Multi-working-condition thighbone mechanical parameter generation method and device based on machine learning

Abstract

The embodiment of the invention discloses a multi-working-condition thighbone mechanical parameter generation method and device based on machine learning. According to one specific embodiment, the method comprises the steps that an image preprocessing end in an image processing system is configured to segment an original femoral bone near-end image set to obtain a three-dimensional femoral bone near-end image set; the image preprocessing end is configured to model the three-dimensional femur near-end image set to obtain a label data set; the image preprocessing end is configured to generate a first training set and a first test set corresponding to a three-dimensional proximal femur image set; the model training end is configured to generate an image omics feature training set and a deep learning feature training set, and generate an image omics feature test set and a deep learning feature test set; the model training end is configured to train the initial working condition parameter prediction model to obtain a working condition parameter prediction model; the model application end is configured to generate a multi-working-condition parameter prediction result set. According to the embodiment, the multi-working-condition thighbone mechanical parameters can be quickly and accurately predicted.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology, and in particular to a method, device, electronic device, and computer-readable medium for generating multi-working-condition femoral mechanical parameters based on machine learning. Background Art

[0002] In femoral mechanics analysis, high-precision image processing is the core foundation for the generation of femoral mechanical parameters. Submillimeter three-dimensional images obtained through quantitative scanning by QCT (Quantitative Computed Tomography) can accurately characterize the spatial distribution of bone density and the microstructural characteristics of trabecular bone. Existing clinical evaluation mainly relies on bone density measurement, but its ability to predict bone strength is limited and cannot reflect key mechanical influencing factors such as bone material distribution and microstructure. It is of great significance to develop more advanced, image-based bone mechanical parameter prediction methods. In the existing technology, there are three main technologies for evaluating bone mechanical parameters based on imaging: 1. Use dual-energy X-ray absorptiometry to measure bone density to indirectly reflect bone strength.

[0003] 2. Calculate bone mechanical parameters using finite element method based on QCT (Quantitative Computed Tomography) images.

[0004] 3. Establish a machine learning model based on image features.

[0005] However, when the above method is used to generate femoral mechanical parameters, the following problems often occur: Prediction accuracy is low, the modeling process is complex, the computation time is long, and the operator's technical skills are highly demanding. Existing methods typically only consider the prediction of a single mechanical parameter (such as bone strength) under a single working condition. Image feature extraction in existing methods is often manual, a time-consuming process with low feature reproducibility. Furthermore, a large number of image features reflecting bone material distribution and microstructure have not been fully explored. Summary of the Invention

[0006] The content of this disclosure is used to briefly introduce concepts that will be described in detail in the detailed description section below. The content of this disclosure is not intended to identify key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0007] Some embodiments of the present disclosure propose a method, device, electronic device and computer-readable medium for generating multi-working-condition femoral mechanical parameters based on machine learning to solve one or more of the technical problems mentioned in the above background technology section.

[0008] In a first aspect, some embodiments of the present disclosure provide a method for generating multi-working-condition femoral mechanical parameters based on machine learning, comprising: an image preprocessing end in an image processing system is configured to perform region-of-interest segmentation on an original proximal femoral image set to obtain a three-dimensional proximal femoral image set, and the image processing system further comprises: a model training end and a model application end; the image preprocessing end is configured to perform multi-working-condition finite element modeling on the three-dimensional proximal femoral image set to obtain a label data set under multiple working conditions, wherein each working condition is represented by adjusting a loading direction, and the loading direction is a direction parameter for applying an external force to the femur; the image preprocessing end is configured to generate a first training set and a first test set corresponding to the three-dimensional proximal femoral image set; the model training end is configured to generate an imaging genomics feature training set and a deep learning feature training set corresponding to the first training set, and to generate an imaging genomics feature test set and a deep learning feature test set corresponding to the first test set; The above-mentioned model training end is configured to train the initial working condition parameter prediction model based on the above-mentioned imaging genomics feature training set, the above-mentioned deep learning feature training set, the clinical feature information set corresponding to the above-mentioned first training set, and the corresponding loading direction set, to obtain the working condition parameter prediction model; the above-mentioned model application end is configured to generate multiple working condition parameter prediction result sets using the above-mentioned working condition parameter prediction model based on the above-mentioned imaging genomics feature test set, the above-mentioned deep learning feature test set, the clinical feature information set corresponding to the above-mentioned first test set, and the corresponding loading direction set.

[0009] In a second aspect, some embodiments of the present disclosure provide an electronic device comprising: one or more processors; a storage device on which one or more programs are stored, and when the one or more programs are executed by one or more processors, the one or more processors implement the method described in any implementation manner in the first aspect.

[0010] In a third aspect, some embodiments of the present disclosure provide a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method described in any implementation manner in the first aspect is implemented.

[0011] The various embodiments disclosed above have the following beneficial effects: A multi-condition femoral mechanical parameter generation method based on machine learning, as described in some embodiments of the present disclosure, can achieve rapid and accurate prediction of multiple femoral mechanical parameters under various working conditions. The decision-making process of the prediction model is explained using interpretable analysis methods, further clarifying the most important features in the femoral mechanical parameter prediction process. Specifically, the aforementioned technical issues arise from the fact that existing methods typically employ indirect measurement, resulting in low measurement accuracy. Existing methods typically extract image features manually, a time-consuming process with low feature reproducibility. Furthermore, a large number of image features reflecting bone material distribution and microstructure are not fully explored, and existing prediction models are typically black-box models, with opaque prediction processes. Based on this, the multi-condition femoral mechanical parameter generation method based on machine learning in some embodiments of the present disclosure includes the following steps: first, an image preprocessing end in an image processing system is configured to segment an original proximal femoral image set into a region of interest (ROI) to obtain a three-dimensional proximal femoral image set. The image processing system further includes a model training end and a model application end. By segmenting the proximal femoral ROI, the target region can be focused, reducing computational complexity. The three-dimensional reconstruction retains the spatial information of the bone structure, which is convenient for quantifying features. Then, the above-mentioned image preprocessing end is configured to perform multi-condition finite element modeling on the above-mentioned three-dimensional proximal femur image set to obtain a label data set under multiple working conditions, and each working condition is represented by adjusting the loading direction, and the above-mentioned loading direction is the direction parameter of the external force applied to the femur. The generated label data set provides high-quality training samples for the deep learning model, enhancing the generalization ability of the model. Then, the above-mentioned image preprocessing end is configured to generate a first training set and a first test set corresponding to the above-mentioned three-dimensional proximal femur image set. The training set and the test set are processed independently to avoid the test set information interfering with the training process. Secondly, the above-mentioned model training end is configured to generate an imaging genomics feature training set and a deep learning feature training set corresponding to the first training set, and to generate an imaging genomics feature test set and a deep learning feature test set corresponding to the first test set. Multimodal feature input can reduce the impact of single data noise and enhance generalization ability. Independently generating training / test set features can ensure the objectivity of model evaluation. Secondly, the model training end is configured to train the initial working condition parameter prediction model based on the imaging genomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set, to obtain the working condition parameter prediction model. The initial model structure can accelerate convergence and improve prediction accuracy through joint training of multi-source data. Adding clinical features and loading directions as model input can improve the robustness of the model. Finally, the model application end is configured to generate a multi-working condition parameter prediction result set based on the imaging genomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set using the working condition parameter prediction model.Independently validating model performance with a test set can reduce the risk of misjudgment. Furthermore, through the collaborative work of the image preprocessing, model training, and model application ends within the image processing system, the core technical goal of rapidly and accurately predicting femoral mechanical parameters under multiple working conditions was achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that components and elements are not necessarily drawn to scale.

[0013] Figure 1 is a flowchart of some embodiments of a method for generating multi-working condition femoral mechanical parameters based on machine learning according to the present disclosure; Figure 2 is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0014] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0015] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0017] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, they should be understood as "one or more".

[0018] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0019] The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0020] refer to Figure 1 , shows a process 100 of some embodiments of the method for generating multi-working condition femoral mechanical parameters based on machine learning according to the present disclosure. The method for generating multi-working condition femoral mechanical parameters based on machine learning includes the following steps: In step 101 , an image pre-processing end in an image processing system is configured to segment an original proximal femur image set into a region of interest to obtain a three-dimensional proximal femur image set.

[0021] In some embodiments, the image preprocessing end of the image processing system is configured to perform region-of-interest segmentation on the original proximal femur image set to obtain a three-dimensional proximal femur image set. The image processing system further includes a model training end and a model application end. The image preprocessing end may be a processing terminal (e.g., a server) that performs standardized segmentation and three-dimensional reconstruction on the original proximal femur images and outputs a labeled dataset. The model training end may be a processing terminal (e.g., a server) that trains a working condition parameter prediction model. The model application end may be a processing terminal (e.g., a server) that deploys the trained working condition parameter prediction model to generate femoral mechanical parameters. The original proximal femur images may be proximal femur QCT (Quantitative Computed Tomography) images. The three-dimensional proximal femur images may be digital three-dimensional models extracted through medical image segmentation. The proximal femur region of interest (ROI) may be, in machine vision or image processing, a box, circle, ellipse, irregular polygon, or other form of outline to delineate the proximal femur region to be processed from the processed proximal femur images.

[0022] In some optional implementations of some embodiments, the image preprocessing end of the image processing system can perform region-of-interest segmentation on the original proximal femur image set to obtain a three-dimensional proximal femur image set. In practice, TotalSegmentator v2 is first used to segment the proximal femur region of interest in the original proximal femur images to obtain segmented images. Then, two femoral image processing personnel manually adjust the segmented images using 3D Slicer to obtain three-dimensional proximal femur images.

[0023] In step 102 , the image preprocessing end is configured to perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label data set under multiple conditions.

[0024] In some embodiments, the image preprocessing end is configured to perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label data set under multiple conditions, and each condition is represented by adjusting the loading direction, and the loading direction is a directional parameter of the external force applied to the femur. Among them, the multiple conditions can be a variety of different stress states. For example, the multiple conditions may include: a standing condition and three falling conditions. Among them, the label data set may be a set of femoral mechanical parameters generated by multi-condition finite element modeling, including a set of femoral mechanical parameters corresponding to a standing condition and three falling conditions. The subset of femoral mechanical parameters for each condition includes: ultimate strength, yield strength and failure energy value. The loading direction can be used and Indicates that, It represents the angle between the loading direction on the coronal plane and the femoral shaft axis. In practice, for standing working conditions, the loading angle of the displacement load is , And points to the center of the femoral head. For the three fall conditions, the loading angles of the displacement load are ( , )、( , )and( , ). ( , ) indicates that the outside of the greater trochanter is hit), simulating a sideways fall, where the body's center of gravity is biased to the outside, and the impact force acts on the greater trochanter area along the outside of the femur. , ) indicates that the posterolateral aspect of the greater trochanter is impacted, simulating a fall with the trunk leaning back and externally rotating. The impact force acts on the posterolateral aspect of the greater trochanter, close to the base of the femoral neck. ( , ) indicates that the posterior side of the greater trochanter is hit, simulating a backward fall in which the posterior side of the femur is directly hit, and the torque acts on the posterior side of the proximal femur.

[0025] In some optional implementations of some embodiments, the image preprocessing end may perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label data set under multiple working conditions, where each working condition is represented by adjusting a loading direction, where the loading direction is a directional parameter for applying an external force to the femur, and the steps may include: The first step is to generate a mesh model set corresponding to the above-mentioned 3D proximal femur image set. The mesh model can be generated by performing voxel-level meshing on the 3D proximal femur image. In practice, first, the 3D proximal femur image in the above-mentioned 3D proximal femur image set can be re-cut using Mimics 21.0 software (for example, the size of a single voxel is Then, the “Create Voxel Mesh” function in Mimics 21.0 software was used to directly convert the single QCT voxel into Hexahedral elements are used to obtain the mesh model set corresponding to the three-dimensional proximal femur image set.

[0026] However, when the above method is used to obtain a mesh model set, the following technical problem often occurs: "The constructed mesh model is not accurate enough and wastes computing resources." The reason for the above technical problems is that the uniform mesh model uses the same unit size throughout the model and cannot be locally optimized for differences in mechanical properties. In complex structures such as the proximal femur, high stress gradient areas (for example, the cortical bone transition zone below the femoral neck) require finer mesh resolution to accurately capture stress mutations, and the uniform mesh model still maintains high-density division in mechanically insensitive areas (for example, the medullary cavity in the middle of the shaft), resulting in a large amount of computing resources being used in insignificant areas. The technical problem faced when obtaining a mesh model set is how to improve the insufficient accuracy of the mesh model and save computing resources. Therefore, it can be decided to adopt the following solution: Optionally, after generating the mesh model set corresponding to the three-dimensional proximal femur image set, the execution entity may include the following steps: The first step is to obtain a grayscale image set corresponding to the above-mentioned grid model set. The above-mentioned grid model set and the grayscale image set are in a one-to-one correspondence. The above-mentioned grayscale image set can be a set of three-dimensional digital images of the proximal femur obtained by quantitative QCT scanning. The above-mentioned three-dimensional digital images of the proximal femur can be an acquired tomographic image sequence. In practice, the original grayscale values ​​in the above-mentioned three-dimensional proximal femur image set can be converted into standard Hounsfield units by calibrating the phantom data to obtain the grayscale image set corresponding to the grid model set.

[0027] The second step is to establish a correspondence between the mesh model set and the grayscale image set, generating an enhanced mesh dataset with grayscale attributes. This enhanced mesh dataset can be an extended dataset based on the mesh model, with corresponding grayscale values ​​added to each node and element. In practice, image processing software can first be used to read the grayscale image set and mesh model set, extracting shared anatomical landmarks (e.g., the center of the femoral head, the apex of the greater trochanter, etc.), and then performing a coarse alignment using a feature point-based registration algorithm. Then, an ICP (Iterative Closest Point) algorithm is used for fine registration (e.g., determining the spatial distance between the surface nodes of the mesh model and the nearest voxel in the corresponding grayscale image, optimizing the rotation and translation matrices using singular value decomposition, and iterating until the average registration error is less than 0.3 mm, completing coordinate system 1). Finally, a voxel-to-mesh mapping is established. This mapping may include the following steps: first, assigning the HU value of the nearest grayscale image voxel to each mesh node. Then, trilinear interpolation is performed on the internal points of the element. Finally, an enhanced mesh dataset with grayscale attributes is generated, preserving the original topology while adding density field data.

[0028] In the third step, the enhanced mesh dataset is input into a pre-trained critical area identification model to obtain a stress thermodynamic atlas. The pre-trained critical area identification model can be a deep learning architecture based on a 3D convolutional neural network, which is specifically used to predict the mechanical critical area of ​​the proximal femur. The model is trained with a large amount of finite element simulation data and can directly infer the stress distribution from the geometric and density characteristics of the mesh. The pre-trained critical area identification model can include: an input layer, a feature extraction module, a mechanical prediction head, and a post-processing module. The stress thermodynamic atlas can be a color-coded image generated by finite element analysis or deep learning prediction, which intuitively displays the stress distribution of various regions of the femur. For example, red represents a high-stress danger zone, blue represents a low-stress safety zone, and yellow represents a transition zone.

[0029] The fourth step is to perform dynamic mesh optimization on the mesh model set using the stress thermodynamic atlas to obtain an optimized non-uniform mesh model set. The dynamic mesh optimization process includes high-stress area processing, low-stress area processing, and transition area processing. The optimized non-uniform mesh model set can be a set of mechanically adaptive mesh models obtained through dynamic mesh optimization. In practice, the high-stress area processing can be performed on the red areas in the thermodynamic atlas. First, the coordinates of the red high-risk areas can be extracted based on a preset stress threshold (e.g., 0.85) to generate a marker matrix. For example, an element of 1 in the marker matrix indicates an encrypted area, while an element of 0 in the marker matrix indicates a preserved area. Then, using the marker matrix and the corresponding mesh model, local network encryption (e.g., using a quadtree or octree algorithm) is performed to obtain an encrypted transition network. Finally, Laplace smoothing is performed on the encrypted transition network to obtain the mesh corresponding to the red areas in the thermodynamic atlas. The low-stress area processing can be performed on the blue areas in the thermodynamic atlas. In practice, low-stress areas (blue areas, corresponding to stress values ​​less than 0.3) can be identified based on the stress thermogram and the mesh elements to be simplified can be marked. Next, an edge-collapse algorithm is used to merge adjacent elements, controlling the size expansion rate to ≤300%, and outputting an intermediate simplified mesh. The aforementioned transition zone processing can be performed on the yellow areas in the thermogram. In practice, first, mesh elements in the yellow transition zone with stress values ​​of 0.3-0.7 are extracted based on the stress thermogram. Then, a gradient density algorithm is used to optimize mesh size transitions, inserting pyramid elements to ensure smooth connections. Finally, mechanical continuity is verified, distorted elements are corrected, and the final optimized mesh is output.

[0030] In the fifth step, the optimized non-uniform grid model set is determined as the grid model set corresponding to the three-dimensional proximal femur image set.

[0031] The above-mentioned optional steps and their related contents, as an inventive feature of an embodiment of the present disclosure, address the aforementioned technical problem of "the constructed mesh model lacks accuracy and wastes computing resources." Factors contributing to this technical problem are often as follows: uniform mesh models use the same cell size throughout the entire model, making it impossible to perform local optimization for differences in mechanical properties. In complex structures such as the proximal femur, high-stress gradient regions (e.g., the cortical bone transition zone below the femoral neck) require finer mesh resolution to accurately capture stress mutations. Furthermore, uniform mesh models maintain high-density meshing in mechanically insensitive regions (e.g., the medullary cavity in the mid-diaphysis), resulting in a significant amount of computing resources being spent on insignificant areas. Addressing these factors ensures improved mesh model accuracy and conserves computing resources when the mesh model is obtained. This solution uses intelligent mesh optimization to precisely allocate computing resources, ensuring that the cell density in high-stress and low-stress regions matches actual mechanical requirements. Mechanical continuity is ensured by using gradient cell sizes and pyramid-shaped transition cells in the transition region, thereby improving mesh model accuracy and conserving computing resources.

[0032] In the second step, material assignment is performed on each mesh model in the above mesh model set to obtain an assigned three-dimensional model set. Among them, material assignment can give the geometric structure of the mesh model the mechanical properties of real bone tissue, so that the finite element model can simulate the stress and strain behavior of real bone. In practice, first, the gray density is determined based on the equivalent density. Then, the inhomogeneous linear elastic and nonlinear post-yield material models used by other researchers are used to describe the stress and strain relationship of each unit based on the gray density. The bone material properties of the entire proximal femur are defined by 120 groups of materials, and the Poisson's ratio is set to 0.4.

[0033] In practice, the ash density can be determined by the following formula: ,in, represents the ash density, represents the equivalent density.

[0034] The third step is to use finite element analysis software to set boundary conditions for each of the above-mentioned working conditions in the assigned 3D model set to obtain a working condition configuration file. The finite element analysis software can be ABAQUS 14.1. The working condition configuration file can be a set of parameters defining different working conditions in the finite element analysis. Boundary conditions can refer to the constraints and load settings that simulate the real physical environment in the finite element analysis. These parameters can include: loading direction, displacement load, constraint settings, and bone material properties. In practice, the displacement load parameters for one standing condition and three fall conditions can be a 7mm displacement applied 30mm from the femoral head. The constraint settings for one standing condition can be set to completely fix the distal femur. When simulating the three fall conditions, the following constraints are applied to the proximal femoral model: a 6mm thick constraint layer (equivalent to two elements) is set on the surface of the greater trochanter opposite the femoral head loading area to resist the applied displacement load while allowing the model to move freely in the lateral direction. In addition, the elements in the model, including the nodes in the femoral head loading region and the nodes in the greater trochanter constraint region, were assigned specific material properties: an elastic modulus of 20 GPa and a yield strength of 200 MPa. The bone material properties of the remaining elements in the model were defined by another 120 sets of material parameters.

[0035] The fourth step is to generate a set of load-displacement curves based on the aforementioned working condition configuration file using the target yield criterion. The set includes load-displacement curves corresponding to each working condition. The target yield criterion can be the vos-Mises yield criterion. In practice, based on the working condition configuration file and the vos-Mises yield criterion, the reaction force in the loaded region of the femoral head in each increment can be obtained, thereby deriving the load-displacement curves for the proximal femur. The set of load-displacement curves represents a collection of mechanical responses under different working conditions from finite element analysis. Each load-displacement curve includes load-displacement curves corresponding to one standing posture and three fall conditions. Each curve is plotted using the reaction force (vertical axis) and displacement (horizontal axis) data points for the loaded region of the femoral head. For example, the standing posture curve shows an ultimate strength of 7510 N, while the posterior trochanter impact curve reduces the strength to 3500 N.

[0036] The fifth step is to extract the femoral mechanical parameters based on the above-mentioned load and displacement curve set to generate the femoral mechanical parameters corresponding to each working condition and obtain a label data set, wherein the above-mentioned femoral mechanical parameters include: ultimate strength, yield strength and failure energy. Among them, the above-mentioned ultimate strength is defined as the maximum total reaction force on the loading area of ​​the femoral head. The above-mentioned yield strength is defined as the load when at least one solid unit yields. The failure energy is defined as the area under the load and displacement curve at the maximum total reaction force. In practice, the ultimate strength, yield strength, and failure energy extracted under the four working conditions can be integrated into a label data set.

[0037] In step 103 , the image preprocessing end is configured to generate a three-dimensional proximal femur image set corresponding to a first training set and a first test set.

[0038] In some embodiments, the image preprocessing end is configured to generate a first training set and a first test set corresponding to the three-dimensional proximal femur image set, wherein the first training set is used to train an initial working condition parameter prediction model, and the first test set is used to verify the working condition parameter prediction model.

[0039] In some optional implementations of some embodiments, the image preprocessing end may generate the first training set and the first test set corresponding to the three-dimensional proximal femur image set, which may include the following steps: The first step is to pre-process the above three-dimensional proximal femur image set to obtain a pre-processed image set. In practice, the pre-processing of the above three-dimensional proximal femur image set can be performed by processing each three-dimensional proximal femur image in the above three-dimensional proximal femur image set according to 1×1×1mm. 3 The voxel size is resampled to obtain the preprocessed image set.

[0040] The second step is to perform two-dimensional projection and cropping on the preprocessed image set to obtain a two-dimensional projection image set. First, the resampled three-dimensional image is projected two-dimensionally in the anteroposterior position using the proximal femur mask obtained by semi-automatic segmentation. The anteroposterior position can be a standard projection orientation for medical imaging. The two-dimensional projection image is then cropped to 120×120 to ensure that the entire proximal femur is included. The value of each pixel in the two-dimensional projection image is obtained by accumulating the pixel values ​​of all femurs in the anteroposterior position.

[0041] The third step is to divide the preprocessed image set and the two-dimensional projection image set into a first training set and a first test set.

[0042] In step 104 , the model training end is configured to generate a radiomics feature training set and a deep learning feature training set corresponding to the first training set, and to generate a radiomics feature test set and a deep learning feature test set corresponding to the first test set.

[0043] In some embodiments, the above-mentioned model training end can be configured to generate an imaging omics feature training set and a deep learning feature training set corresponding to the first training set, and to generate an imaging omics feature test set and a deep learning feature test set corresponding to the first test set. Among them, the above-mentioned imaging omics feature training set and imaging omics feature test set can be a training set and a test set including imaging omics features. In practice, the above-mentioned imaging omics feature training set and test set can be divided into three groups according to size and shape features, intensity features and texture features. The above-mentioned deep learning feature training set and deep learning feature test set can be a training set and a test set including deep learning features. In practice, deep learning features can include: edge and texture features (for example, the edge contour of cortical bone and the microstructural texture of cancellous bone), grayscale distribution (for example, the intensity mean and variance of pixels) and morphological features (for example, the two-dimensional morphology of bone structure).

[0044] In some optional implementations of some embodiments, the model training end may generate a radiomics feature training set and a deep learning feature training set corresponding to the first training set, and generate a radiomics feature test set and a deep learning feature test set corresponding to the first test set, which may include the following steps: The first step is to extract radiomic features from the preprocessed image subset corresponding to the first training set, obtaining a radiomic feature training set. The preprocessed image subset is the subset of the preprocessed image set corresponding to the first training set. In practice, radiomic features from the radiomic feature training set can be extracted using the open-source Python package Pyradiomics.

[0045] The second step is to extract radiomic features from the preprocessed image subset corresponding to the first test set, generating a radiomics feature test set. The preprocessed image subset is the subset of the preprocessed image set corresponding to the first test set. In practice, radiomics features from the radiomics feature test set can be extracted using the open-source Python package Pyradiomics.

[0046] The third step is to divide the corresponding two-dimensional projection image set in the first training set into a two-dimensional training set and a two-dimensional test set. In practice, the ratio of the two-dimensional training set to the two-dimensional test set can be 4:1.

[0047] The fourth step is to normalize and enhance the above-mentioned two-dimensional training set, and to normalize the above-mentioned two-dimensional test set, to obtain a normalized and enhanced two-dimensional training set and a normalized two-dimensional test set. First, the two-dimensional projection images in the two-dimensional training set are Min-Max normalized to obtain a normalized two-dimensional training set. Then, each image in the above-mentioned normalized two-dimensional training set is transformed by horizontal and vertical movement, horizontal flipping, and grayscale inversion to perform data enhancement and obtain a normalized and enhanced two-dimensional test set. The above-mentioned method for normalizing the two-dimensional test set refers to the corresponding processing method for the two-dimensional training set and will not be repeated here.

[0048] In the fifth step, the pre-set initial deep feature information extraction model is fine-tuned using the normalized and augmented two-dimensional training set, the normalized two-dimensional test set, the target loss function, and the target optimizer to obtain a deep feature information extraction model. The initial deep feature information extraction model can be a modified ResNet50 model. The modified ResNet50 model includes 49 convolutional layers and one fully connected layer and can be divided into six modules. The first module includes one convolutional block and a maximum pooling layer. The second, third, fourth, and fifth modules contain 3, 4, 6, and 3 residual blocks, respectively, and the fifth module contains a global average pooling layer. To extract intermediate features from the network output, the last fully connected layer of the ResNet-50 model is deleted, and the features in the penultimate layer (global average pooling layer) are used as the output features.

[0049] Specifically, during training, the parameters of the first four modules of the deep feature information extraction model are frozen, and only the parameters of the last residual module and global average pooling layer are trained. The target optimizer can be an Adamax optimizer, with a learning rate set to 0.0001 and then decayed by a factor of 0.1 at the 50th epoch. Eight training samples are randomly selected from the training set and eight validation samples are randomly selected from the validation set for each batch, for a total of 100 epochs. To combat overfitting, L2 regularization is used with a weight decay of 0.01. The final model weights are determined based on the lowest validation set loss after each epoch. The weighted multi-task mean squared error (MSE) can be used as the loss function. All fine-tuning was performed on a personal computer with an Intel Core i9-14900K CPU, 32GB of RAM, and an NVIDIA RTX4070TI GPU.

[0050] In practice, the above loss function can be determined by the following formula: ,in, It represents the mean square error between the ultimate strength of the femur predicted by the model and the ultimate strength of the femur calculated by finite element method. represents the mean square error between the femoral yield strength predicted by the model and the femoral yield strength calculated by finite element method. It represents the mean square error between the failure energy predicted by the model and the failure energy calculated by the finite element method.

[0051] In the sixth step, the normalized and enhanced image training set corresponding to the first training set is input into the deep feature information extraction model to obtain a deep learning feature training set. The deep learning features in the deep learning feature training set are 2048-dimensional feature vectors. The model parameters corresponding to the deep learning feature test set remain consistent with the parameters determined for the deep learning feature training set.

[0052] Step 7: Normalize the corresponding two-dimensional projection image set in the first test set to obtain a normalized image test set. The specific steps are as described in the steps for normalizing the two-dimensional training set, which will not be repeated here.

[0053] In the eighth step, the normalized image test set is input into the deep feature information extraction model to obtain a deep learning feature test set, wherein the deep learning features in the deep learning feature test set are 2048-dimensional feature vectors.

[0054] In step 105 , the model training end is configured to train an initial working condition parameter prediction model based on the radiomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set to obtain a working condition parameter prediction model.

[0055] In some embodiments, the model training end is configured to train the initial working condition parameter prediction model based on the imaging genomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set to obtain the working condition parameter prediction model. The clinical features corresponding to the first training set may include: height, weight, age, and bone density. The loading direction may include Value and value.

[0056] In some optional implementations of some embodiments, the model training end may train an initial operating condition parameter prediction model based on the radiomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set to obtain the operating condition parameter prediction model, which may include the following steps: The first step is to preprocess the radiomics feature training set and the deep learning feature training set to obtain a processed radiomics feature training set and a processed deep learning feature training set. In practice, the radiomics feature training set and the deep learning feature training set are first subjected to feature variance filtering to obtain a first radiomics feature training set and a first deep learning feature training set. For example, the variance filtering method in the filter filtering method can be used to remove low-information features with a variance less than 0.4. Then, the first radiomics feature training set and the first deep learning feature training set are subjected to feature filtering using a Person correlation filtering algorithm to obtain a second radiomics feature training set and a second deep learning feature training set. For example, the Pearson correlation filtering algorithm can be used to retain features with high Person correlation coefficient rankings (e.g., the top 50). Then, the second radiomics feature training set and the second deep learning feature training set are subjected to feature filtering using a least absolute shrinkage and selection operator regression algorithm to obtain a third radiomics feature training set and a third deep learning feature training set. For example, the penalty coefficients in the least absolute shrinkage and selection operator regression are adjusted by 5-fold cross validation. Finally, the third imaging omics feature training set and the third deep learning feature training set are subjected to multicollinearity processing to obtain a processed imaging omics feature training set and a processed deep learning feature training set. Specifically, the Pearson correlation coefficient between each pair of features is determined. When the correlation coefficient is greater than or equal to 0.6, one of the features can be identified as a redundant feature, and thus one of them is removed. Among them, the imaging omics features in the processed imaging omics feature training set can be 20-dimensional features, and the deep learning features in the processed deep learning feature training set are 20-dimensional features.

[0057] In the second step, the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the subject set corresponding to the first training set, and the corresponding loading direction set are fused to obtain the target feature training set. The clinical features are 4-dimensional features, and the loading directions in the loading direction set are 2-dimensional features. The radiomics features (20 dimensions) in the processed radiomics feature training set, the deep learning features (20 dimensions) in the processed deep learning feature training set, the clinical features (4 dimensions), and the loading directions (2 dimensions) are vertically concatenated to obtain the target features (46 dimensions) in the target feature training set.

[0058] However, when using the above method to obtain the target feature training set, the following technical problem often arises: "High-dimensional features lead to excessive computational complexity and waste of computing resources." This is due to the fact that directly concatenating traditional high-dimensional features (46 dimensions) can lead to an excessive number of parameters in the operating condition parameter prediction model and increase the inefficient computational load. The technical challenge faced when obtaining the target feature training set is how to conserve computing resources. Therefore, the following solution can be adopted: Optionally, the execution entity may perform feature fusion on the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the object set corresponding to the first training set, and the corresponding loading direction set to obtain a target feature training set, which may include the following steps: The first step is to concatenate the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the subject set corresponding to the first training set, and the corresponding loading direction set to obtain an original feature matrix. The original feature matrix can be a 46-dimensional matrix obtained by vertically concatenating four types of standardized features, with each row corresponding to a sample and each column representing a feature, providing standardized input for subsequent graph structure construction. In practice, the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the subject set corresponding to the first training set, and the corresponding loading direction set can first be standardized (e.g., using Z-score normalization) to obtain the standardized processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the subject set corresponding to the first training set, and the corresponding loading direction set. Then, the standardized processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the subject set corresponding to the first training set, and the corresponding loading direction set are vertically concatenated to obtain the original feature matrix.

[0059] The second step is to determine the correlation coefficient matrix corresponding to the original feature matrix. The correlation coefficient matrix can be a 46×46 symmetric matrix obtained by determining the Pearson correlation coefficient of the correlation coefficient between two features in the original feature matrix (46 dimensions). For example, the elements of the correlation coefficient matrix are , represents the linear correlation strength between features i and j, the diagonal value is 1 (indicating autocorrelation), and the off-diagonal value reflects the correlation between cross-modal features (e.g., the correlation between radiomics features and clinical features).

[0060] The third step is to filter the correlation coefficient matrix to obtain a filtered matrix. The filtered matrix can be a sparse matrix obtained by retaining the elements (strong correlation feature pairs) with a preset absolute value threshold (e.g., 0.5) in the correlation coefficient matrix, removing weakly correlated or meaningless feature connections. This matrix only retains significant feature interactions (e.g., bone density and texture features, ), as the basis for building the graph structure, the dimension is still 46×46 but the number of non-zero elements is greatly reduced.

[0061] Step 4: Deduplication is performed on the filtered matrix to obtain a deduplication matrix. The deduplication matrix can be an undirected graph connection matrix obtained by eliminating symmetric duplicate edges in the filtered matrix. In practice, deduplication can be performed on the filtered matrix using undirected edge processing methods (e.g., deleting the symmetric edge (j, i) corresponding to edge (i, j)) to obtain a deduplication matrix.

[0062] Step 5: Supplement the self-loop edges in the deduplicated matrix to obtain an edge index matrix. This edge index matrix can be the edge connectivity table ultimately used in the pre-built fusion network model and can be formatted as a two-dimensional array. The first row represents the source node number, and the next row represents the target node number. This matrix contains the filtered valid feature connections and necessary self-loop edges. In practice, the deduplicated matrix can first be used to detect isolated nodes (e.g., nodes with degree = 0). Then, the self-loop edges (i, i) are added to obtain the edge index matrix.

[0063] Step 6: Encapsulate the correlation coefficient matrix and the edge index matrix to obtain graph data. The graph data may include node features and edge connections. In practice, the correlation coefficient matrix and the edge index matrix may be encapsulated in PyG format to obtain graph data.

[0064] In the seventh step, the graph data is fed into a pre-built fusion network model to obtain a target feature training set. Each target feature in the target feature training set can be an 8-dimensional fusion feature, with each dimension corresponding to a cross-modal feature combination pattern (e.g., a joint representation of "bone density + texture uniformity + loading direction"). In practice, the pre-built fusion network model can be a pre-built GraphSAGE model, which can include convolutional layers, skip connection layers, and normalization layers.

[0065] The above optional steps and their related contents serve as an inventive point of an embodiment of the present disclosure, which solves the above technical problem that "high-dimensional features lead to excessive computational complexity and waste of computing resources". The factors that lead to the above technical problems are often as follows: Direct splicing of traditional high-dimensional features (46 dimensions) may lead to excessive number of parameters in the working condition parameter prediction model and increase invalid computing load. Solving the above factors ensures that computing resources are saved when obtaining the target feature training set. This solution compresses the fusion features from 46 dimensions to 8 dimensions through a pre-built fusion network model, greatly reducing the number of downstream parameters. Then, by screening and eliminating irrelevant edges and weakly correlated edges in the correlation coefficient matrix, invalid calculations are reduced, thereby saving computing resources.

[0066] The third step is to randomly divide the standardized feature training set and the standardized label data set corresponding to the above-mentioned target feature training set and the corresponding label data set into mutually exclusive subsets of the target number. In practice, first, the above-mentioned target feature training set is Z-score standardized to eliminate the scale sensitivity in different features to obtain a standardized feature training set. Similarly, the label data set corresponding to the target feature training set is Z-score standardized to obtain a standardized label data set. Among them, the above-mentioned target feature training set and the corresponding label data set are subsets of the above-mentioned label data set. Among them, each label data includes three femoral mechanical parameters: femoral ultimate strength, yield strength and failure energy. Then, the above-mentioned standardized feature training set and the corresponding standardized label data set are randomly divided into mutually exclusive subsets of the target number. For example, the above-mentioned target number can be 5. Among them, the above-mentioned mutually exclusive subsets can be 5 non-overlapping subsets of similar size.

[0067] The fourth step is to determine the parameter range group corresponding to each hyperparameter of the hyperparameter set corresponding to the initial working condition parameter prediction model, and obtain the parameter range group set. Among them, the hyperparameter set includes the learning rate parameter, the weight decay parameter, the random inactivation rate parameter and the batch size parameter. Among them, the parameter range group corresponding to the learning rate parameter is (Select according to the order of magnitude). The parameter range group corresponding to the weight decay parameter is (Select according to the order of magnitude). The parameter range corresponding to the random inactivation rate parameter is , the batch size parameter can be 8 or 16 or 32.

[0068] In the sixth step, parameter ranges are selected from each parameter range group in the parameter range group set to generate parameter range combinations, thereby obtaining a parameter range combination set, wherein the parameter ranges in the parameter range combination correspond one-to-one with the hyperparameters in the hyperparameter set. For example, the parameter range combination set may include: a learning rate of 0.01, a weight decay of 0.01, a dropout rate of 0.1, and a batch size of 16. Alternatively, the parameter range combination set may include: a learning rate of 0.001, a weight decay of 0.001, a dropout rate of 0.3, and a batch size of 16.

[0069] Step 5: Randomly divide the standardized feature training set and the corresponding standardized label dataset into a target number of mutually exclusive subsets. For example, the target number can be 5. The mutually exclusive subsets can be 5 non-overlapping subsets of similar size.

[0070] In the sixth step, each hyperparameter range combination in the above parameter range combination set is cross-validated using the above mutually exclusive subsets to obtain the target hyperparameter combination.

[0071] Specifically, first, the above mutually exclusive subsets can be divided into training subsets and validation subsets. In practice, 5 mutually exclusive subsets can be used for 5 iterations, and 4 subsets can be selected as training subsets and 1 subset can be selected as validation subsets each time. The average value of the weighted multi-task mean square error obtained in the 5 iterations is determined as the validation result set. Then, the hyperparameter combination corresponding to the minimum validation result in the above validation result set is determined as the target hyperparameter combination. Among them, the above minimum validation result corresponds to the optimal hyperparameter combination, and the optimal hyperparameter combination is used as the target hyperparameter combination. In practice, the target hyperparameter combination can be a learning rate parameter of 1e-4 and a weight decay parameter of 1e-5. The batch size parameter is 32.

[0072] In the seventh step, based on the standardized feature training set and the standardized label dataset, the initial operating parameter prediction model corresponding to the target hyperparameter combination is trained using a target optimization algorithm and a target loss function to obtain the operating parameter prediction model. First, the standardized feature training set is divided into a standard training set and a standard validation set. In practice, the ratio of the standard training set to the standard validation set can be 4:1. Then, the operating parameter model is trained based on the standard training set, the standardized label dataset corresponding to the standard training set, the standard validation set, and the standardized label dataset corresponding to the standard validation set. The target optimization algorithm can be an Adamax optimization algorithm, and the target loss function can be a weighted multi-task mean squared error. The training cycle is 100 rounds, and the early stopping mechanism threshold is when the weighted multi-task mean squared error does not decrease for five consecutive rounds. The loss function is the loss determined based on the standard validation set and the corresponding standardized label dataset after each round of training.

[0073] In practice, the weighted multi-task mean square error can be obtained by the following formula: ,in, It represents the mean square error between the ultimate strength of the femur predicted by the model and the ultimate strength of the femur calculated by finite element method. represents the mean square error between the femoral yield strength predicted by the model and the femoral yield strength calculated by finite element method. It represents the mean square error between the failure energy predicted by the model and the failure energy calculated by the finite element method.

[0074] As an example, the imaging omics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the above-mentioned first training set, and the corresponding loading direction set are first preprocessed to obtain the preprocessed imaging omics feature training set, the preprocessed deep learning feature training set, the preprocessed clinical feature information set, and the preprocessed loading direction set. Then, the preprocessed imaging omics feature training set, the preprocessed deep learning feature training set, the preprocessed clinical feature information set, and the preprocessed loading direction set are spliced ​​to obtain a unified feature vector set. Then, the feature vectors in the unified feature vector set are fused through the Transformer encoder to obtain a fused feature vector set. Finally, based on the above-mentioned fused feature vector set and the corresponding label data set, the weighted multi-task loss function is used to train the initial working condition parameter prediction model to obtain the working condition parameter prediction model. Among them, the initial working condition parameter prediction model can be a Transformer model.

[0075] In step 106, the model application end is configured to generate a multi-operating condition parameter prediction result set using the operating condition parameter prediction model based on the imaging genomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set.

[0076] In some embodiments, the model application is configured to generate a multi-condition parameter prediction result set using the working condition parameter prediction model based on the radiomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set. The multi-condition parameter prediction result set may include three mechanical parameters (ultimate strength, yield strength, and failure energy) predicted for each sample in the first test set under four working conditions.

[0077] In some optional implementations of some embodiments, the model application end may generate a multi-operating condition parameter prediction result set using the operating condition parameter prediction model based on the radiomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set, and may include the following steps: The first step is to preprocess the radiomics feature test set and the deep learning feature test set to obtain a processed radiomics feature test set and a processed deep learning feature test set. The specific methods are described in the section "Generation of the Processed Radiomics Feature Training Set and the Processed Deep Learning Feature Training Set" and will not be repeated here.

[0078] The second step is to perform feature fusion on the processed radiomics feature test set, the processed deep learning feature test set, the clinical feature set corresponding to the subject set corresponding to the first test set, and the corresponding loading direction set to obtain the target feature test set. The specific method is described in "Generation of the Target Feature Training Set" and will not be repeated here.

[0079] In the third step, the target feature test set is input into the working condition parameter prediction model to obtain a multi-condition parameter prediction result set. Each multi-condition prediction result in this multi-condition prediction result set includes three femoral mechanical parameters predicted for the corresponding sample under multiple working conditions. For example, for the standing working condition, the three predicted femoral mechanical parameters are: 7510N, 3500N, and 16J.

[0080] As an example, based on the radiomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set, the above-mentioned working condition parameter prediction model is used to generate a multi-working condition parameter prediction result set, which may also include the following steps: The first step is to preprocess the above-mentioned imaging omics feature test set, the above-mentioned deep learning feature test set, and the clinical feature information set corresponding to the above-mentioned first test set to obtain a standardized imaging omics feature test set, a standardized deep learning feature test set, and a standardized clinical feature information set, and encode the above-mentioned loading direction set to obtain an encoded loading direction set. In practice, the mean and standard deviation corresponding to the above-mentioned imaging omics feature test set can be Min-Max normalized to obtain a standardized imaging omics feature test set. In practice, the mean and standard deviation corresponding to the above-mentioned deep learning feature test set can be Z-score normalized to obtain a standardized deep learning feature test set. In practice, the clinical feature information set corresponding to the above-mentioned first test set can be Z-score normalized to obtain a standardized clinical feature information set. In practice, the above-mentioned loading direction set can be one-hot encoded to obtain an encoded loading direction set. For example, the α and β corresponding to the loading direction can be converted into one-hot vectors (for example, for the standing working condition ( , ) is [0, 1, 0, 0, 1, 0, 0]).

[0081] The second step is to input the standardized radiomics feature test set and the corresponding labeled dataset into a first prediction model to obtain a first prediction result set. In practice, the first prediction model can be a random forest regression model. The prediction results in the first prediction result set can include the ultimate strength, yield strength, and failure energy for one standing posture and three fall conditions corresponding to the sample.

[0082] The third step is to input the standardized deep learning feature test set and the corresponding label data set into the second prediction model to obtain a second prediction result set. In practice, the second prediction model can be an XGBoost regression model.

[0083] In the fourth step, the standardized clinical feature information set, the encoded loading direction set, and the corresponding label data set are input into a third prediction model to obtain a third prediction result set. In practice, the third prediction model can be a multi-layer perceptron prediction model.

[0084] In the fifth step, the first, third, and fourth prediction result sets are weighted and integrated to obtain a multi-operating condition parameter prediction result set. In practice, the weight of the first prediction result set can be 0.3, the weight of the second prediction result set can be 0.4, and the weight of the third prediction result set can be 0.3.

[0085] In some optional implementations of the embodiments, the execution entity may input the target feature test set into the operating condition parameter prediction model to obtain a multi-operating condition parameter prediction result set, which may include the following steps: In the first step, the structure of the working condition parameter prediction model includes: a shared feature extraction module and an independent prediction module for the target number, wherein the above-mentioned shared feature extraction module includes: a residual connection layer, a layered normalization and a dynamic deactivation layer, and the above-mentioned independent prediction module for the target number includes: an ultimate strength prediction layer, a yield strength prediction layer and a failure energy prediction layer. Among them, the ultimate strength prediction layer is used to predict the ultimate strength parameters and may include: an input layer, a fully connected layer and an output layer. The yield strength prediction layer is used to predict the yield strength parameters and may include: an input layer, an attention mechanism module, a segmented regression head and a loss function. The failure energy prediction layer is used to predict the failure energy parameters and may include: an input layer, an LSTM layer, a fully connected layer and a loss function.

[0086] In the second step, the target feature test set is fed into the shared feature extraction module to generate a deep shared feature set. This module integrates radiomics features, deep learning features, and clinical features to extract common features across tasks. Specifically, the target feature training set is passed through a residual connection layer, hierarchical normalization, and dynamic dropout layer to generate a deep shared feature set.

[0087] The third step is to input the deep shared features into the ultimate strength prediction layer to obtain an ultimate strength prediction parameter set, which includes the ultimate strength prediction parameters under multiple working conditions for each sample in the target feature test set.

[0088] The fourth step is to input the deep shared features into the yield strength prediction layer to obtain a yield strength prediction parameter set, which includes the yield strength prediction parameters under multiple working conditions for each sample in the target feature test set.

[0089] In the fifth step, the deep shared features are input into the failure energy prediction layer to obtain a failure energy prediction parameter set, which includes the failure energy prediction parameters under multiple working conditions for each sample in the target feature test set.

[0090] In the sixth step, the ultimate strength prediction parameter set, the yield strength prediction parameter set and the failure energy set prediction parameter set are determined as a multi-operating condition parameter prediction result set.

[0091] Optionally, the above execution entity may further perform the following steps: The first step is to perform interpretability analysis on the multi-condition parameter prediction results set and the test feature information set corresponding to the first test set to obtain global analysis results. In practice, the SHAP (Shapley Additive Explanations) method can be used to determine global feature importance and generate a global importance ranking of key features. For example, the entropy of the gray-level co-occurrence matrix in radiomics has the greatest impact on ultimate strength.

[0092] The second step is to perform interpretability analysis on the samples corresponding to each test feature in the above test feature information set to obtain a local analysis summary. In practice, for each sample, the SHAP value can be used to calculate the local contribution of each feature in its prediction result. The local feature contribution distribution of all samples (for example, the characteristic patterns of high-frequency key features and abnormal samples) can be statistically analyzed to obtain a local analysis summary. For example, the low ultimate strength of a patient is mainly due to abnormal grayscale features.

[0093] The third step is to screen key features from the global analysis results and the local analysis summary results to obtain a key feature list. Specifically, statistical methods can be combined to screen high-contribution, high-frequency features. In practice, the above statistical method can be a variance filtering method or a Pearson correlation coefficient method. The above key feature list can include grayscale entropy from radiomics, texture features from deep learning, and clinical bone density.

[0094] The fourth step is to generate a medical explanation table based on the key feature list using a medical knowledge base. In practice, the medical knowledge base can be a collection of literature or guidelines that link features to bone mechanics and diseases. In practice, the medical explanation table can present the feature name, medical explanation, and associated disease risk level.

[0095] Step 5: Generate a key feature report using the medical interpretation table and the corresponding sample's clinical data. In practice, a patient-level key feature report can be generated by combining the corresponding sample's clinical data. For example, Patient A's advanced age (70 years) and low bone density (T-score -2.5) lead to a low ultimate strength prediction.

[0096] Step 6: Using the key feature reports, combine the global analysis results and the local analysis summary results to generate visualizations. In practice, visualizations can be generated by integrating global views (e.g., waterfall charts showing the total impact of each feature on the prediction), local views (e.g., scatter plot matrices showing the correlation between features and predicted values), and medical views (e.g., heat maps) with medical interpretations, along with feature importance and key feature reports.

[0097] The various embodiments disclosed above have the following beneficial effects: A multi-condition femoral mechanical parameter generation method based on machine learning, as described in some embodiments of the present disclosure, can achieve rapid and accurate prediction of multiple femoral mechanical parameters under various working conditions. The decision-making process of the prediction model is explained using interpretable analysis methods, further clarifying the most important features in the femoral mechanical parameter prediction process. Specifically, the aforementioned technical issues arise from the fact that existing methods typically employ indirect measurement, resulting in low measurement accuracy. Existing methods typically extract image features manually, a time-consuming process with low feature reproducibility. Furthermore, a large number of image features reflecting bone material distribution and microstructure are not fully explored, and existing prediction models are typically black-box models, with opaque prediction processes. Based on this, the multi-condition femoral mechanical parameter generation method based on machine learning in some embodiments of the present disclosure includes the following steps: first, an image preprocessing end in an image processing system is configured to segment an original proximal femoral image set into a region of interest (ROI) to obtain a three-dimensional proximal femoral image set. The image processing system further includes a model training end and a model application end. By segmenting the proximal femoral ROI, the target region can be focused, reducing computational complexity. The three-dimensional reconstruction retains the spatial information of the bone structure, which is convenient for quantifying features. Then, the above-mentioned image preprocessing end is configured to perform multi-condition finite element modeling on the above-mentioned three-dimensional proximal femur image set to obtain a label data set under multiple working conditions, and each of the above-mentioned working conditions is represented by adjusting the loading direction, and the above-mentioned loading direction is the direction parameter of the external force applied to the femur. The generated label data set provides high-quality training samples for the deep learning model, enhancing the generalization ability of the model. Then, the above-mentioned image preprocessing end is configured to generate a first training set and a first test set corresponding to the above-mentioned three-dimensional proximal femur image set. The training set and the test set are processed independently to avoid the test set information interfering with the training process. Secondly, the above-mentioned model training end is configured to generate an imaging genomics feature training set and a deep learning feature training set corresponding to the first training set, and to generate an imaging genomics feature test set and a deep learning feature test set corresponding to the first test set. Multimodal feature input can reduce the impact of single data noise and enhance generalization ability. Independently generating training / test set features can ensure the objectivity of model evaluation. Secondly, the model training end is configured to train the initial working condition parameter prediction model based on the imaging genomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set, to obtain the working condition parameter prediction model. The initial model structure can accelerate convergence and improve prediction accuracy through joint training of multi-source data. Adding clinical features and loading directions as model input can improve the robustness of the model. Finally, the model application end is configured to generate a multi-working condition parameter prediction result set based on the imaging genomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set using the working condition parameter prediction model.Independently validating model performance with a test set can reduce the risk of misjudgment. Furthermore, through the collaborative work of the image preprocessing, model training, and model application ends within the image processing system, the core technical goal of rapidly and accurately predicting femoral mechanical parameters under multiple working conditions was achieved.

[0098] Reference below Figure 2 , which shows a structural schematic diagram of an electronic device (eg, an electronic device) 200 suitable for implementing some embodiments of the present disclosure. Figure 2 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0099] like Figure 2 As shown, electronic device 200 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 201, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 202 or programs loaded from a storage device 208 into a random access memory (RAM) 203. Various programs and data required for the operation of electronic device 200 are also stored in RAM 203. Processing device 201, ROM 202, and RAM 203 are connected to each other via a bus 204. An input / output (I / O) interface 205 is also connected to bus 204.

[0100] Typically, the following devices may be connected to the I / O interface 205: an input device 206 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 207 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 208 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 209. The communication device 209 may allow the electronic device 200 to communicate with other devices wirelessly or by wire to exchange data. Figure 2 The electronic device 200 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead. Figure 2 Each block shown in the figure may represent one device, or may represent multiple devices as needed.

[0101] In particular, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In some such embodiments, the computer program can be downloaded and installed from a network via the communication device 209, or installed from the storage device 208, or installed from the ROM 202. When the computer program is executed by the processing device 201, the above-mentioned functions defined in the method of some embodiments of the present disclosure are performed.

[0102] It should be noted that in some embodiments of the present disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable storage media may include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or components, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In some embodiments of the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component. Furthermore, in some embodiments of the present disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wire, optical cable, RF (radio frequency), or any suitable combination thereof.

[0103] In some embodiments, the client and server can communicate using any currently known or later developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or later developed network.

[0104] The computer-readable medium may be included in the electronic device; or it may exist independently without being assembled into the electronic device. The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: performs region of interest segmentation on the original proximal femur image set to obtain a three-dimensional proximal femur image set; performs multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label data set under multiple conditions, where each condition is represented by adjusting the loading direction, and the loading direction is a direction parameter for applying external force to the femur; generates a first training set and a first test set corresponding to the three-dimensional proximal femur image set; generates an imaging genomics feature training set and a deep learning feature training set corresponding to the first training set, and generates an imaging genomics feature test set and a deep learning feature test set corresponding to the first test set; Based on the above-mentioned imaging genomics feature training set, the above-mentioned deep learning feature training set, the clinical feature information set corresponding to the above-mentioned first training set, and the corresponding loading direction set, the initial working condition parameter prediction model is trained to obtain the working condition parameter prediction model; based on the above-mentioned imaging genomics feature test set, the above-mentioned deep learning feature test set, the clinical feature information set corresponding to the above-mentioned first test set, and the corresponding loading direction set, the above-mentioned working condition parameter prediction model is used to generate a multi-working condition parameter prediction result set.

[0105] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0106] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the prescribed logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that the combination of each box diagram and / or box in the flow chart can be implemented with a dedicated hardware-based system that performs the prescribed function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0107] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0108] The above descriptions are merely some preferred embodiments of the present disclosure and illustrate the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A method for generating multi-condition femoral mechanical parameters based on machine learning, characterized in that: include: The image preprocessing end of the image processing system is configured to perform region of interest segmentation on the original proximal femur image set to obtain a three-dimensional proximal femur image set. The image processing system further includes: a model training end and a model application end; The image preprocessing end is configured to perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label data set under multiple conditions, each condition being represented by adjusting a loading direction, wherein the loading direction is a directional parameter of an external force applied to the femur; The image preprocessing end is configured to generate a first training set and a first test set corresponding to the three-dimensional proximal femur image set; The model training end is configured to generate a radiomics feature training set and a deep learning feature training set corresponding to the first training set, and to generate a radiomics feature test set and a deep learning feature test set corresponding to the first test set; The model training end is configured to train an initial working condition parameter prediction model based on the radiomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set to obtain a working condition parameter prediction model; The model application end is configured to generate a multi-operating condition parameter prediction result set based on the imaging genomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set, using the operating condition parameter prediction model.

2. The method according to claim 1, characterized in that The multi-condition finite element modeling is performed on the three-dimensional proximal femur image set to obtain a label data set under multiple conditions, including: generating a grid model set corresponding to the three-dimensional proximal femur image set; Assigning material to each grid model in the grid model set to obtain an assigned three-dimensional model set; By using finite element analysis software, boundary conditions for each working condition in the assigned three-dimensional model set are set to obtain a working condition configuration file; generating a load and displacement curve set according to the working condition configuration file using a target yield criterion, wherein the load and displacement curve set includes a load and displacement curve corresponding to each working condition; Femoral mechanical parameters are extracted according to the load and displacement curve set to generate femoral mechanical parameters corresponding to each working condition, thereby obtaining a label data set, wherein the femoral mechanical parameters include ultimate strength, yield strength, and failure energy.

3. The method according to claim 1, characterized in that Generating the three-dimensional proximal femur image set corresponding to the first training set and the first test set includes: Preprocessing the three-dimensional proximal femur image set to obtain a preprocessed image set; Performing two-dimensional projection on the preprocessed image set to obtain a two-dimensional projection image set; The preprocessed image set and the two-dimensional projection image set are divided to obtain a first training set and a first test set.

4. The method according to claim 1, wherein Generating a radiomics feature training set and a deep learning feature training set corresponding to the first training set, and generating a radiomics feature test set and a deep learning feature test set corresponding to the first test set, includes: Performing radiomics feature extraction on a corresponding subset of preprocessed images in the first training set to obtain a radiomics feature training set; Performing radiomics feature extraction on a corresponding subset of preprocessed images in the first test set to obtain a radiomics feature test set; Dividing the corresponding two-dimensional projection image set in the first training set to obtain a two-dimensional training set and a two-dimensional test set; Normalizing and enhancing the two-dimensional training set, and normalizing the two-dimensional test set to obtain a normalized and enhanced two-dimensional training set and a normalized two-dimensional test set; Fine-tuning a preset initial deep feature information extraction model using the normalized and enhanced two-dimensional training set, the target loss function, and the target optimizer to obtain a deep feature information extraction model; Inputting the normalized and enhanced image training set corresponding to the first training set into the deep feature information extraction model to obtain a deep learning feature training set; Normalizing the corresponding two-dimensional projection image set in the first test set to obtain a normalized image test set; The normalized image test set is input into the deep feature information extraction model to obtain a deep learning feature test set.

5. The method according to claim 1, characterized in that The initial operating condition parameter prediction model is trained based on the radiomics feature training set, the deep learning feature training set, the clinical feature information set corresponding to the first training set, and the corresponding loading direction set to obtain the operating condition parameter prediction model, including: Preprocessing the radiomics feature training set and the deep learning feature training set to obtain a processed radiomics feature training set and a processed deep learning feature training set; Performing feature fusion on the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the object set corresponding to the first training set, and the corresponding loading direction set to obtain a target feature training set; Randomly divide the standardized feature training set and the standardized label data set corresponding to the target feature training set and the corresponding label data set into mutually exclusive subsets of the target number; Determine a parameter range group corresponding to each hyperparameter of the hyperparameter set corresponding to the initial operating condition parameter prediction model to obtain a parameter range group set; Selecting parameter ranges from each parameter range group in the parameter range group set to generate parameter range combinations, thereby obtaining a parameter range combination set; cross-validating each parameter range combination in the parameter range combination set using the mutually exclusive subsets to obtain a target hyperparameter combination; According to the standardized feature training set and the standardized label data set, the initial operating parameter prediction model corresponding to the target hyperparameter combination is trained using a target optimization algorithm and a target loss function to obtain an operating parameter prediction model.

6. The method according to claim 1, characterized in that The method of generating a multi-operating condition parameter prediction result set based on the radiomics feature test set, the deep learning feature test set, the clinical feature information set corresponding to the first test set, and the corresponding loading direction set, and using the operating condition parameter prediction model, includes: Preprocessing the radiomics feature test set and the deep learning feature test set to obtain a processed radiomics feature test set and a processed deep learning feature test set; Performing feature fusion on the processed radiomics feature test set, the processed deep learning feature test set, the clinical feature set corresponding to the object set corresponding to the first test set, and the corresponding loading direction set to obtain a target feature test set; The target feature test set is input into the operating condition parameter prediction model to obtain a multi-operating condition parameter prediction result set.

7. The method according to claim 6, characterized in that The structure of the working condition parameter prediction model includes: a shared feature extraction module and an independent prediction module for the target number, wherein the shared feature extraction module includes: a residual connection layer, a layered normalization and a dynamic deactivation layer, and the independent prediction module for the target number includes: an ultimate strength prediction layer, a yield strength prediction layer and a failure energy prediction layer; and Inputting the target feature test set into the operating condition parameter prediction model to obtain a multi-operating condition parameter prediction result set includes: Inputting the target feature test set into a shared feature extraction module to obtain a deep shared feature set; Inputting the deep shared features into the ultimate strength prediction layer to obtain an ultimate strength prediction parameter set; Inputting the deep shared features into the yield strength prediction layer to obtain a yield strength prediction parameter set; Inputting the deep shared features into the failure energy prediction layer to obtain a failure energy prediction parameter set; The ultimate strength prediction parameter, the yield strength prediction parameter, and the failure energy prediction parameter are determined as a multi-operating condition parameter prediction result set.

8. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

9. A computer-readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images