-dimensional tomographic images, which is complicated and time-consuming, making it difficult to meet the needs of efficient detection of large batches of samples.

[0007] 2) Poor consistency of test results: The test results are highly dependent on the subjective experience of the operators, and the repeatability and consistency of measurements between different operators are low, lacking standardized quantitative basis.

[0008] 3) Incomplete extraction of structural features: Existing methods mostly only calculate the mean of thickness parameters in preset local areas, which cannot fully extract the complex morphological heterogeneity features of cortical bone in three-dimensional space, making it difficult to comprehensively and objectively reflect the overall structural state of the skeleton.

[0009] 4) Insufficient adaptability of quantitative analysis models: Existing bone parameter analysis models mostly use traditional statistical methods or simple machine learning algorithms, which cannot efficiently model the nonlinear mapping relationship between the high-dimensional features of three-dimensional bones and the structural state, resulting in insufficient accuracy and comprehensiveness in the quantitative analysis of bone structural parameters.

[0010] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0011] To address the problems in related technologies, this application proposes a method and system for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, in order to solve the aforementioned technical problems existing in the existing related technologies.

[0012] Therefore, the specific technical solution adopted in this application is as follows:

[0013] Firstly, this application proposes a method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, including:

[0014] S1. Obtain the initial three-dimensional medical image of the femur to be analyzed and perform preprocessing to obtain the preprocessed three-dimensional medical image of the femur.

[0015] S2. Segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional mask of the cortical bone; calculate the local thickness of the cortical bone based on the three-dimensional mask of the cortical bone to generate a three-dimensional thickness distribution map of the femoral cortical bone.

[0016] S3. Based on the feature extraction model and the target region of interest, extract the global depth feature vector and local statistical feature vector that represent the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map, and combine the feature vectors to obtain the comprehensive feature vector.

[0017] The feature extraction model is self-supervised pre-training based on anatomical perception occlusion and multi-scale reconstruction strategies.

[0018] S4. Using the pre-trained deep learning model, the comprehensive feature vector is quantitatively analyzed to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0019] Further, the segmented preprocessed three-dimensional medical image of the femur generates a bone tissue mask for the femoral region, and the bone tissue mask is separated using an algorithm combining region growing and level set evolution to obtain a three-dimensional mask of cortical bone; based on the three-dimensional mask of cortical bone, the local thickness of cortical bone is calculated, and a three-dimensional thickness distribution map of the femoral cortical bone is generated, including:

[0020] S21. Based on the 3D nnU-Net framework, residual connections and attention gate mechanism, a lightweight segmentation model is constructed and supervised training is performed using a femur 3D medical image dataset with cortical bone segmentation annotations.

[0021] S22. Use the trained lightweight segmentation model to perform region segmentation on the preprocessed femoral 3D medical image dataset to obtain the bone tissue mask of the femoral region.

[0022] S23. Based on an algorithm combining region growing and level set evolution, the cortical bone region in the bone tissue mask is separated to obtain a three-dimensional cortical bone mask.

[0023] S24. Using the three-dimensional Euclidean distance transformation method, calculate the local thickness of the cortical bone at each voxel point in the three-dimensional mask of cortical bone, and generate a three-dimensional thickness distribution map of the femoral cortical bone based on the local thickness of the cortical bone of all voxels.

[0024] Furthermore, the algorithm based on the combination of region growing and level set evolution separates the cortical bone region in the bone tissue mask, resulting in a three-dimensional cortical bone mask including:

[0025] S231. Based on the gray value distribution of computed tomography scan, an initial seed point is automatically selected within a preset range of the endosteal membrane in the bone tissue mask.

[0026] S232. Using a three-dimensional region growth algorithm based on grayscale threshold, starting from the initial seed point, the region is gradually grown towards adjacent voxels to initially separate the cortical bone region and obtain a cortical bone roughness mask.

[0027] S233. Using the roughened mask of cortical bone as the initial contour for the evolution of the level set, the level set algorithm based on edge and region information is used to generate the three-dimensional mask of cortical bone.

[0028] Furthermore, the energy functional expression for the evolution of the level set is:

[0029] ;

[0030] In the formula, Let be the energy functional used to control the evolution of the level set in the contour evolution process. H( is the level set function) ) is the Heaviside function used to distinguish between the inner and outer regions of a contour. Let Ω be the gradient of the Heaviside function used to represent the contour boundary, and Ω be the image domain. Let I be the gradient vector of the CT grayscale image I, representing the direction and intensity of grayscale changes in space. Let μ be an edge detection function used to reduce energy at image edges and guide the contour to fit the anatomical boundary, where μ, λ, and ν are weight parameters that control the strength of the contour length smoothing term, the edge attraction term, and the region area constraint term, respectively.

[0031] Furthermore, the step of calculating the local cortical bone thickness at each voxel point in the three-dimensional cortical bone mask using the three-dimensional Euclidean distance transformation method, and generating a three-dimensional thickness distribution map of the femoral cortical bone based on the local cortical bone thickness of all voxels, includes:

[0032] S241. Using three-dimensional morphological erosion operations, the boundary of the bone marrow cavity in the three-dimensional mask of cortical bone is extracted to obtain the voxel set of the boundary of the bone marrow cavity.

[0033] S242. Extract the outer boundary of the bone in the three-dimensional mask of cortical bone through three-dimensional morphological dilatation operation to obtain the voxel set of the outer boundary of the bone;

[0034] S243. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the boundary of the medullary cavity, and select the minimum distance as the Euclidean distance from that voxel to the nearest surface of the medullary cavity.

[0035] S244. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the outer boundary of bone, and select the minimum distance as the Euclidean distance from that voxel to the nearest outer surface of bone.

[0036] S245. Calculate the local cortical bone thickness at each voxel point based on the Euclidean distance from the voxel to the nearest inner surface of the medullary cavity and the Euclidean distance from the voxel to the nearest outer surface of the bone.

[0037] S246. Map the local thickness values ​​of cortical bone at all voxel points back to their corresponding positions in the original three-dimensional image space to form a three-dimensional thickness distribution map of the femoral cortical bone.

[0038] Furthermore, based on the feature extraction model and the target region of interest, the global depth feature vector and local statistical feature vector representing the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map are extracted, and the comprehensive feature vector obtained by concatenating the feature vectors includes:

[0039] S31. Based on the three-dimensional deep convolutional autoencoder architecture, a feature extraction model is constructed, and self-supervised pre-training is performed using an anatomically perceptual occlusion and multi-scale reconstruction strategy.

[0040] S32. Using a pre-trained feature extraction model, extract the global depth feature vector of the overall distribution pattern of cortical bone thickness in three-dimensional space from the three-dimensional thickness distribution map.

[0041] S33. Based on the preset target region of interest, extract the statistical thickness features within the target region of interest in the 3D thickness distribution map to form a local statistical feature vector;

[0042] S34. Concatenate the global deep feature vector with the local statistical feature vector to obtain the comprehensive feature vector.

[0043] Furthermore, the self-supervised pre-training using an anatomically-aware occlusion and multi-scale reconstruction strategy includes:

[0044] Construct a multi-level region of interest probability map corresponding to the femoral anatomy, and assign different initial occlusion probabilities to regions at different levels in the multi-level region of interest probability map.

[0045] Based on the multi-level region of interest probability map, the three-dimensional thickness distribution map of the input femoral cortical bone is generatively sampled to obtain the initial masking layer.

[0046] The initial masking mask is optimized using a three-dimensional conditional random field model to generate spatially coherent masking blocks. The original three-dimensional thickness distribution map is then masked on a voxel-by-voxel basis using the spatially coherent masking blocks to obtain the masked thickness map.

[0047] A 3D deep convolutional autoencoder is trained using the occluded thickness map for reconstruction. The reconstruction loss function includes pixel-level reconstruction loss and structural consistency loss based on multi-scale feature matching.

[0048] Furthermore, the step of using a pre-trained deep learning model to perform quantitative analysis on the comprehensive feature vector, generating quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, thereby achieving parameterized quantification of femoral cortical bone thickness characteristics, includes:

[0049] S41. Construct a deep learning model based on a multi-layer fully connected neural network and perform supervised training using a dataset containing comprehensive features and quantization results.

[0050] S42. Input the comprehensive feature vector into the trained deep learning model to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0051] Furthermore, the loss function of the deep learning model is a binary classification focus loss function, and the expression of the binary classification focus loss function is:

[0052] ;

[0053] In the formula, L focal Let y represent the binary classification focus loss function, where N represents the number of samples, and y represents the number of samples.i p represents the true class label of the i-th sample. i Let represent the probability that the i-th sample belongs to the positive class, α represent the class weight coefficient of the positive class, and (1-α) represent the class weight coefficient of the negative class. This indicates the focus parameter.

[0054] Secondly, this application also provides a system for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, including:

[0055] The femoral image preprocessing module is used to acquire the initial three-dimensional medical image of the femur to be analyzed and to preprocess it to obtain the preprocessed three-dimensional medical image of the femur.

[0056] The cortical bone thickness calculation module is used to segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional cortical bone mask; based on the three-dimensional cortical bone mask, the local thickness of the cortical bone is calculated to generate a three-dimensional thickness distribution map of the femoral cortical bone.

[0057] The femoral feature extraction and fusion module is used to extract global depth feature vectors and local statistical feature vectors that characterize the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map based on the feature extraction model and the target region of interest. The feature vectors are then combined to obtain a comprehensive feature vector. The feature extraction model is self-supervised pre-trained based on the anatomical perception occlusion and multi-scale reconstruction strategy.

[0058] The quantitative analysis module is used to perform quantitative analysis on the comprehensive feature vector using a pre-trained deep learning model, and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0059] The beneficial effects of this application are as follows:

[0060] 1) This application realizes the fully automated processing from CT images to quantitative analysis of femoral cortical bone parameters, thereby eliminating errors caused by human operation and improving detection consistency and repeatability.

[0061] 2) This application can not only extract local statistical features, but also automatically learn global cortical bone thickness spatial distribution patterns that are strongly correlated with the biomechanical mechanisms of the femur by utilizing a three-dimensional deep convolutional autoencoder that is self-supervised pre-trained based on anatomical perception occlusion and multi-scale reconstruction strategies, thereby achieving the extraction of global depth features. Simultaneously, this pre-training method makes the model's feature learning more targeted, accurately uncovering structural heterogeneity changes in key anatomical regions, thus extracting feature representations that are more discriminative and better suited to the quantitative analysis task of femoral cortical bone.

[0062] 3) Based on the three-dimensional thickness distribution map, this application can intuitively present the spatial location and distribution pattern of areas with abnormal cortical bone thickness, providing a clear basis for the analysis of femoral structural parameters. Furthermore, for the target region of interest and features that contribute most to the quantitative analysis results, it can clearly indicate the source of changes in key structural parameters such as thickness distribution and local mean, improving the interpretability and accuracy of the overall analysis process.

[0063] 4) This application can accurately capture subtle structural changes in cortical bone, identifying early spatial pattern changes in cortical bone thickness before differences appear in macroscopic skeletal parameters, providing a refined and quantitative parametric analysis basis for the three-dimensional structure of the femur. Simultaneously, this application integrates image depth features with relevant reference information to achieve individualized, high-precision comprehensive analysis of femoral cortical bone parameters, improving the accuracy and comprehensiveness of structural quantification results. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of a method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, according to an embodiment of this application.

[0066] Figure 2 This is a schematic diagram of the femoral cortical bone parameter quantification analysis system based on three-dimensional medical images, according to an embodiment of this application.

[0067] Figure 3 This is a schematic diagram of the cortical bone segmentation and thickness map generation process according to an embodiment of this application.

[0068] Figure 4 This is a schematic diagram illustrating the principle of a method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, according to an embodiment of this application.

[0069] Figure 5 This is a schematic diagram of the anatomical perception self-supervised pre-training process of a three-dimensional depth convolutional autoencoder according to an embodiment of this application.

[0070] Figure 6 This is a flowchart of deep learning model training and application according to an embodiment of this application. Detailed Implementation

[0071] To further illustrate the various embodiments, this application provides accompanying drawings, which are part of the disclosure of this application. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of this application.

[0072] According to embodiments of this application, a method and system for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images are proposed.

[0073] This application will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 and Figure 4 As shown, the method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to an embodiment of this application includes:

[0074] S1. Obtain the initial three-dimensional medical image of the femur to be analyzed and perform preprocessing to obtain the preprocessed three-dimensional medical image of the femur.

[0075] Specifically, data acquisition and preprocessing include: acquiring a three-dimensional quantitative CT image dataset of the femoral region of the target object; and preprocessing the three-dimensional CT images, including: image normalization, noise reduction, and standardizing the image spatial coordinates to a unified anatomical coordinate system, such as registration based on the femoral head center and the bone shaft axis.

[0076] S2. Segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional mask of the cortical bone; calculate the local thickness of the cortical bone based on the three-dimensional mask of the cortical bone to generate a three-dimensional thickness distribution map of the femoral cortical bone.

[0077] The segmented and preprocessed three-dimensional medical image of the femur generates a bone tissue mask for the femoral region. An algorithm combining region growing and level set evolution is used to separate the bone tissue mask, resulting in a three-dimensional mask of the cortical bone. Based on the three-dimensional mask of the cortical bone, the local thickness of the cortical bone is calculated, generating a three-dimensional thickness distribution map of the femoral cortical bone, including:

[0078] S21. Based on the 3D nnU-Net framework, residual connections and attention gate mechanism, a lightweight segmentation model is constructed and supervised training is performed using a femur 3D medical image dataset with cortical bone segmentation annotations.

[0079] Specifically, the lightweight segmentation model is based on the 3D nnU-Net framework and introduces residual connections and attention gate mechanisms in its encoder part to accurately segment the bone tissue mask of the entire femoral region (including the femoral head, neck, shaft, and intertrochanteric region). The loss function of the lightweight segmentation model adopts a combination of Dice loss and cross-entropy loss.

[0080] S22. Use the trained lightweight segmentation model to perform region segmentation on the preprocessed femoral 3D medical image dataset to obtain the bone tissue mask of the femoral region.

[0081] S23. Based on an algorithm combining region growing and level set evolution, the cortical bone region in the bone tissue mask is separated to obtain a three-dimensional cortical bone mask.

[0082] Specifically, the algorithm based on the combination of region growing and level set evolution separates the cortical bone region in the bone tissue mask to obtain a three-dimensional cortical bone mask, including:

[0083] S231. Based on the gray value distribution of computed tomography scan, an initial seed point is automatically selected within a preset range of the endosteal membrane in the bone tissue mask.

[0084] Initialize the seed region: In the bone tissue mask, an initial seed point is automatically selected near the endosteal region based on the CT gray value distribution (i.e., computed tomography gray value). Typically, the seed point is located in the transition region where the gray value is between the typical HU value of cortical bone (e.g., 300-1000 HU) and the typical HU value of medullary cavity (e.g., -100 to 100 HU).

[0085] S232. Using a three-dimensional region growth algorithm based on grayscale threshold, starting from the initial seed point, the region is gradually grown towards adjacent voxels to initially separate the cortical bone region and the cancellous bone region, thus obtaining a roughened cortical bone mask.

[0086] Region growing: A 3D region growing algorithm based on grayscale thresholding is used, starting from the seed point and gradually expanding to adjacent voxels. The expansion condition is that the voxel grayscale value is within the preset cortical bone grayscale range. This step initially separates the cortical bone region, but may result in rough boundaries or infiltration into cancellous bone due to uneven grayscale or partial volume effects.

[0087] Specifically, a grayscale threshold range matching the CT value range of cortical bone is set. Starting from the initial seed point, all voxels within its 26 neighborhoods are recursively checked. If the grayscale value of a neighboring voxel falls within this preset threshold range, it is included in the cortical bone region and used as a new growth point. This process is repeated until no new voxels are included, thereby generating a complete and connected cortical bone roughness mask.

[0088] S233. Using the roughened cortical bone mask as the initial contour for level set evolution, the level set algorithm based on edge and region information is used to generate a high-precision three-dimensional cortical bone mask.

[0089] Refined segmentation through level set evolution: The coarse mask obtained from region growing is used as the initial contour for level set evolution. An algorithm based on edge and region information is employed for evolution to accurately locate the inner and outer boundaries of the cortical bone. Specifically, the energy functional is defined as follows:

[0090] ;

[0091] in, Let be the energy functional used to control the evolution of the level set in the contour evolution process. The level set function is a scalar function defined on the image domain Ω, and its zero contour line is... =0 indicates the contour in the evolution, H( ) is used to distinguish the interior of the contour ( >0) and external ( Heaviside function for the region <0), The gradient of the Heaviside function used to represent the contour boundary is actually in The value at 0 is approximated by the Dirac function, and Ω represents the image domain, referring to the spatial extent of the entire 3D image. The processing function is an auxiliary function constructed based on the image gradient. This is a CT grayscale image (HU value). Let I be the gradient vector of the CT grayscale image I, representing the direction and intensity of grayscale changes in space. The edge detection function, defined as follows, is used to reduce energy at image edges and guide the contour to fit anatomical boundaries. μ, λ, and ν are weighting parameters that control the strength of the contour length smoothing term, edge attraction term, and region area constraint term, respectively.

[0092] In this embodiment, μ∈[0.01, 0.5], λ∈[0.5, 5.0], ν∈[-1.0, 1.0], preferably μ is 0.1, λ is 1.0, and ν is 0.1. It should be noted that these parameter values ​​are initial values ​​set based on conventional experimental experience in the field. In practical applications, μ, λ, and ν can be adaptively adjusted and iteratively optimized using hyperparameter optimization methods such as grid search, random search, or Bayesian optimization, according to the specific femoral CT image quality, cortical bone boundary clarity, and segmentation accuracy requirements, to obtain the optimal parameter combination suitable for a specific dataset. The above parameter optimization process is a conventional technique used by those skilled in the art and will not be elaborated upon here.

[0093] The gradient descent method is used for iterative solution, causing the contour to converge to the actual boundary between the intima and adventitia of the cortical bone, ultimately yielding a high-precision 3D mask of the cortical bone. The gradient descent iterative solution process is briefly described below:

[0094] Initialization: Convert the rough mask obtained from region growing into an initial level set function. The process involves setting iteration stopping conditions (e.g., maximum number of iterations or energy change threshold); iterative update: in each iteration, the energy functional gradient direction corresponding to the current level set function is calculated, and the level set function is updated along the gradient descent direction to evolve it in the direction of energy reduction. The level set function is periodically reinitialized to maintain its numerical stability; convergence judgment: iteration stops when the energy functional change is lower than the preset threshold or the maximum number of iterations is reached; contour extraction: the zero isosurface is extracted from the final level set function as the inner and outer boundaries of the cortical bone to generate a high-precision 3D mask; post-processing: morphological closing operation is performed on the initial 3D contour determined by the zero isosurface output from the level set evolution process. This operation aims to smooth the contour boundaries and fill any possible small voids, thereby finally obtaining a continuous, complete, and anatomically consistent high-precision 3D mask of cortical bone. This mask will be directly used for subsequent thickness map generation.

[0095] S24. Using the three-dimensional Euclidean distance transformation method, calculate the local thickness of the cortical bone at each voxel point in the three-dimensional mask of cortical bone, and generate a three-dimensional thickness distribution map of the femoral cortical bone based on the local thickness of the cortical bone of all voxels.

[0096] Specifically, the step of using the three-dimensional Euclidean distance transformation method to calculate the local thickness of the cortical bone at each voxel point in the three-dimensional cortical bone mask, and generating a three-dimensional thickness distribution map of the femoral cortical bone based on the local cortical bone thickness of all voxels, includes:

[0097] S241. Using three-dimensional morphological erosion operations, the boundary of the bone marrow cavity in the three-dimensional mask of cortical bone is extracted to obtain the voxel set of the boundary of the bone marrow cavity.

[0098] S242. Extract the outer boundary of the bone in the three-dimensional mask of cortical bone through three-dimensional morphological dilatation operation to obtain the voxel set of the outer boundary of the bone;

[0099] S243. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the boundary of the medullary cavity, and select the minimum distance as the Euclidean distance from that voxel to the nearest surface of the medullary cavity.

[0100] S244. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the outer boundary of bone, and select the minimum distance as the Euclidean distance from that voxel to the nearest outer surface of bone.

[0101] S245. Calculate the local cortical bone thickness at each voxel point based on the Euclidean distance from the voxel to the nearest inner surface of the medullary cavity and the Euclidean distance from the voxel to the nearest outer surface of the bone.

[0102] S246. Map the local thickness values ​​of cortical bone at all voxel points back to their corresponding positions in the original three-dimensional image space to form a three-dimensional thickness distribution map of the femoral cortical bone.

[0103] Specifically, based on a three-dimensional cortical bone mask, the local cortical bone thickness at each voxel is calculated. The three-dimensional Euclidean distance transformation method is used: the local cortical bone thickness value at voxel v, denoted as T(v), in mm, is defined as follows:

[0104] ;

[0105] Where, d in (v) and d out (v) represents the Euclidean distance from the voxel to the nearest inner and outer surfaces of the medullary cavity (i.e., in units of the same image spatial resolution, typically in mm). The thickness values ​​of all voxels are mapped back to three-dimensional space to form a three-dimensional thickness distribution map of the femoral cortex (the voxel values ​​are the thickness values).

[0106] S3. Based on the feature extraction model and the target region of interest, extract the global depth feature vector and local statistical feature vector that represent the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map, and combine the feature vectors to obtain the comprehensive feature vector.

[0107] Specifically, the step of extracting global depth feature vectors and local statistical feature vectors representing the three-dimensional spatial distribution pattern of cortical bone from the three-dimensional thickness distribution map based on the feature extraction model and the target region of interest, and then concatenating these feature vectors to obtain a comprehensive feature vector includes:

[0108] S31. Based on the three-dimensional deep convolutional autoencoder architecture, a feature extraction model is constructed, and self-supervised pre-training is performed using an anatomically perceptual occlusion and multi-scale reconstruction strategy.

[0109] Specifically, the feature extraction model employs a 3D deep convolutional autoencoder based on anatomically perceptual occlusion and multi-scale reconstruction strategies. Its encoder part uses multiple downsampling convolutional layers to compress the thickness map and abstract it into a low-dimensional, dense global depth feature vector. This vector not only contains the overall distribution pattern of cortical bone thickness in 3D space but also strengthens the encoding of the spatial contextual relationships of key biomechanical regions, thereby extracting deep features with stronger discriminative power for structural categories.

[0110] The self-supervised pre-training using an anatomically-aware occlusion and multi-scale reconstruction strategy includes: constructing a multi-level region-of-interest (ROI) probability map corresponding to the femoral anatomy, with different initial occlusion probabilities assigned to regions at different levels within the multi-level ROI probability map; generatively sampling the input three-dimensional thickness distribution map of the femoral cortex based on the multi-level ROI probability map to obtain an initial occlusion mask; optimizing the initial occlusion mask using a three-dimensional conditional random field model to generate spatially coherent occlusion blocks, and using these spatially coherent occlusion blocks to perform a voxel-by-voxel occlusion operation on the original three-dimensional thickness distribution map to obtain an occluded thickness map; and training a three-dimensional deep convolutional autoencoder for reconstruction using the occluded thickness map. The feature extraction model employs a reconstruction loss (such as mean squared error) to learn effective feature representations, and the reconstruction loss function includes pixel-level reconstruction loss and structural consistency loss based on multi-scale feature matching.

[0111] Among them, the 3D Conditional Random Field (CRF) model is a probabilistic graphical model for structured prediction. Its core function is to jointly optimize the masking labels by modeling the spatial consistency and thickness similarity between adjacent voxels, given an initial masking mask, thereby generating spatially coherent 3D masking blocks and avoiding isolated, scattered masking points. In this method, the energy function of the model consists of univariate and binary potential energy: univariate potential energy encourages the final masking label of a voxel to be consistent with its initial sampling label; binary potential energy penalizes adjacent voxels with similar thickness values ​​being assigned different masking labels, thus promoting the formation of a continuous blocky structure in the masking region. By minimizing this energy function using the graph cut algorithm, the optimized spatially coherent masking mask can be obtained. S32. Using a pre-trained feature extraction model, extract the global depth feature vector containing the overall distribution pattern of cortical bone thickness in 3D space from the 3D thickness distribution map;

[0112] S33. Based on the preset target region of interest, extract the statistical thickness features within the target region of interest in the 3D thickness distribution map to form a local statistical feature vector;

[0113] Specifically, based on anatomical knowledge, multiple regions of interest (ROIs) are defined from the three-dimensional thickness distribution map of cortical bone, such as the upper, lower, anterior, and posterior parts of the femoral neck, the medial and lateral parts of the intertrochanteric region, and the proximal femoral shaft. Statistical thickness features (such as mean thickness, standard deviation of thickness, skewness, kurtosis, and thickness distribution histogram quantiles) are extracted from each ROI to form a local statistical feature vector.

[0114] S34. Concatenate the global deep feature vector with the local statistical feature vector to obtain the comprehensive feature vector.

[0115] S4. Using the pre-trained deep learning model, the comprehensive feature vector is quantitatively analyzed to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0116] The step of using a pre-trained deep learning model to perform quantitative analysis on the comprehensive feature vector and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, thereby achieving parameterized quantification of femoral cortical bone thickness characteristics, includes:

[0117] S41. Construct a deep learning model based on a multi-layer fully connected neural network and perform supervised training using a dataset containing comprehensive features and quantization results.

[0118] S42. Input the comprehensive feature vector into the trained deep learning model to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0119] Specifically, the comprehensive feature vector is input into a pre-trained deep learning model, which is a multi-layer fully connected neural network. Its last layer uses the Sigmoid activation function to output the quantitative analysis result of the target object. This result is a probability value between 0 and 1, which is used to characterize the confidence level of the current spatial distribution pattern of femoral cortical bone thickness belonging to a preset category.

[0120] Deep learning models employ weighted cross-entropy loss to address the imbalance problem caused by scarce samples. Data augmentation techniques (such as rotation, scaling, and elastic deformation) are used during training to improve the model's generalization ability.

[0121] The loss function of the deep learning model is the binary classification focus loss function, and the expression of the binary classification focus loss function is:

[0122] ;

[0123] In the formula, L focal Let y represent the binary classification focus loss function, where N represents the number of samples, and y represents the number of samples. i p represents the true class label of the i-th sample. i Let represent the probability that the i-th sample belongs to the positive class, α represent the class weight coefficient of the positive class, and (1-α) represent the class weight coefficient of the negative class. This indicates the focus parameter.

[0124] like Figure 2 As shown, according to another embodiment of this application, a femoral cortical bone parameter quantification analysis system based on three-dimensional medical images is also provided, comprising:

[0125] The femoral image preprocessing module is used to acquire the initial three-dimensional medical image of the femur to be analyzed and to preprocess it to obtain the preprocessed three-dimensional medical image of the femur.

[0126] The cortical bone thickness calculation module is used to segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional cortical bone mask; based on the three-dimensional cortical bone mask, the local thickness of the cortical bone is calculated to generate a three-dimensional thickness distribution map of the femoral cortical bone.

[0127] The femoral feature extraction and fusion module is used to extract global depth feature vectors and local statistical feature vectors that characterize the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map based on the feature extraction model and the target region of interest. The feature vectors are then combined to obtain a comprehensive feature vector. The feature extraction model is self-supervised pre-trained based on the anatomical perception occlusion and multi-scale reconstruction strategy.

[0128] The quantitative analysis module is used to perform quantitative analysis on the comprehensive feature vector using a pre-trained deep learning model, and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

[0129] The cortical bone thickness calculation module, after segmenting the preprocessed 3D medical image of the femur, generates a bone tissue mask for the femur region and separates the bone tissue mask using an algorithm combining region growing and level set evolution to obtain a 3D cortical bone mask. The calculation of local cortical bone thickness based on the 3D cortical bone mask to generate a 3D thickness distribution map of the femoral cortex includes: constructing a lightweight segmentation model based on the 3D nnU-Net framework, residual connections, and attention gating mechanism, and performing supervised training using a 3D medical image dataset of the femur with cortical bone segmentation annotations; using the trained lightweight segmentation model to segment the preprocessed 3D medical image dataset of the femur to obtain a bone tissue mask for the femur region; separating the cortical bone region from the bone tissue mask using an algorithm combining region growing and level set evolution to obtain a 3D cortical bone mask; and calculating the local cortical bone thickness at each voxel point in the 3D cortical bone mask using the 3D Euclidean distance transformation method, and generating a 3D thickness distribution map of the femoral cortical bone based on the local cortical bone thickness of all voxels.

[0130] The cortical bone thickness calculation module, when separating the cortical bone region from the bone tissue mask and obtaining the three-dimensional cortical bone mask using an algorithm combining region growth and level set evolution, includes: automatically selecting an initial seed point within a preset range of the endostea in the bone tissue mask based on the gray value distribution of computed tomography scans; using a three-dimensional region growth algorithm based on gray thresholds, starting from the initial seed point, gradually growing towards adjacent voxels to initially separate the cortical bone region and obtain a cortical bone roughness mask; using the cortical bone roughness mask as the initial contour for level set evolution, and using a level set algorithm based on edge and region information to evolve and generate the three-dimensional cortical bone mask.

[0131] Specifically, the energy functional expression for level set evolution is:

[0132]

[0133] In the formula, Let be the energy functional used to control the evolution of the level set in the contour evolution process. H( is the level set function) ) is the Heaviside function used to distinguish between the inner and outer regions of a contour. Let Ω be the gradient of the Heaviside function used to represent the contour boundary, and Ω be the image domain. Let I be the gradient vector of the CT grayscale image I, representing the direction and intensity of grayscale changes in space. Let μ be an edge detection function used to reduce energy at image edges and guide the contour to fit the anatomical boundary, where μ, λ, and ν are weight parameters that control the strength of the contour length smoothing term, the edge attraction term, and the region area constraint term, respectively.

[0134] Specifically, the cortical bone thickness calculation module, when calculating the local cortical bone thickness at each voxel point in the three-dimensional cortical bone mask using the three-dimensional Euclidean distance transformation method, and generating a three-dimensional thickness distribution map of the femoral cortical bone based on the local cortical bone thickness of all voxels, includes: extracting the intramedullary boundary of the cortical bone three-dimensional mask using a three-dimensional morphological erosion operation to obtain the intramedullary boundary voxel set; extracting the outer bone boundary of the cortical bone three-dimensional mask using a three-dimensional morphological dilation operation to obtain the outer bone boundary voxel set; and calculating the three-dimensional Euclidean distance from each voxel in the cortical bone three-dimensional mask to all voxels in the intramedullary boundary voxel set. The distances are calculated, and the minimum distance is selected as the Euclidean distance from the voxel to the nearest medullary cavity surface. The three-dimensional Euclidean distances from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the outer boundary of bone are calculated, and the minimum distance is selected as the Euclidean distance from the voxel to the nearest outer bone surface. Based on the Euclidean distances from the voxel to the nearest medullary cavity surface and from the voxel to the nearest outer bone surface, the local cortical bone thickness value at each voxel point is calculated. The local cortical bone thickness values ​​at all voxel points are mapped back to the corresponding positions in the original three-dimensional image space to form a three-dimensional thickness distribution map of the femoral cortical bone.

[0135] The femoral feature extraction and fusion module, based on a feature extraction model and a target region of interest, extracts global depth feature vectors and local statistical feature vectors representing the three-dimensional spatial distribution pattern of cortical bone in a three-dimensional thickness distribution map, and concatenates these feature vectors to obtain a comprehensive feature vector. This process includes: constructing a feature extraction model based on a three-dimensional deep convolutional autoencoder architecture and performing self-supervised pre-training using an anatomically perceptual occlusion and multi-scale reconstruction strategy; using the pre-trained feature extraction model to extract global depth feature vectors from the three-dimensional thickness distribution map that represent the overall distribution pattern of cortical bone thickness in three-dimensional space; extracting statistical thickness features from the target region of interest in the three-dimensional thickness distribution map based on a preset target region of interest to form a local statistical feature vector; and concatenating the global depth feature vector with the local statistical feature vector to obtain the comprehensive feature vector.

[0136] Specifically, the femoral feature extraction and fusion module, when performing self-supervised pre-training using an anatomically-aware occlusion and multi-scale reconstruction strategy, includes: constructing a multi-level region of interest probability map corresponding to the femoral anatomical structure, where different levels of regions in the multi-level region of interest probability map are assigned different initial occlusion probabilities; generatively sampling the input three-dimensional thickness distribution map of the femoral cortex based on the multi-level region of interest probability map to obtain an initial occlusion mask; optimizing the initial occlusion mask using a three-dimensional conditional random field model to generate spatially coherent occlusion blocks; and training a three-dimensional deep convolutional autoencoder for reconstruction using the occluded thickness map, wherein the reconstruction loss function includes pixel-level reconstruction loss and structural consistency loss based on multi-scale feature matching.

[0137] The quantitative analysis module, which utilizes a pre-trained deep learning model to perform quantitative analysis on the comprehensive feature vector and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, thereby achieving parameterized quantification of femoral cortical bone thickness features, includes: constructing a deep learning model based on a multi-layer fully connected neural network and performing supervised training using a dataset containing comprehensive features and quantification results; inputting the comprehensive feature vector into the trained deep learning model to generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, thereby achieving parameterized quantification of femoral cortical bone thickness features.

[0138] Specifically,

[0139] The loss function of the deep learning model is the binary classification focus loss function, and the expression of the binary classification focus loss function is:

[0140] ;

[0141] In the formula, L focal Let y represent the binary classification focus loss function, where N represents the number of samples, and y represents the number of samples. i p represents the true class label of the i-th sample. i Let represent the probability that the i-th sample belongs to the positive class, α represent the class weight coefficient of the positive class, and (1-α) represent the class weight coefficient of the negative class. This indicates the focus parameter.

[0142] To facilitate understanding of the above technical solutions of this application, the following further explains the above technical solutions from the perspective of architecture and principles, as follows:

[0143] This application aims to overcome the shortcomings of existing technologies and provide a highly automated, accurate, and comprehensive method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images. This method comprehensively includes the following core steps: image acquisition and preprocessing, thickness map generation, depth feature extraction, and feature quantification analysis. The specific implementation details of each step are as follows:

[0144] I. Data Preparation and Preprocessing Scheme

[0145] 1.1 Data Sources and Acquisition Strategies:

[0146] This method is adaptable to different data sources and can use one or more of the following strategies to construct training and testing datasets:

[0147] Strategy A (Multi-center Collaboration): Collaborate with 2-3 hospitals possessing imaging research capabilities to retrospectively collect hip CT scan data. The minimum feasible data volume is recommended to include at least 500 well-defined cases to meet the basic needs for initial training and validation of the deep learning model.

[0148] Strategy B (Utilization of Public Datasets and Incremental Learning): Prioritize the use of existing publicly available medical image datasets for model pre-training. Subsequently, fine-tune or adapt the pre-trained model using a smaller local dataset collected by the institution (e.g., 100-200 samples with fine annotations) to adapt it to the target device and image features.

[0149] Strategy C (Synthetic Data Augmentation): When real data is limited, synthetic data generation techniques based on generative adversarial networks (such as 3D-StyleGAN) can be used. First, a 3D medical image synthesizer is trained using a small amount of high-quality labeled data (e.g., 50-100 cases) to generate synthetic femoral CT images with diversity, realistic anatomical structures, and cortical bone changes, along with corresponding labels, to expand the training set.

[0150] 1.2 Data Labeling Process (High-Efficiency Semi-Automatic):

[0151] To address the issue of high annotation costs, this application employs an iterative annotation optimization process based on prediction uncertainty:

[0152] Initial annotation: A physician uses a first deep learning model (i.e., a segmentation model, the initial version of which can be pre-trained on a public dataset) to automatically segment a small number of images (e.g., 50 cases). Then, manual correction is performed primarily on automatically segmented error regions (i.e., the cortical-endostrate boundary), forming a high-quality seed annotation set. Model iteration: The segmentation model is trained or fine-tuned using the seed annotation set. Active learning loop: The trained model is used to perform quantitative analysis on unlabeled data, and the uncertainty of the model analysis is calculated (e.g., calculating voxel-level analysis entropy using Monte Carlo Dropout). The samples with the most model uncertainty (e.g., 20-30 cases) are selected and given to the physician for targeted annotation.

[0153] The specific method for calculating voxel-level analysis entropy using Monte Carlo Dropout is as follows:

[0154] Enable Dropout: During the first analysis phase of the trained deep learning model, keep the Dropout layer in the model enabled (i.e., do not disable Dropout during testing) to introduce randomness in multiple forward propagations and simulate the uncertainty of the model.

[0155] Multiple sampling analysis: Perform T independent forward propagations on the same input image (T=30 is recommended). Due to the randomness of Dropout, different output probability maps P will be obtained each time. t (v), where t=1,2,...,T, and v represents the voxel position.

[0156] Calculate the analysis entropy: For each voxel v, calculate the mean of its category probability (e.g., cortical bone category) across T analyses.

[0157] ;

[0158] Then, the analytical entropy H(v) is calculated based on this mean, which is used to quantify the uncertainty of the model at this voxel:

[0159] ;

[0160] Uncertainty measurement: The voxel-level analysis entropy of each sample is averaged across all voxels to obtain the overall uncertainty score for that sample. Samples with the highest overall uncertainty are selected and given priority annotation by physicians. The newly annotated data is added to the training set, and the above steps are repeated. Typically, after 3-5 iterations, a high-performance segmentation model can be obtained with relatively low manual annotation costs (i.e., a total of approximately 150-200 finely annotated cases).

[0161] 1.3 Data Preprocessing (Standardized Process):

[0162] Regardless of the data size, the following preprocessing steps are consistent and necessary:

[0163] CT value standardization: If the scan includes a QCT phantom (Quantitative Computed Tomography phantom), the HU values ​​are converted to bone mineral density values ​​(mg / cm³) according to standard procedures. If no phantom is available, a Z-score normalization method based on image statistics is used, with the following formula:

[0164] ;

[0165] Among them, I norm (v) is the normalized value of voxel v, I raw (v) represents the original HU value of voxel v, μ muscle and These represent the mean and standard deviation of HU values ​​for muscle tissue at the same or adjacent levels. Muscle tissue ROIs can be automatically located using a pre-trained soft tissue segmentation model, or sampled at fixed positions within a standard registration space based on anatomical atlases, ensuring cross-sample consistency. This method effectively corrects for differences between different scanners and protocols.

[0166] Spatial standardization (i.e., key steps):

[0167] Landmark-based registration, even with small datasets, needs to be automated. A lightweight variant of 3DU-Net can be used for keypoint detection, with the output layer being a keypoint heatmap and the loss function being mean squared error. Training should use at least 50 examples of labeled keypoints. This network can quickly analyze keypoint coordinates in new images.

[0168] Femoral head center P head By fitting the point cloud of the femoral head articular surface into a sphere, the center of the sphere is P. head The femoral neck axis can be calculated on the point cloud of the femoral neck region using principal component analysis (PCA), and the direction of its first principal component is the direction of the neck axis.

[0169] Based on the key points of the analysis, similarity transformations (such as rotation, translation, and scaling) are used to register all images to a common template. This step does not require large-scale data; its purpose is to ensure that all femurs have similar spatial orientation and scale, greatly reducing the difficulty of subsequent model learning.

[0170] ROI extraction: In the registered space, with the femoral head center as a reference, a cubic region containing key anatomical structures of the proximal femur (such as head, neck, intertrochanteric region, and part of the diaphysis) is fixedly cut out, for example, 160×160×200 voxels.

[0171] II. Core Model Architecture and Training Details

[0172] 2.1 First Deep Learning Model (Segmentation Model, Lightweight Design):

[0173] Given that data may be limited, model design should avoid overparameterization to prevent overfitting.

[0174] Network architecture: Figure 3 The detailed process of cortical bone segmentation and thickness map generation is demonstrated. The first deep learning model (segmentation model) is based on the 3D nnU-Net framework, and residual connections and attention gate mechanisms are introduced into the encoder to enhance feature extraction capabilities. nnU-Net can automatically configure hyperparameters such as network depth and number of convolutional kernels according to the given dataset, making it a robust choice for situations with limited data. The introduced improved mechanisms help to more accurately locate the cortical bone boundaries.

[0175] 2.2 Second Deep Learning Model (Feature Extraction Model, Pre-training and Fine-tuning Strategies):

[0176] The second deep learning model uses, as follows: Figure 5The diagram shows a 3D deep convolutional autoencoder architecture. It is recommended that the encoder output dimension of the second deep learning model (i.e., the length of the global deep feature vector) be set to 512 dimensions, as a compressed representation to match the input dimension of the subsequent classification head. This can be achieved even with limited data through transfer learning.

[0177] Model architecture: The second deep learning model adopts a 3D convolutional autoencoder architecture.

[0178] Key Strategy—Self-Supervised Pre-training Based on Anatomical Perception Covering and Multi-Scale Reconstruction

[0179] The core improvement of this application lies in the design of a hierarchical anatomically-aware generative occlusion and multi-scale structural-aware reconstruction pre-training task. This task aims to guide the feature representation learned by the 3D deep convolutional autoencoder to more accurately reflect the spatial distribution pattern of cortical bone thickness parameters, rather than general image texture.

[0180] The self-supervised pre-training process is as follows:

[0181] 1) Construction of prior anatomical atlas: Before the start of pre-training, a multi-level anatomical importance probability atlas M is constructed offline based on literature and clinical knowledge of femoral biomechanics. importance The atlas is aligned with the standard registered thickness map space, and each voxel v is assigned an importance weight w(v) and an initial target occlusion probability p. mask (v), the calculation formula is as follows:

[0182] ;

[0183] ;

[0184] An example of importance hierarchy is shown below:

[0185] Level 1 (i.e., the biomechanical critical region): includes the posterosuperior part of the femoral neck and the lateral wall of the intertrochanteric region, such as p mask (v)∈[0.1,0.2], to ensure that the model input retains as much of its anatomical context as possible.

[0186] Level 2 (i.e., the important transitional area): includes the anterior and inferior portions of the femoral neck and the proximal weight-bearing area of ​​the femoral shaft. Assign a moderate occlusion probability, e.g., p. mask (v)∈[0.3,0.5].

[0187] Level 3 (i.e., relatively stable region): includes the non-major weight-bearing area in the center of the femoral head and the distal portion of the femoral shaft. A higher occlusion probability is assigned, e.g., p. mask (v)∈[0.6,0.8], which encourages the model to reconstruct these parts by reasoning about the surrounding regions in order to learn the global distribution pattern.

[0188] 2) Spatially coherent generative masking: For each input thickness map T, a masking mask is generated according to the following process:

[0189] Probability sampling: For each voxel v, with probability p mask (v) Mark it as to be covered.

[0190] 3D CRF Optimization (i.e., Achieving Spatial Coherence): To avoid generating overly fragmented occlusion patterns that disrupt 3D structure perception, a lightweight 3D conditional random field model is used to optimize the initial binary occlusion mask. The energy function E(X) of the CRF is defined as:

[0191] ;

[0192] in, The model's total energy function is represented by X; X represents the masking label for each voxel (0 indicates preservation, 1 indicates masking); x i The masking label represents the i-th voxel, with a value of 0 (retained) or 1 (masked); x j The masking label represents the j-th voxel, with a value of 0 (retained) or 1 (masked). Represents unary potential energy, reflecting the label x of voxel i. i The cost of this is to encourage voxels to retain their initial sampling labels; It represents the binary potential energy, reflecting the smoothness cost of adjacent voxels i and j taking different labels.

[0193] binary potential energy Defined as:

[0194] ;

[0195] Among them, w s The spatial similarity weights control the strength of the influence of location distance on potential energy; w c This represents the similarity weight of thickness values, controlling the influence of strength differences on potential energy; w s and w c This represents the weighting coefficient, usually denoted as w. s =1.0, w c =0.5, to further emphasize spatial coherence; loc i Represents the spatial coordinates of voxel i in a 3D image; loc j This represents the spatial coordinates of voxel j in the 3D image; This represents the spatial distance scale parameter, typically taken as 3 to 5 voxels. The thickness variation scale parameter is usually taken as 0.5 to 1 times the standard deviation of the thickness distribution; This represents the cortical bone thickness value of voxel i in the three-dimensional thickness distribution map of cortical bone; represents the cortical bone thickness value of voxel j in the three-dimensional cortical bone thickness distribution map; ||.|| represents the Euclidean norm, used to calculate the difference between coordinates or values; The tag compatibility function takes the following values:

[0196] .

[0197] This binary term encourages spatially adjacent voxels with similar thickness values ​​(i.e., likely belonging to the same anatomical substructure) to receive the same masking label, thereby generating larger and more coherent 3D masking blocks. The final optimized masking mask is obtained by minimizing the energy function E(X) using a graph cut algorithm.

[0198] The graph cut algorithm is briefly described below:

[0199] Graph construction: The 3D thickness distribution map is modeled as a weighted undirected graph G=(V,E), where: V represents the set of all voxels, and each voxel corresponds to a node in the graph; E represents the set of edges between adjacent voxels.

[0200] Edge weight setting: The weight of each edge (i,j)∈E is set to the binary potential energy value. This reflects the different penalty costs associated with adjacent voxel labels. The graph also introduces two virtual nodes: a source node (where the corresponding label is preserved) and a sink node (where the corresponding label is masked), each connected to a voxel node, with edge weights equal to unary potential energy. This reflects the cost of a voxel taking a certain label.

[0201] Minimum cut solution: The minimum cut of the graph is solved using the Boykov-Kolmogorov algorithm, which divides the nodes into two subsets: nodes connected to the source (i.e., those with labels kept) and nodes connected to the sink (i.e., those with labels covered).

[0202] Output result: The node partitioning result corresponding to the minimum cut is the optimized binary occlusion mask.

[0203] Dynamic adjustment: Based on the reconstruction difficulty of the model on the validation set, the global average occlusion ratio can be dynamically fine-tuned within a training cycle, following a course learning strategy from easy to difficult.

[0204] 3) Multi-scale structure-aware reconstruction task: The masked thickness map T⊙Mask is input into the autoencoder, and the reconstructed map T' is output. The training objective is not only to recover the thickness value, but also to ensure the anatomical rationality of the reconstructed structure. ⊙ represents element-wise multiplication.

[0205] Pixel-level reconstruction loss: Focuses on the reconstruction accuracy of the occluded area, pixel-level reconstruction loss L pixel The formula is as follows:

[0206] ;

[0207] in, T(v) represents the value of the reconstructed thickness map at voxel v; T(v) represents the value of the original thickness map at voxel v; N masked The value represents the total number of voxels that are covered; Mask represents the set of covered regions.

[0208] Multi-scale feature matching loss: A pre-trained, parameter-frozen 3D convolutional neural network on a medical image dataset is introduced as a structure perceptron F(·). L intermediate feature maps at different scales (resolutions) are computed during the autoencoder decoding process. } and the corresponding scale feature map extracted by the structure perceptron F from the original complete thickness map T The difference between}. Loss is defined as:

[0209] ;

[0210] Where L represents the total number of scales selected; This represents the feature map of the decoder at the l-th scale; This represents the l-th scale feature map extracted by the pre-trained structure perceptron from the original complete thickness map; This represents the square of the L2 norm (i.e., the square of the Euclidean distance). This represents the weight coefficients of the l-th layer, which typically decrease as the scale coarsens (i.e., the resolution decreases). For example, let be... .

[0211] Total loss: ;

[0212] in, Represents the total loss of self-supervised pre-training, which refers to the overall loss function used to optimize the 3D deep convolutional autoencoder during the self-supervised pre-training stage; This represents pixel-level reconstruction loss and measures the reconstruction accuracy by evaluating the thickness of the occluded region. This represents the multi-scale feature matching loss; This represents the weighting coefficient that balances the two losses. It is usually set to 1.0, but can also be adjusted in the range of [0.5, 2.0] based on the reconstruction quality of the validation set.

[0213] 4) Supervised Fine-Tuning: After completing the above self-supervised pre-training, two strategies can be used for supervised fine-tuning:

[0214] Strategy 1 (i.e., freeze the encoder): Freeze all parameters of encoder E and train only the subsequent multilayer perceptron classifier head C. This strategy is suitable for scenarios with limited labeled data (e.g., less than 500 examples) and can prevent overfitting.

[0215] Strategy 2 (i.e., partial fine-tuning): Unfreeze the parameters of the last 1-2 convolutional layers of encoder E and fine-tune them together with the classification head C. This strategy is suitable for scenarios with relatively sufficient labeled data (e.g., more than 500 examples) and can further improve feature adaptability.

[0216] Regardless of the strategy employed, supervised training is performed using a dataset with sample labels to adapt the high-quality general features obtained from pre-training to the task of quantitative analysis of femoral cortical bone structure.

[0217] 2.3 Calculation of cortical bone thickness (stable algorithm):

[0218] This step, based on the segmented 3D mask of cortical bone, uses the 3D Euclidean distance transformation method to calculate the local thickness at each cortical bone voxel. The specific implementation steps are as follows:

[0219] Step 1: Define and extract the boundary

[0220] Bone marrow cavity boundary extraction: A three-dimensional morphological erosion operation is performed on the cortical bone mask, and then the eroded mask is subtracted from the original mask to obtain the voxel set of the bone marrow cavity boundary (i.e. the inner surface of the cortical bone adjacent to the bone marrow cavity).

[0221] Extraction of the outer bone boundary: Perform a three-dimensional morphological expansion operation on the cortical bone mask, and then subtract the original mask from the expanded mask to obtain the voxel set of the outer bone boundary (i.e. the outer surface of the cortical bone adjacent to the soft tissue).

[0222] Step 2, Calculate the distance field: for each voxel v within the cortical bone mask:

[0223] Calculate the three-dimensional Euclidean distance from it to all voxels in the set of intramedullary boundaries, and take the minimum value among them, denoted as d. in (v), which represents the distance of the voxel to the nearest intramedullary surface of the bone marrow cavity.

[0224] Calculate the three-dimensional Euclidean distance from it to all voxels in the set of bones at the outer boundary, and take the minimum value among them, denoted as d. out (v), which represents the distance of the voxel to the nearest bone surface.

[0225] Step 3, Calculate the local thickness: The local cortical bone thickness T(v) at voxel v is defined as follows:

[0226] ;

[0227] The physical meaning of this is that the thickness of the cortical bone at this point is approximately the sum of its vertical distances to the inner and outer surfaces.

[0228] Step 4: Generate thickness distribution map: Traverse all voxels within the cortical bone mask, calculate their T(v) values, and map these thickness values ​​back to their corresponding positions in the original three-dimensional image space to form a three-dimensional thickness distribution map of the femoral cortex (a three-dimensional array with the same size as the original image but with voxel values ​​as thickness values).

[0229] 2.4 Training and Optimization Strategies for Various Deep Learning Models

[0230] 1) Regularization: The following strategies are widely used in the training of all deep learning models to prevent overfitting:

[0231] In the supervised training phase of all deep learning models in this application, including the segmentation training of the first model, the supervised fine-tuning of the second model, and the classification training of the third model, several mature regularization techniques are systematically applied to improve the generalization ability of the model under limited data and prevent overfitting.

[0232] Dropout: Use Dropout after the fully connected layer, with a drop rate of 0.3-0.5.

[0233] Weight decay (i.e., L2 regularization): The coefficient is set to 1×10 -4 .

[0234] Early stop: Monitor the loss on the validation set, and stop training if there is no improvement after 10 consecutive epochs.

[0235] 2) Detailed explanation of the loss function:

[0236] Output layer design: The last layer of the network uses the Sigmoid activation function to output the probability map of each voxel belonging to the cortical bone, with a value range of [0,1].

[0237] Loss function: Hybrid loss is used. ;

[0238] in, This represents the total loss of the task splitting; This represents the weighted cross-entropy loss, used to measure the difference between the quantitative analysis data for each voxel classified as cortical bone and the true label, and to alleviate the class imbalance problem by assigning higher weights to cortical bone voxels. Dice loss is used to measure the overlap between the segmentation mask and the real mask. It is particularly sensitive to the alignment of the segmentation boundaries and helps to improve the contour accuracy of the segmentation. This represents the weighting coefficient, with an empirical value of 0.7. It can also be adjusted within the range of [0.5, 0.9] based on the segmentation performance.

[0239] 3) The third deep learning model (i.e., the deep learning model):

[0240] Output layer design: The last layer of the network uses the Sigmoid activation function (for binary classification probability analysis) to obtain the quantitative analysis results.

[0241] Application and calculation of the Sigmoid activation function in this application: In the third deep learning model of this application, the Sigmoid activation function is used to convert the linear output of the last layer of the network into a probability value between 0 and 1. Let the output of the last layer of the model be z, then the probability p is calculated as follows:

[0242] ;

[0243] Where e is the natural constant. In this application, the output p-value can be directly used as a quantitative analysis value.

[0244] Loss function: Focal Loss is used to address the data imbalance problem of predefined structural pattern category labels. Inaccurately segmented regions (such as endothelial images and endothelial boundary images) are corrected and used as segmentation labels. This loss function is applied to the activation function to output the corrected structural category labels.

[0245] Application and calculation of the Focal Loss function (i.e., the binary classification focal loss function) in this application: L focal This is the formula for calculating the loss function (i.e., binary classification form) used to optimize the model:

[0246] ;

[0247] Where N represents the number of samples, y i Let y represent the true class label of the i-th sample. i ∈{0,1}, p i p represents the probability of predicting that the i-th sample belongs to the positive class. i ∈[0,1], α represents the class weight coefficient of the positive class, and (1-α) represents the class weight coefficient of the negative class. This represents the focusing parameter (usually set to 2.0), used to reduce the loss contribution of easily classified samples. It is adjusted... and During training, the model can learn discriminative features more effectively from a limited range of structural pattern categories.

[0248] 4) Optimizer and learning rate scheduling strategy:

[0249] Optimizer settings: The first, second, and third deep learning models all use the AdamW optimizer (i.e., an adaptive moment estimator with weight decay) independently for parameter updates. For the second deep learning model, different optimizer instances are used for its self-supervised pre-training phase and supervised fine-tuning phase.

[0250] Learning rate initialization: The initial learning rate for all models is uniformly set to l. r =3×10 -4 .

[0251] Cosine annealing scheduler implementation: During the training of each model, a cosine annealing learning rate scheduler is used. Within each training cycle, the learning rate smoothly decays from its initial value to its minimum value according to the following cosine function:

[0252] ;

[0253] in, Let η represent the learning rate in the t-th training epoch. max =3×10 -4 η represents the initial learning rate (i.e., the maximum value). min =3×10 -5 T represents the minimum learning rate (set to 1 / 10 of the initial value). max This represents the total number of training cycles, typically set to 100-200 epochs, where t represents the current epoch number, ranging from 0 to T. max -1.

[0254] The training strategies (e.g., whether to freeze or fine-tune) for each deep learning model in this application can be flexibly adjusted according to the actual data scale, computing resources and performance requirements.

[0255] III. System Integration and Deployment Examples (e.g.) Figure 6 (As shown)

[0256] 3.1 Minimum Feasible System Flow:

[0257] The user uploads a DICOM file of a proximal femoral CT scan. The system automatically performs preprocessing (i.e., intensity normalization, keypoint detection and registration, and target interest (ROI) cropping). A first deep learning model (segmentation model) is run to obtain a cortical bone mask. In the registered space, local statistical features are automatically calculated from 12 predefined ROIs (for each ROI, the extracted statistical thickness features include mean thickness, thickness standard deviation, skewness, kurtosis, and the 5th, 25th, 50th, 75th, and 95th percentiles. These statistics constitute the local statistical feature vector). The cortical bone thickness map is input into a pre-trained second deep learning model (feature extraction model) to extract a 512-dimensional global depth feature vector. The global and local features are concatenated and input into a trained third deep learning model (i.e., a deep learning model, a classifier based on a multilayer perceptron (MLP)). The output is a continuous numerical value (range 0-1) generated by the sigmoid activation function, which characterizes the similarity between the cortical bone thickness distribution pattern of the input sample and a specific pattern category in a predefined reference atlas library.

[0258] The definition method for ROI is as follows:

[0259] In the registered standard space, 12 key ROIs were defined based on the anatomical atlas for local statistical feature extraction:

[0260] (1) Femoral neck (4 zones): The femoral neck is divided into 4 quadrants in the anterior-posterior direction and the superior-inferior direction, namely the anterior superior (AS), anterior inferior (AI), posterosuperior (PS), and posteroinferior (PI). The partition plane is the anatomical axis of the femoral neck and the plane perpendicular to it.

[0261] (2) Intertrochanteric region (4 regions): Along the femoral neck axis, the anteromedial (AM), anterolateral (AL), posteromedial (PM), and posterolateral (PL) regions are defined on the inner and outer sides of the intertrochanteric line.

[0262] (3) Proximal femoral shaft (zone 2): 20 mm below the lower edge of the lesser trochanter, defining the medial and lateral cortical regions.

[0263] (4) Ward's triangle and subcapital region of the femoral head (1 region each): defined based on standard anatomical location. Each ROI is defined by a three-dimensional binary mask, which is generated and stored before training and applied to all samples.

[0264] In the registered standard space, each ROI has a fixed three-dimensional coordinate range, with the center of the femoral head as the spatial origin.

[0265] The three-dimensional binary mask of the ROI is pre-constructed using the following method:

[0266] A standard femoral CT image is selected and spatially standardized and registered to a universal template to serve as a reference image. On the reference image, the three-dimensional voxel regions corresponding to each of the 12 ROIs are delineated according to their anatomical definitions. The delineation results of each ROI are saved as a three-dimensional binary mask with the same size as the reference image, where the voxel values ​​within the ROI are set to 1 and the values ​​of the remaining voxels are set to 0. For each sample to be analyzed, after completing the same spatial standardization and registration, the predefined ROI mask is directly loaded for feature extraction without the need for re-delineation.

[0267] Compared to traditional self-supervised pre-training methods with random occlusion, the anatomically-aware occlusion and multi-scale reconstruction strategy employed in this application enables the second deep learning model to focus more on learning the complete morphological patterns of key biomechanical regions during the pre-training stage, and ensures feature quality through structural consistency constraints. With the same data scale, this method is expected to extract more discriminative features, thus exhibiting higher evaluation metrics such as AUC in cross-validation, particularly demonstrating superior feature discrimination ability in distinguishing structural features such as "normal bone density but abnormal spatial distribution pattern of cortical bone thickness".

[0268] 3.2 Performance Validation (i.e., evaluation methods for small datasets):

[0269] Five-fold cross-validation is used to make full use of the limited data for evaluation.

[0270] Key evaluation metrics: AUC (area under the curve), sensitivity (recall rate), and specificity.

[0271] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, characterized in that, include: S1. Obtain the initial three-dimensional medical image of the femur to be analyzed and perform preprocessing to obtain the preprocessed three-dimensional medical image of the femur. S2. Segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional mask of cortical bone. The local thickness of the cortical bone is calculated based on the three-dimensional mask of the cortical bone, and a three-dimensional thickness distribution map of the femoral cortical bone is generated. S3. Based on the feature extraction model and the target region of interest, extract the global depth feature vector and local statistical feature vector that represent the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map, and combine the feature vectors to obtain the comprehensive feature vector. The feature extraction model is self-supervised pre-training based on anatomical perception occlusion and multi-scale reconstruction strategies. S4. Using the pre-trained deep learning model, the comprehensive feature vector is quantitatively analyzed to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

2. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 1, characterized in that, The segmented and preprocessed three-dimensional medical image of the femur generates a bone tissue mask for the femur region. The bone tissue mask is then separated using an algorithm that combines region growing and level set evolution to obtain a three-dimensional mask of the cortical bone. Based on the calculation of local cortical bone thickness using a 3D mask, a 3D thickness distribution map of the femoral cortical bone is generated, including: S21. Based on the 3D nnU-Net framework, residual connections and attention gate mechanism, a lightweight segmentation model is constructed and supervised training is performed using a femur 3D medical image dataset with cortical bone segmentation annotations. S22. Use the trained lightweight segmentation model to perform region segmentation on the preprocessed femoral 3D medical image dataset to obtain the bone tissue mask of the femoral region. S23. Based on an algorithm combining region growing and level set evolution, the cortical bone region in the bone tissue mask is separated to obtain a three-dimensional cortical bone mask. S24. Using the three-dimensional Euclidean distance transformation method, calculate the local thickness of the cortical bone at each voxel point in the three-dimensional mask of cortical bone, and generate a three-dimensional thickness distribution map of the femoral cortical bone based on the local thickness of the cortical bone of all voxels.

3. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 2, characterized in that, The algorithm, which combines region growing and level set evolution, separates the cortical bone region from the bone tissue mask, resulting in a three-dimensional cortical bone mask including: S231. Based on the gray value distribution of computed tomography scan, an initial seed point is automatically selected within a preset range of the endosteal membrane in the bone tissue mask. S232. Using a three-dimensional region growth algorithm based on grayscale threshold, starting from the initial seed point, the region is gradually grown towards adjacent voxels to initially separate the cortical bone region and obtain a cortical bone roughness mask. S233. Using the roughened mask of cortical bone as the initial contour for the evolution of the level set, the level set algorithm based on edge and region information is used to generate the three-dimensional mask of cortical bone.

4. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 3, characterized in that, The energy functional expression for the evolution of the level set is: ； In the formula, Let be the energy functional used to control the evolution of the level set in the contour evolution process. H( is the level set function) ) is the Heaviside function used to distinguish between the inner and outer regions of a contour. Let Ω be the gradient of the Heaviside function used to represent the contour boundary, and Ω be the image domain. Let I be the gradient vector of the CT grayscale image I, representing the direction and intensity of grayscale changes in space. Let μ be an edge detection function used to reduce energy at image edges and guide the contour to fit the anatomical boundary, where μ, λ, and ν are weight parameters that control the strength of the contour length smoothing term, the edge attraction term, and the region area constraint term, respectively.

5. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 2, characterized in that, The method of using three-dimensional Euclidean distance transformation to calculate the local thickness of the cortical bone at each voxel point in the three-dimensional cortical bone mask, and generating a three-dimensional thickness distribution map of the femoral cortical bone based on the local cortical bone thickness of all voxels, includes: S241. Using three-dimensional morphological erosion operations, the boundary of the bone marrow cavity in the three-dimensional mask of cortical bone is extracted to obtain the voxel set of the boundary of the bone marrow cavity. S242. Extract the outer boundary of the bone in the three-dimensional mask of cortical bone through three-dimensional morphological dilatation operation to obtain the voxel set of the outer boundary of the bone; S243. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the boundary of the medullary cavity, and select the minimum distance as the Euclidean distance from that voxel to the nearest surface of the medullary cavity. S244. Calculate the three-dimensional Euclidean distance from each voxel in the three-dimensional mask of cortical bone to all voxels in the set of voxels at the outer boundary of bone, and select the minimum distance as the Euclidean distance from that voxel to the nearest outer surface of bone. S245. Calculate the local cortical bone thickness at each voxel point based on the Euclidean distance from the voxel to the nearest inner surface of the medullary cavity and the Euclidean distance from the voxel to the nearest outer surface of the bone. S246. Map the local thickness values ​​of cortical bone at all voxel points back to their corresponding positions in the original three-dimensional image space to form a three-dimensional thickness distribution map of the femoral cortical bone.

6. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 1, characterized in that, The method, based on the feature extraction model and the target region of interest, extracts global depth feature vectors and local statistical feature vectors representing the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map. These feature vectors are then concatenated to obtain a comprehensive feature vector, which includes: S31. Based on the three-dimensional deep convolutional autoencoder architecture, a feature extraction model is constructed, and self-supervised pre-training is performed using an anatomically perceptual occlusion and multi-scale reconstruction strategy. S32. Using a pre-trained feature extraction model, extract the global depth feature vector of the overall distribution pattern of cortical bone thickness in three-dimensional space from the three-dimensional thickness distribution map. S33. Based on the preset target region of interest, extract the statistical thickness features within the target region of interest in the 3D thickness distribution map to form a local statistical feature vector; S34. Concatenate the global deep feature vector with the local statistical feature vector to obtain the comprehensive feature vector.

7. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 6, characterized in that, The self-supervised pre-training using a strategy based on anatomical perception occlusion and multi-scale reconstruction includes: Construct a multi-level region of interest probability map corresponding to the femoral anatomy, and assign different initial occlusion probabilities to regions at different levels in the multi-level region of interest probability map. Based on the multi-level region of interest probability map, the three-dimensional thickness distribution map of the input femoral cortical bone is generatively sampled to obtain the initial masking layer. The initial masking mask is optimized using a three-dimensional conditional random field model to generate spatially coherent masking blocks. The original three-dimensional thickness distribution map is then masked on a voxel-by-voxel basis using the spatially coherent masking blocks to obtain the masked thickness map. A 3D deep convolutional autoencoder is trained using the occluded thickness map for reconstruction. The reconstruction loss function includes pixel-level reconstruction loss and structural consistency loss based on multi-scale feature matching.

8. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 1, characterized in that, The process of using a pre-trained deep learning model to perform quantitative analysis on the comprehensive feature vector and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, thereby achieving parameterized quantification of femoral cortical bone thickness characteristics, includes: S41. Construct a deep learning model based on a multi-layer fully connected neural network and perform supervised training using a dataset containing comprehensive features and quantization results. S42. Input the comprehensive feature vector into the trained deep learning model to generate quantitative analysis results that characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

9. The method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to claim 8, characterized in that, The loss function of the deep learning model is a binary classification focus loss function, and the expression of the binary classification focus loss function is: ； In the formula, L focal Let y represent the binary classification focus loss function, where N represents the number of samples, and y represents the number of samples. i p represents the true class label of the i-th sample. i Let represent the probability that the i-th sample belongs to the positive class, α represent the class weight coefficient of the positive class, and (1-α) represent the class weight coefficient of the negative class. This indicates the focus parameter.

10. A system for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images, used to implement the method for quantitative analysis of femoral cortical bone parameters based on three-dimensional medical images according to any one of claims 1-9, characterized in that, include: The femoral image preprocessing module is used to acquire the initial three-dimensional medical image of the femur to be analyzed and to preprocess it to obtain the preprocessed three-dimensional medical image of the femur. The cortical bone thickness calculation module is used to segment the preprocessed three-dimensional medical image of the femur, generate a bone tissue mask for the femur region, and use an algorithm combining region growing and level set evolution to separate the bone tissue mask to obtain a three-dimensional cortical bone mask. The local thickness of the cortical bone is calculated based on the three-dimensional mask of the cortical bone, and a three-dimensional thickness distribution map of the femoral cortical bone is generated. The femoral feature extraction and fusion module is used to extract global depth feature vectors and local statistical feature vectors that characterize the three-dimensional spatial distribution pattern of cortical bone in the three-dimensional thickness distribution map based on the feature extraction model and the target region of interest. The feature vectors are then combined to obtain a comprehensive feature vector. The feature extraction model is self-supervised pre-trained based on the anatomical perception occlusion and multi-scale reconstruction strategy. The quantitative analysis module is used to perform quantitative analysis on the comprehensive feature vector using a pre-trained deep learning model, and generate quantitative analysis results to characterize the spatial distribution pattern of femoral cortical bone thickness parameters, so as to achieve parameterized quantification of femoral cortical bone thickness characteristics.

Images