Machine learning based multi-condition femur mechanical parameter generation method and device

Abstract

Embodiments of the present disclosure disclose a machine learning-based multi-condition femur mechanical parameter generation method and device. A specific embodiment of the method includes: an image preprocessing end in an image processing system is configured to segment an original proximal femur image set to obtain a three-dimensional proximal femur image set; the image preprocessing end is configured to model the three-dimensional proximal femur 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 the three-dimensional proximal femur image set; a model training end is configured to generate an imageomics feature training set and a deep learning feature training set, and generate an imageomics feature test set and a deep learning feature test set; the model training end is configured to train an initial condition parameter prediction model to obtain a condition parameter prediction model; and a model application end is configured to generate a multi-condition parameter prediction result set. The embodiment can quickly and accurately predict multi-condition femur mechanical parameters.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computer technology, and more specifically to a method, apparatus, electronic device, and computer-readable medium for generating multi-condition femoral biomechanical parameters based on machine learning. Background Technology

[0002] In femoral biomechanics analysis, high-precision image processing is the core foundation for generating femoral biomechanical parameters. Sub-millimeter-level three-dimensional images acquired through quantitative CT (Quantitative Computed Tomography) scanning can accurately characterize the spatial distribution of bone mineral density and the microstructural features of trabecular bone. Current clinical assessments primarily rely on bone mineral density measurement, but its predictive ability for bone strength is limited, failing to reflect key mechanical influencing factors such as bone material distribution and microstructure. Developing more advanced, image-based methods for predicting bone biomechanical parameters is of great significance. Existing technologies for image-based assessment of bone biomechanical parameters mainly include three types:

[0003] 1. Bone mineral density is measured using dual-energy X-ray absorptiometry to indirectly reflect bone strength.

[0004] 2. Calculate bone biomechanical parameters using finite element method based on QCT (Quantitative Computed Tomography) images.

[0005] 3. Establish machine learning models based on image features.

[0006] However, when using the above method to generate femoral biomechanical parameters, the following problems often occur:

[0007] The prediction accuracy is low, the modeling process is complex, the computation time is long, and the technical requirements for operators are high. Existing methods usually only consider the prediction of one mechanical parameter (such as bone strength) under one working condition. When extracting image features, existing methods usually use manual extraction, which is time-consuming and has a low feature repetition rate. In addition, a large number of image features reflecting bone material distribution and bone microstructure have not been fully explored. Summary of the Invention

[0008] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0009] Some embodiments of this disclosure propose a method, apparatus, electronic device, and computer-readable medium for generating multi-condition femoral biomechanical parameters based on machine learning to solve one or more of the technical problems mentioned in the background section above.

[0010] In a first aspect, some embodiments of this disclosure provide a method for generating femoral biomechanical parameters under multiple working conditions based on machine learning, comprising: an image preprocessing end in an image processing system is configured to segment the region of interest in an original proximal femoral image set to obtain a three-dimensional proximal femoral image set; 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 dataset under multiple working conditions, wherein each working condition is represented by adjusting the loading direction, and the loading direction is a directional parameter for applying 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 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;

[0011] The model training end is configured to train an initial operating condition parameter prediction model based on the aforementioned radiomics feature training set, the aforementioned deep learning feature training set, the clinical feature information set corresponding to the aforementioned first training set, and the corresponding loading direction set, thereby obtaining the operating condition parameter prediction model; the model application end is configured to generate a multi-operating condition parameter prediction result set based on the aforementioned radiomics feature test set, the aforementioned deep learning feature test set, the clinical feature information set corresponding to the aforementioned first test set, and the corresponding loading direction set, using the aforementioned operating condition parameter prediction model.

[0012] In a second aspect, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0013] Thirdly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0014] The various embodiments of this disclosure have the following beneficial effects: Through a machine learning-based multi-condition femoral biomechanical parameter generation method according to some embodiments of this disclosure, multiple femoral biomechanical parameters under various working conditions can be predicted quickly and accurately. Furthermore, interpretability analysis methods are used to explain the decision-making process of the prediction model, further clarifying the most important features in the prediction process of these femoral biomechanical parameters. Specifically, the reasons for the above-mentioned technical problems are: existing methods usually use indirect measurement, resulting in low measurement accuracy. Existing methods typically extract image features manually, which is time-consuming and has a low feature repetition rate. In addition, a large number of image features reflecting bone material distribution and bone microstructure have not been fully explored, and existing prediction models are usually black-box models, with opaque prediction processes. Based on this, the machine learning-based multi-condition femoral biomechanical parameter generation method of some embodiments of this disclosure first configures the image preprocessing end of the image processing system to segment the region of interest (ROI) of the original proximal femoral image set to obtain a three-dimensional proximal femoral image set. The image processing system also 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. 3D reconstruction preserves spatial information of the bone structure, facilitating feature quantification. Then, the image preprocessing unit is configured to perform multi-condition finite element modeling on the aforementioned 3D proximal femoral image set, obtaining a label dataset under multiple conditions. Each condition is represented by adjusting the loading direction, which is the directional parameter of the external force applied to the femur. The generated label dataset provides high-quality training samples for the deep learning model, enhancing its generalization ability. Next, the image preprocessing unit is configured to generate a first training set and a first test set corresponding to the aforementioned 3D proximal femoral image set. Independent processing of the training and test sets avoids interference from test set information during the training process. Furthermore, the model training unit is configured to generate a radiomics feature training set and a deep learning feature training set corresponding to the first training set, as well as a radiomics feature test set and a deep learning feature test set corresponding to the first test set. Multimodal feature input reduces the influence of single-data noise, enhancing generalization ability. Independent generation of training / test set features ensures the objectivity of model evaluation. Secondly, the model training end is configured to train an initial operating condition parameter prediction model based on the aforementioned radiomics feature training set, the aforementioned deep learning feature training set, the clinical feature information set corresponding to the aforementioned first training set, and the corresponding loading direction set, thus obtaining the operating condition parameter prediction model. The initial model structure, trained jointly with multi-source data, can accelerate convergence and improve prediction accuracy. Incorporating clinical features and loading direction as inputs to the model can enhance its robustness. Finally, the model application end is configured to generate a multi-operating condition parameter prediction result set using the aforementioned operating condition parameter prediction model based on the aforementioned radiomics feature test set, the aforementioned deep learning feature test set, the clinical feature information set corresponding to the aforementioned first test set, and the corresponding loading direction set.Independently validating model performance using a test set reduces the risk of misjudgment. Furthermore, by leveraging the collaborative work of the image preprocessing, model training, and model application ends within the image processing system, the core technical objective of rapidly and accurately predicting femoral biomechanical parameters under various working conditions was achieved. Attached Figure Description

[0015] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0016] Figure 1 This is a flowchart of some embodiments of the machine learning-based multi-condition femoral biomechanical parameter generation method according to the present disclosure;

[0017] Figure 2 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0018] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0019] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0020] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0023] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] refer to Figure 1 The diagram illustrates a flow 100 of some embodiments of a machine learning-based multi-condition femoral biomechanical parameter generation method according to the present disclosure. This machine learning-based multi-condition femoral biomechanical parameter generation method includes the following steps:

[0025] Step 101: 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.

[0026] In some embodiments, the image preprocessing end of the above-described image processing system is configured to perform region of interest segmentation on the original proximal femoral image set to obtain a three-dimensional proximal femoral image set. The image processing system further includes a model training end and a model application end. The image preprocessing end can be a processing terminal (e.g., a server) that performs standardized segmentation and three-dimensional reconstruction on the original proximal femoral images and outputs a labeled dataset. The model training end can be a processing terminal (e.g., a server) that trains a condition parameter prediction model. The model application end can be a processing terminal (e.g., a server) that deploys the trained condition parameter prediction model to generate femoral biomechanical parameters. The original proximal femoral image can be a proximal femoral QCT (Quantitative Computed Tomography) image. The three-dimensional proximal femoral image can be a digitized three-dimensional model extracted from medical image segmentation. The proximal femoral region of interest (ROI) can be defined in machine vision and image processing by delineating the proximal femoral region to be processed from the processed proximal femoral image using methods such as rectangles, circles, ellipses, and irregular polygons.

[0027] In some optional implementations of certain embodiments, the image preprocessing unit of the aforementioned image processing system can perform region-of-interest (ROI) segmentation on the original proximal femoral image set to obtain a three-dimensional proximal femoral image set. In practice, firstly, TotalSegmentator v2 is used to segment the proximal femoral RIO region of the original proximal femoral image to obtain the segmented image. Then, two femoral image processing personnel manually adjust the segmented image using a 3D Slicer to obtain the three-dimensional proximal femoral image.

[0028] 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 dataset under multiple conditions.

[0029] In some embodiments, the image preprocessing terminal is configured to perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a label dataset under multiple conditions. Each condition is represented by adjusting the loading direction, which is a directional parameter of the external force applied to the femur. The multiple conditions can be various different stress states. For example, the multiple conditions may include: one standing condition and three fall conditions. The label dataset can be a set of femoral mechanical parameters generated through multi-condition finite element modeling, including the femoral mechanical parameter sets corresponding to one standing condition and three fall conditions. The subset of femoral mechanical parameters for each condition includes: ultimate strength, yield strength, and failure energy value. The loading direction can be... and It means that, among them, This indicates the angle between the loading direction on the coronal plane and the femoral shaft axis. This represents the angle between the loading direction on the cross-section and the femoral neck axis. In practice, for standing load conditions, the loading angle of the displacement load is... , And pointing towards the center of the femoral head. For the three fall scenarios, the loading angles of the displacement load are respectively ( , ), ( , )and( , ). , (This indicates an impact to the outer side of the greater trochanter), simulating a lateral fall where the body's center of gravity shifts outward, and the impact force acts along the outer side of the femur onto the greater trochanter region. , This indicates an impact to the posterolateral aspect of the greater trochanter, simulating a fall where the trunk leans backward with external rotation. The impact force acts on the posterolateral aspect of the greater trochanter, near the base of the femoral neck. , This indicates that the posterior side of the greater trochanter is impacted, simulating a backward fall where the posterior side of the femur is directly impacted, with the torque acting on the posterior proximal end of the femur.

[0030] In some optional implementations of certain embodiments, the image preprocessing unit can perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a labeled dataset under multiple conditions. Each condition is represented by adjusting the loading direction, which is a directional parameter of the external force applied to the femur. This can include the following steps:

[0031] The first step is to generate a mesh model set corresponding to the aforementioned 3D proximal femoral image set. This mesh model can be generated by dividing the 3D proximal femoral images into voxel-level meshes. In practice, firstly, Mimics 21.0 software can be used to re-segment the 3D proximal femoral images in the aforementioned 3D proximal femoral image set (e.g., with a single voxel size of...). Then, using the "Create Voxel Mesh" function in Mimics 21.0 software, a single QCT voxel is directly converted into... Hexahedral elements are used to obtain the mesh model set corresponding to the three-dimensional proximal femoral image set.

[0032] However, when obtaining a mesh model set using the above method, the following technical problem often exists: "The constructed mesh model has insufficient accuracy and wastes computational resources." The reason for this is that uniform mesh models use the same element size throughout the model, making local optimization for differences in mechanical properties impossible. 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 abrupt changes. Furthermore, uniform mesh models maintain high-density partitioning in mechanically insensitive regions (e.g., the medullary canal in the middle of the femoral shaft), resulting in a large amount of computational resources being used for insignificant areas. The technical challenge in obtaining the mesh model set is how to improve the mesh model's accuracy while conserving computational resources. Therefore, the following solution can be adopted:

[0033] Optionally, after generating the mesh model set corresponding to the three-dimensional proximal femoral image set, the execution entity may include the following steps:

[0034] The first step is to obtain the grayscale image set corresponding to the aforementioned mesh model set. There is a one-to-one correspondence between the mesh model set and the grayscale image set. The grayscale image set can be a collection of three-dimensional digital images of the proximal femur obtained through quantitative QCT scanning. These three-dimensional digital images of the proximal femur can be acquired tomographic image sequences. In practice, the original grayscale values ​​in the aforementioned three-dimensional proximal femur image set can be converted to standard Hounsfield units using calibration phantom data to obtain the grayscale image set corresponding to the mesh model set.

[0035] The second step is to establish the correspondence between the aforementioned mesh model set and the aforementioned grayscale image set, resulting in 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 cell. In practice, firstly, image processing software can be used to read the grayscale image set and the mesh model set, extracting common anatomical landmarks (e.g., the center of the femoral head, the apex of the greater trochanter, etc.), and performing coarse alignment using a feature-point-based registration algorithm. Then, the 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 voxels in the corresponding grayscale images, optimizing the rotation and translation matrices through singular value decomposition, iterating until the average registration error is less than 0.3 mm, completing coordinate system one). Finally, the mapping relationship from voxels to the mesh is established. Establishing the mapping relationship may include the following steps: First, assigning the HU value of the nearest grayscale image voxel to each mesh node. Then, performing trilinear interpolation on the points within the cells. Finally, an enhanced mesh dataset with grayscale attributes is generated, preserving the original topology while adding density field data.

[0036] The third step involves inputting the enhanced mesh dataset into a pre-trained critical region identification model to obtain a stress heatmap. This pre-trained critical region identification model can be a deep learning architecture based on a 3D convolutional neural network, specifically designed to predict the biomechanical critical regions of the proximal femur. The model is trained using a large amount of finite element simulation data and can directly infer stress distribution from the geometric and density features of the mesh. The pre-trained critical region identification model may include an input layer, a feature extraction module, a biomechanical prediction head, and a post-processing module. The stress heatmap can be a color-coded image generated through finite element analysis or deep learning prediction, visually displaying the stress distribution in different regions of the femur. For example, red represents high-stress danger zones, blue represents low-stress safe zones, and yellow represents transition zones.

[0037] The fourth step involves dynamically optimizing the aforementioned mesh model set using the stress-thermal map, resulting in an optimized non-uniform mesh model set. This dynamic optimization includes processing high-stress areas, low-stress areas, and transition areas. The optimized non-uniform mesh model set can be a mechanically adapted mesh model set obtained through dynamic mesh optimization. In practice, the high-stress area processing can be applied to the red areas in the heatmap. 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 = 1 in the marker matrix indicates a densified area, and an element = 0 indicates a preserved area. Then, local network densification (e.g., using a quadtree or octree algorithm) is performed using the marker matrix and the corresponding mesh model to obtain a densified transition network. Finally, Laplace smoothing is applied to the densified transition network to obtain the mesh corresponding to the red areas in the heatmap. The low-stress area processing can be applied to the blue areas in the heatmap. In practice, firstly, low-stress areas (blue areas, corresponding to stress values ​​less than 0.3) can be identified based on the stress heatmap, and the mesh elements to be simplified are marked. Then, an edge-folding algorithm is used to merge adjacent elements, controlling the dimensional expansion rate to ≤300%, and an intermediate simplified mesh is output. The above transition zone processing can be performed on the yellow areas in the heatmap. In practice, firstly, mesh elements with stress values ​​of 0.3-0.7 in the yellow transition zone are extracted based on the stress heatmap. Then, a gradient density algorithm is used to optimize the mesh size transition, and pyramid elements are inserted to ensure smooth connections. Finally, mechanical continuity is verified and distorted elements are corrected, and the final optimized mesh is output.

[0038] The fifth step is to determine the optimized non-uniform mesh model set as the mesh model set corresponding to the above three-dimensional proximal femoral image set.

[0039] The aforementioned optional steps and related content, as an inventive point of this disclosure, solve the aforementioned technical problem of "insufficient accuracy and wasted computational resources in the constructed mesh model." Factors leading to this technical problem often include: uniform mesh models use the same element size throughout the model, making local optimization impossible 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 abrupt changes; and uniform mesh models maintain high-density partitioning in mechanically insensitive regions (e.g., the medullary canal in the middle of the bone shaft), resulting in a large amount of computational resources being used for insignificant areas. Solving these factors ensures improved mesh model accuracy and conserved computational resources when obtaining the mesh model. This solution uses intelligent mesh optimization to precisely allocate computational resources, matching the element density of high-stress and low-stress areas with actual mechanical requirements. Furthermore, it employs gradually changing element sizes and pyramidal transition elements in transition regions to ensure mechanical continuity, thereby improving mesh model accuracy and conserving computational resources.

[0040] The second step involves assigning material properties to each mesh model in the aforementioned mesh model set, resulting in a set of assigned 3D models. Material assignment imparts the mechanical properties of real bone tissue to the geometry of the mesh model, enabling the finite element model to simulate the stress and strain behavior of real bone. In practice, firstly, the gray density is determined based on the equivalent density. Then, based on the gray density, non-homogeneous linear elastic and nonlinear yielding material models used by other researchers are employed to describe the stress and strain relationships of each element. The bone material properties of the entire proximal femur are defined by 120 material sets, all with a Poisson's ratio set to 0.4.

[0041] In practice, ash density can be determined using the following formula: ,in, Indicates ash density, This represents the equivalent density.

[0042] The third step involves using finite element analysis (FEM) software to set the boundary conditions for each of the aforementioned working conditions in the 3D model set, thus obtaining a working condition configuration file. The FEM software can be ABAQUS 14.1. The working condition configuration file can be a set of parameters defining different working conditions in the FEM. Boundary conditions refer to the constraints and load settings that simulate the real physical environment in the FEM. These parameters can include: loading direction, displacement load, constraint settings, and bone material properties. In practice, the displacement load parameters for one standing working condition and three fall working conditions can be a 7mm displacement applied at 30mm from the femoral head. The constraint settings for one standing working condition can be set to completely fix the distal femur. When simulating the three fall working conditions, the following constraint is applied to the proximal femoral model: a constraint layer with a thickness of 6mm (equivalent to 2 elements) is set on the greater trochanter surface region opposite the femoral head loading area to resist the applied displacement load while allowing the model to move freely laterally. Furthermore, the elements in the model that include the femoral head loading zone nodes and the aforementioned greater trochanter constraint zone nodes are assigned specific material properties, namely an elastic modulus of 20 GPa and a yield strength of 200 MPa. The bone material properties of the remaining elements in the model are defined by another 120 sets of material parameters.

[0043] The fourth step involves generating a load-displacement curve set based on the aforementioned work condition configuration file and the target yield criterion. This set includes load-displacement curves for each work condition. The target yield criterion can be the Vos Mises yield criterion. In practice, based on the work condition configuration file and the Vos Mises yield criterion, the reaction force in the femoral head loading region for each increment can be obtained, thus deriving the load-displacement curves for the proximal femur. This load-displacement curve set represents a collection of mechanical responses under different work conditions from the finite element analysis. Each load-displacement curve includes load and 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 of the femoral head loading region. For example, the standing posture curve shows an ultimate strength of 7510 N, while the posterior impact curve of the greater trochanter shows a strength reduced to 3500 N.

[0044] The fifth step involves extracting femoral mechanical parameters from the aforementioned load-displacement curve set to generate femoral mechanical parameters corresponding to each working condition, resulting in a labeled dataset. These femoral mechanical parameters include ultimate strength, yield strength, and failure energy. Ultimate strength is defined as the maximum total reaction force on the loading region of the femoral head. Yield strength is defined as the load at which at least one solid element yields. Failure energy is defined as the area under the load-displacement curve at the point of maximum total reaction force. In practice, the ultimate strength, yield strength, and failure energy extracted from the four working conditions can be integrated into a labeled dataset.

[0045] Step 103: The image preprocessing end is configured to generate a three-dimensional proximal femoral image set corresponding to the first training set and the first test set.

[0046] In some embodiments, the image preprocessing unit is configured to generate a first training set and a first test set corresponding to the three-dimensional proximal femoral image set. The first training set is used to train an initial condition parameter prediction model, and the first test set is used to validate the condition parameter prediction model.

[0047] In some optional implementations of certain embodiments, the image preprocessing terminal can 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:

[0048] The first step is to preprocess the aforementioned three-dimensional proximal femoral image set to obtain a preprocessed image set. In practice, preprocessing the aforementioned three-dimensional proximal femoral image set can be performed by dividing each three-dimensional proximal femoral image in the set into 1×1×1mm segments. 3 The voxel sizes are resampled to obtain a preprocessed image set.

[0049] The second step involves two-dimensional projection and cropping of the preprocessed image set to obtain a two-dimensional projected image set. First, using a proximal femoral mask obtained through semi-automatic segmentation, the resampled three-dimensional image is projected in two dimensions in the anteroposterior view. This anteroposterior view can be a standard medical imaging orientation. Then, the two-dimensional projected image is cropped to 120×120 to ensure the entire proximal femur is included. The value of each pixel in the obtained two-dimensional projected image is obtained by summing the femoral pixel values ​​in both the anteroposterior and anterior views.

[0050] The third step is to divide the above-mentioned preprocessed image set and the above-mentioned two-dimensional projection image set into a first training set and a first test set.

[0051] Step 104: The model training end is configured to generate the image omics feature training set and deep learning feature training set corresponding to the first training set, and to generate the image omics feature test set and deep learning feature test set corresponding to the first test set.

[0052] In some embodiments, the model training end can be configured to generate a radiomics feature training set and a deep learning feature training set corresponding to a first training set, and to generate a radiomics feature test set and a deep learning feature test set corresponding to a first test set. The radiomics feature training set and the radiomics feature test set can be training sets and test sets that include radiomics features. In practice, the radiomics feature training set and test set can be divided into three groups according to size and shape features, intensity features, and texture features. The deep learning feature training set and the deep learning feature test set can be training sets and test sets that include deep learning features. In practice, deep learning features can include: edge and texture features (e.g., the edge contour of cortical bone and the microstructure texture of cancellous bone), grayscale distribution (e.g., the mean and variance of pixel intensity), and morphological features (e.g., the two-dimensional morphology of bone structures).

[0053] In some optional implementations of certain embodiments, the model training end can 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:

[0054] The first step involves extracting radiomics features from the preprocessed image subset corresponding to the first training set, resulting in a radiomics feature training set. The preprocessed image subset is the subset of the preprocessed image set corresponding to the first training set. In practice, the radiomics features in the training set can be extracted using the open-source Python package Pyradiomics.

[0055] The second step involves extracting radiomics features from the preprocessed image subset corresponding to the first test set, resulting in 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, the radiomics features in the above radiomics feature test set can be extracted using the open-source Python package Pyradiomics.

[0056] The third step is to divide the two-dimensional projected image set corresponding to the first training set into two-dimensional training and two-dimensional test sets. In practice, the ratio of the two-dimensional training set to the two-dimensional test set can be 4:1.

[0057] The fourth step involves normalizing and enhancing the aforementioned two-dimensional training set, as well as normalizing the aforementioned two-dimensional test set, to obtain normalized and enhanced two-dimensional training sets and normalized two-dimensional test sets. First, the two-dimensional projected images in the training set are subjected to Min-Max normalization to obtain a normalized two-dimensional training set. Then, each image in the normalized two-dimensional training set undergoes transformations such as horizontal and vertical shifting, horizontal flipping, and grayscale inversion to perform data enhancement, resulting in a normalized and enhanced two-dimensional test set. The method for normalizing the two-dimensional test set is the same as the method for processing the two-dimensional training set, and will not be repeated here.

[0058] The fifth step involves fine-tuning the pre-set initial deep feature extraction model using the normalized and enhanced 2D training set, the normalized 2D test set, the objective loss function, and the objective optimizer to obtain the final deep feature extraction model. This initial deep feature extraction model can be a modified ResNet50 model. The modified ResNet50 model comprises 49 convolutional layers and 1 fully connected layer, which can be divided into six modules. The first module includes one convolutional block and a max-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 removed, and the features from the penultimate layer (the global average pooling layer) are used as the output features.

[0059] Specifically, during training, the parameters of the first four modules of the deep feature extraction model are frozen, and only the parameters of the last residual module and the global average pooling layer are trained. The target optimizer can be the Adamax optimizer, with a learning rate of 0.0001, and then decayed by a decay factor of 0.1 at the 50th epoch. Eight training samples are randomly selected from the training set and eight validation samples are selected from the validation set in each batch, for a total of 100 training epochs. To overcome overfitting, L2 regularization is used, with a weight decay value of 0.01. The final model weights are determined based on minimizing the loss on the validation set after each epoch. The loss function can be the weighted multi-task mean squared error (MSE). All the fine-tuning was performed on a personal computer with an Intel Core i9-14900K CPU, 32GB of RAM, and an NVIDIA RTX 4070 Ti GPU.

[0060] In practice, the above loss function can be determined using the following formula:

[0061] ,in, This represents the mean square error between the femoral ultimate strength value predicted by the model and the femoral ultimate strength value calculated by the finite element method. This represents the mean square error between the femoral yield strength predicted by the model and the femoral yield strength calculated by the finite element method. This represents the mean square error between the failure energy predicted by the model and the failure energy calculated by the finite element method.

[0062] Step 6: Input the normalized and enhanced image training set corresponding to the first training set into the deep feature extraction model to obtain the 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 are consistent with the parameters determined in the deep learning feature training set.

[0063] Step 7: Normalize the two-dimensional projected image set corresponding to the first test set to obtain the normalized image test set. For specific steps, refer to the steps for normalizing the two-dimensional training set; they will not be repeated here.

[0064] Step 8: Input the normalized image test set into the deep feature extraction model to obtain the deep learning feature test set. The deep learning features in the deep learning feature test set are 2048-dimensional feature vectors.

[0065] Step 105: The model training end is configured to train the 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, thereby obtaining the working condition parameter prediction model.

[0066] In some embodiments, 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, thereby obtaining the working condition parameter prediction model. The clinical features corresponding to the first training set may include: height, weight, age, and bone mineral density. The loading direction may include... Value and value.

[0067] In some optional implementations of certain embodiments, the model training end can 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 the working condition parameter prediction model, which may include the following steps:

[0068] The first step involves preprocessing the aforementioned radiomics feature training set and deep learning feature training set to obtain processed radiomics feature training set and processed deep learning feature training set. In practice, firstly, feature variance filtering is applied to the aforementioned radiomics feature training set and deep learning feature training set to obtain the first radiomics feature training set and the first deep learning feature training set. For example, variance filtering in the Filter method can be used to remove low-information features with variance less than 0.4. Then, the Pearson correlation filtering algorithm is used to filter the features of the aforementioned first radiomics feature training set and the aforementioned first deep learning feature training set to obtain the second radiomics feature training set and the second deep learning feature training set. For example, the Pearson correlation filtering algorithm can be used to retain features with the highest Pearson correlation coefficient ranking (e.g., the top 50). Then, the minimum absolute shrinkage and selection operator regression algorithm is used to filter the features of the aforementioned second radiomics feature training set and the aforementioned second deep learning feature training set to obtain the third radiomics feature training set and the third deep learning feature training set. For example, the penalty coefficients in minimum absolute shrinkage and selection operator regression are adjusted using 5-fold cross-validation. Finally, multicollinearity processing is applied to the aforementioned third radiomics feature training set and the aforementioned third deep learning feature training set to obtain the processed radiomics feature training set and the 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 is considered redundant and is therefore removed. The radiomics features in the processed radiomics feature training set can be 20-dimensional features, and the deep learning features in the processed deep learning feature training set are also 20-dimensional features.

[0069] The second step involves fusing the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the object set of the first training set, and the corresponding loading direction set 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-dimensional), deep learning features (20-dimensional), clinical features (4-dimensional), and loading directions (2-dimensional) in the processed radiomics feature training set are then vertically concatenated to obtain the target features (46-dimensional) in the target feature training set.

[0070] However, when obtaining the target feature training set using the above method, the following technical problem often exists: "High-dimensional features lead to excessive computational load, resulting in wasted computational resources." The reason for this technical problem is that directly concatenating traditional high-dimensional features (46 dimensions) may lead to an excessive number of parameters in the working condition parameter prediction model and increase the unnecessary computational load. The technical challenge in obtaining the target feature training set is how to conserve computational resources. Therefore, the following solution can be adopted:

[0071] Optionally, the aforementioned executing entity can 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 of the first training set, and the corresponding loading direction set to obtain the target feature training set, which may include the following steps:

[0072] The first step involves concatenating the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the object set of the first training set, and the corresponding loading direction set to obtain the original feature matrix. This original feature matrix can be a 46-dimensional matrix obtained by vertically concatenating four types of standardized features, with each row corresponding to one sample and each column representing one feature, providing standardized input for subsequent graph structure construction. In practice, firstly, the processed radiomics feature training set, the processed deep learning feature training set, the clinical feature set corresponding to the object set of the first training set, and the corresponding loading direction set can be standardized (e.g., Z-score standardization) to obtain standardized processed radiomics feature training set, processed deep learning feature training set, clinical feature set corresponding to the object set of the first training set, and corresponding loading direction set. Then, the standardized processed radiomics feature training set, processed deep learning feature training set, clinical feature set corresponding to the object set of the first training set, and corresponding loading direction set are vertically concatenated to obtain the original feature matrix.

[0073] The second step is to determine the correlation coefficient matrix corresponding to the original feature matrix. This correlation coefficient matrix can be a 46×46 symmetric matrix obtained by determining the Pearson correlation coefficients between pairwise features in the original 46-dimensional feature matrix. For example, the elements of the correlation coefficient matrix... , representing the linear correlation strength between features i and j, with a diagonal value of 1 (representing autocorrelation) and off-diagonal values ​​reflecting cross-modal feature associations (e.g., the correlation between radiomics features and clinical features).

[0074] The third step is to filter the correlation coefficient matrix to obtain a filtered matrix. This filtered matrix can be a sparse matrix that retains elements (strongly correlated feature pairs) whose absolute values ​​in the correlation coefficient matrix have a preset threshold (e.g., 0.5), 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 constructing the graph structure, the dimension remains 46×46, but the number of non-zero elements is greatly reduced.

[0075] The fourth step is to deduplicate the filtered matrix to obtain a deduplicated matrix. This deduplicated matrix can be an undirected graph connectivity matrix obtained by eliminating symmetric duplicate edges from the filtered matrix. In practice, deduplication can be performed on the filtered matrix using methods such as deleting the symmetric edge (j, i) corresponding to edge (i, j).

[0076] The fifth step involves adding self-loop edges to the deduplicated matrix to obtain the edge index matrix. This edge index matrix can be the final edge connection table used in the pre-built fusion network model, and can be a two-dimensional array. The first row represents the source node number, and the second row represents the target node number, containing the filtered valid feature connections and necessary self-loop edges. In practice, firstly, isolated nodes (e.g., nodes with degree = 0) can be detected in the deduplicated matrix. Then, self-loop edges (i, i) are added to obtain the edge index matrix.

[0077] The sixth step involves encapsulating the aforementioned correlation coefficient matrix and edge index matrix to obtain graph data. This graph data may include node features and edge connections. In practice, the correlation coefficient matrix and edge index matrix can be encapsulated in PyG format to obtain the graph data.

[0078] Step 7: Input the above graph data into the pre-built fusion network model to obtain the target feature training set. Each target feature in the 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 may include convolutional layers, skip connection layers, and normalization layers.

[0079] The aforementioned optional steps and related content, as an inventive point of this disclosure, solve the aforementioned technical problem of "excessive computational load and waste of computing resources due to high-dimensional features." Factors leading to this problem often include: direct concatenation of traditional high-dimensional features (46 dimensions) may result in an excessive number of parameters in the working condition parameter prediction model and increase the load of ineffective computation. Solving these factors ensures that computing resources are saved when obtaining the target feature training set. This solution compresses the fused features from 46 dimensions to 8 dimensions using a pre-built fusion network model, significantly reducing the number of downstream parameters. Then, by filtering and removing irrelevant and weakly correlated edges in the correlation coefficient matrix, ineffective computation is reduced, thereby saving computing resources.

[0080] The third step involves randomly dividing the standardized feature training set and the standardized label dataset corresponding to the target feature training set and the corresponding label dataset into mutually exclusive subsets of the number of targets. In practice, firstly, the target feature training set is Z-score standardized to eliminate scale sensitivity in different features, resulting in a standardized feature training set. Similarly, the label dataset corresponding to the target feature training set is Z-score standardized to obtain a standardized label dataset. The target feature training set and the corresponding label dataset are subsets of the label dataset. Each label dataset includes three femoral mechanical parameters: ultimate femoral strength, yield strength, and failure energy. Then, the standardized feature training set and the corresponding standardized label dataset are randomly divided into mutually exclusive subsets of the number of targets. For example, the number of targets can be 5. The mutually exclusive subsets can be 5 non-overlapping subsets of similar size.

[0081] The fourth step is to determine the parameter range set corresponding to each hyperparameter in the hyperparameter set of the initial working condition parameter prediction model, thus obtaining the parameter range set. The hyperparameter set includes the learning rate parameter, weight decay parameter, random inactivation rate parameter, and batch size parameter. The parameter range set corresponding to the learning rate parameter is as follows: (Select according to order of magnitude). The parameter range group corresponding to the weight decay parameter is as follows: (Select according to order of magnitude). The parameter range corresponding to the random inactivation rate parameter is: The batch size parameter can be 8, 16, or 32.

[0082] Step 6: Select parameter ranges from each parameter range group in the above parameter range set to generate parameter range combinations, resulting in a parameter range combination set. Each parameter range in the combination corresponds one-to-one with a hyperparameter in the hyperparameter set. For example, the parameter range combination set could have: a learning rate of 0.01, a weight decay of 0.01, a random inactivation rate of 0.1, and a batch size of 16. Alternatively, the parameter range combination set could have: a learning rate of 0.001, a weight decay of 0.001, a random inactivation rate of 0.3, and a batch size of 16.

[0083] The fifth step involves randomly dividing the standardized feature training set and the corresponding standardized label dataset into mutually exclusive subsets of the number of targets. For example, the number of targets could be 5. These mutually exclusive subsets could be 5 non-overlapping subsets of similar size.

[0084] The sixth step is to use the mutually exclusive subsets mentioned above to perform cross-validation on each hyperparameter range combination in the parameter range combination set to obtain the target hyperparameter combination.

[0085] Specifically, firstly, the mutually exclusive subsets can be divided into training and validation subsets. In practice, five mutually exclusive subsets can be used for five iterations, with four subsets selected as the training subset and one subset as the validation subset in each iteration. The average weighted multi-task mean squared error obtained from the five iterations is determined as the validation result set. Then, the hyperparameter combination corresponding to the minimum validation result in the validation result set is determined as the target hyperparameter combination. The minimum validation result corresponds to the optimal hyperparameter combination, which is then used as the target hyperparameter combination. In practice, the target hyperparameter combination can be a learning rate parameter of 1e-4, a weight decay parameter of 1e-5, and a batch size parameter of 32.

[0086] Step 7: Based on the standardized feature training set and the standardized label dataset mentioned above, train the initial operating condition parameter prediction model corresponding to the above target hyperparameter combination using the target optimization algorithm and target loss function to obtain the operating condition parameter prediction model. First, divide the standardized feature training set 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, train the operating condition parameter model using the above 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 the Adamax optimization algorithm, and the target loss function can be the 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 5 consecutive rounds. The loss function mentioned above is the loss determined after each training round based on the standard validation set and the corresponding standardized label dataset.

[0087] In practice, the weighted multi-task mean square error can be obtained using the following formula:

[0088] ,in, This represents the mean square error between the femoral ultimate strength value predicted by the model and the femoral ultimate strength value calculated by the finite element method. This represents the mean square error between the femoral yield strength predicted by the model and the femoral yield strength calculated by the finite element method. This represents the mean square error between the failure energy predicted by the model and the failure energy calculated by the finite element method.

[0089] As an example, 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 are first preprocessed to obtain the preprocessed radiomics feature training set, the preprocessed deep learning feature training set, the preprocessed clinical feature information set, and the preprocessed loading direction set. Then, these preprocessed sets are concatenated to obtain a unified feature vector set. Next, a Transformer encoder is used to fuse the feature vectors in the unified feature vector set to obtain a fused feature vector set. Finally, based on the fused feature vector set and the corresponding label dataset, a weighted multi-task loss function is used to train an initial operating condition parameter prediction model, resulting in the operating condition parameter prediction model. The initial operating condition parameter prediction model can be a Transformer model.

[0090] Step 106: The model application is configured to generate a multi-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, using the condition parameter prediction model.

[0091] In some embodiments, the model application is configured to generate a multi-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, using the above-mentioned working condition parameter prediction model. The multi-condition parameter prediction result set can be the three mechanical parameters (ultimate strength, yield strength, and failure energy) predicted for each sample in the first test set under four working conditions.

[0092] In some optional implementations of certain embodiments, the model application can generate a multi-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, using the above-mentioned condition parameter prediction model. This can include the following steps:

[0093] The first step involves preprocessing the aforementioned radiomics feature test set and deep learning feature test set to obtain the processed radiomics feature test set and the processed deep learning feature test set. For specific methods, please refer to the section on generating the processed radiomics feature training set and the processed deep learning feature training set, which will not be repeated here.

[0094] The second step involves fusing the processed radiomics feature test set, the processed deep learning feature test set, the clinical feature set corresponding to the object set of the first test set, and the corresponding loading direction set to obtain the target feature test set. For details on the generation of the target feature training set, please refer to [link to documentation].

[0095] The third step involves inputting the aforementioned target feature test set into the aforementioned working condition parameter prediction model to obtain a multi-working condition parameter prediction result set. Each multi-working condition prediction result in the set includes the three predicted femoral biomechanical parameters for the corresponding sample under various working conditions. For example, for the standing working condition, the three predicted femoral biomechanical parameters are 7510N, 3500N, and 16J.

[0096] 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 can be used to generate a multi-working condition parameter prediction result set, which may also include the following steps:

[0097] The first step involves preprocessing the aforementioned radiomics feature test set, deep learning feature test set, and the clinical feature information set corresponding to the first test set to obtain a standardized radiomics feature test set, a standardized deep learning feature test set, and a standardized clinical feature information set. Additionally, the loading direction set is encoded to obtain an encoded loading direction set. In practice, the mean and standard deviation of the aforementioned radiomics feature test set can be Min-Max normalized to obtain a standardized radiomics feature test set. Similarly, the mean and standard deviation of the aforementioned deep learning feature test set can be Z-score standardized to obtain a standardized deep learning feature test set. Furthermore, the clinical feature information set corresponding to the first test set can be Z-score standardized to obtain a standardized clinical feature information set. Finally, the 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 (e.g., for standing posture conditions...). , The one-hot vector of is [0, 1, 0, 0, 1, 0, 0]).

[0098] The second step involves inputting the standardized image omics feature test set and the corresponding label dataset into the first prediction model to obtain the 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 under one standing posture and three fall conditions for the corresponding sample.

[0099] The third step involves inputting the standardized deep learning feature test set and the corresponding label dataset into the second prediction model to obtain the second prediction result set. In practice, the second prediction model can be an XGBoost regression model.

[0100] The fourth step involves inputting the standardized clinical feature information set, the encoded loading direction set, and the corresponding label dataset into the third prediction model to obtain the third prediction result set. In practice, the third prediction model can be a multilayer perceptron prediction model.

[0101] The fifth step involves weighted integration of the first, third, and fourth prediction result sets to obtain a multi-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.

[0102] In some optional implementations, the execution entity can 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:

[0103] The first step involves the structure of the working condition parameter prediction model, which includes a shared feature extraction module and independent prediction modules for the number of targets. The shared feature extraction module comprises a residual connection layer, a hierarchical normalization layer, and a dynamic deactivation layer. The independent prediction modules for the number of targets comprise a limit strength prediction layer, a yield strength prediction layer, and a failure energy prediction layer. The limit strength prediction layer, used to predict the limit strength parameter, may include an input layer, a fully connected layer, and an output layer. The yield strength prediction layer, used to predict the yield strength parameter, may include an input layer, an attention mechanism module, a piecewise regression head, and a loss function. The failure energy prediction layer, used to predict the failure energy parameter, may include an input layer, an LSTM layer, a fully connected layer, and a loss function.

[0104] The second step involves inputting the aforementioned target feature test set into the shared feature extraction module to obtain a deep shared feature set. This shared feature extraction module can fuse radiomics features, deep learning features, and clinical features to extract universal features across tasks. Specifically, the target feature training set is sequentially passed through a residual connection layer, a hierarchical normalization layer, and a dynamic inactivation layer to obtain the deep shared feature set.

[0105] The third step involves inputting the aforementioned deep shared features into the aforementioned extreme intensity prediction layer to obtain the extreme intensity prediction parameter set. This extreme intensity prediction parameter set includes the extreme intensity prediction parameters for each sample under multiple operating conditions in the target feature test set.

[0106] The fourth step involves inputting the aforementioned deep shared features into the yield strength prediction layer to obtain the yield strength prediction parameter set. This yield strength prediction parameter set includes the yield strength prediction parameters for each sample under multiple working conditions in the target feature test set.

[0107] The fifth step involves inputting the aforementioned deep shared features into the failure energy prediction layer to obtain a failure energy prediction parameter set. This failure energy prediction parameter set includes failure energy prediction parameters for each sample under multiple operating conditions in the target feature test set.

[0108] The sixth step is to determine the above-mentioned ultimate strength prediction parameter set, the above-mentioned yield strength prediction parameter set, and the above-mentioned failure energy set prediction parameters as the multi-condition parameter prediction result set.

[0109] Optionally, the aforementioned implementing entity may also perform the following steps:

[0110] The first step involves performing interpretability analysis on the predicted result set of the multi-condition parameters and the test feature information set corresponding to the first test set to obtain the global analysis results. In practice, the SHAP (Shapley AdditiveExplanations) method can be used to determine the global feature importance and generate a global importance ranking of key features. For example, in radiomics, the gray-level co-occurrence matrix entropy value has the greatest impact on the limiting intensity.

[0111] The second step involves performing interpretability analysis on the samples corresponding to each test feature in the aforementioned test feature information set to obtain a summary of local analysis results. In practice, for each sample, the SHAP value can be used to calculate the local contribution of each feature in its prediction result, and the distribution of local feature contributions for all samples (e.g., high-frequency key features and feature patterns of abnormal samples) can be statistically analyzed to obtain a summary of local analysis results. For example, a patient's low limit intensity may be mainly due to abnormal grayscale features.

[0112] The third step involves filtering key features from the overall global analysis results and the summarized local analysis results to obtain a list of key features. Specifically, statistical methods can be used to filter high-contribution, high-frequency features. In practice, these statistical methods can be variance filtering or Pearson correlation coefficient methods. The list of key features can include grayscale entropy from radiomics, texture features from deep learning, and bone mineral density from clinical applications.

[0113] The fourth step involves generating a medical interpretation table based on the aforementioned list of key features using a medical knowledge base. In practice, this medical knowledge base can be a collection of relevant literature or guidelines that link the features to bone mechanics and diseases. The medical interpretation table can present the feature name, medical interpretation, and associated disease risk level.

[0114] The fifth step is to generate a key feature report using the aforementioned 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 predicted ultimate strength value is low due to advanced age (70 years old) and low bone mineral density (T-score -2.5).

[0115] Step 6: Using the aforementioned key feature reports, integrate the global analysis results and the summary results of the local analysis to obtain a visualization chart. In practice, the global view (e.g., a waterfall chart showing the total impact of each feature on the prediction), the local view (e.g., a scatter plot matrix showing the correlation between features and predicted values), and the medical view (e.g., a heatmap) can be combined with medical interpretations and the feature importance and key feature reports to generate a visualization chart.

[0116] The various embodiments of this disclosure have the following beneficial effects: Through a machine learning-based multi-condition femoral biomechanical parameter generation method according to some embodiments of this disclosure, multiple femoral biomechanical parameters under various working conditions can be predicted quickly and accurately. Furthermore, interpretability analysis methods are used to explain the decision-making process of the prediction model, further clarifying the most important features in the prediction process of these femoral biomechanical parameters. Specifically, the reasons for the above-mentioned technical problems are: existing methods usually use indirect measurement, resulting in low measurement accuracy. Existing methods typically extract image features manually, which is time-consuming and has a low feature repetition rate. In addition, a large number of image features reflecting bone material distribution and bone microstructure have not been fully explored, and existing prediction models are usually black-box models, with opaque prediction processes. Based on this, the machine learning-based multi-condition femoral biomechanical parameter generation method of some embodiments of this disclosure first configures the image preprocessing end of the image processing system to segment the region of interest (ROI) of the original proximal femoral image set to obtain a three-dimensional proximal femoral image set. The image processing system also 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 3D reconstruction preserves the spatial information of the bone structure, facilitating feature quantification. Then, the image preprocessing unit is configured to perform multi-condition finite element modeling on the aforementioned 3D proximal femoral image set, obtaining a label dataset under multiple conditions. Each condition is represented by adjusting the loading direction, which is the directional parameter of the external force applied to the femur. The generated label dataset provides high-quality training samples for the deep learning model, enhancing its generalization ability. Next, the image preprocessing unit is configured to generate a first training set and a first test set corresponding to the aforementioned 3D proximal femoral image set. Independent processing of the training and test sets avoids interference from test set information during the training process. Furthermore, the model training unit is configured to generate a radiomics feature training set and a deep learning feature training set corresponding to the first training set, as well as a radiomics feature test set and a deep learning feature test set corresponding to the first test set. Multimodal feature input reduces the influence of single-data noise, enhancing generalization ability. Independent generation of training / test set features ensures the objectivity of model evaluation. Secondly, the model training end is configured to train an initial operating condition parameter prediction model based on the aforementioned radiomics feature training set, the aforementioned deep learning feature training set, the clinical feature information set corresponding to the aforementioned first training set, and the corresponding loading direction set, thus obtaining the operating condition parameter prediction model. The initial model structure, trained jointly with multi-source data, can accelerate convergence and improve prediction accuracy. Incorporating clinical features and loading direction as inputs to the model can enhance its robustness. Finally, the model application end is configured to generate a multi-operating condition parameter prediction result set using the aforementioned operating condition parameter prediction model based on the aforementioned radiomics feature test set, the aforementioned deep learning feature test set, the clinical feature information set corresponding to the aforementioned first test set, and the corresponding loading direction set.Independently validating model performance using a test set reduces the risk of misjudgment. Furthermore, by leveraging the collaborative work of the image preprocessing, model training, and model application ends within the image processing system, the core technical objective of rapidly and accurately predicting femoral biomechanical parameters under various working conditions was achieved.

[0117] The following is for reference. Figure 2 It shows a schematic diagram of the structure of an electronic device (e.g., an electronic device) 200 suitable for implementing some embodiments of the present disclosure. Figure 2 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0118] 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 a program stored in read-only memory (ROM) 202 or a program loaded from storage device 208 into random access memory (RAM) 203. The RAM 203 also stores various programs and data required for the operation of electronic device 200. The processing device 201, ROM 202, and RAM 203 are interconnected via bus 204. Input / output (I / O) interface 205 is also connected to bus 204.

[0119] Typically, the following devices can be connected to I / O interface 205: input devices 206 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 207 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 208 including, for example, magnetic tapes, hard disks, etc.; and communication devices 209. Communication device 209 allows electronic device 200 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 2 An electronic device 200 with various devices is shown; however, 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 alternatively. Figure 2 Each box shown can represent a device or multiple devices as needed.

[0120] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 209, or installed from storage device 208, or installed from ROM 202. When the computer program is executed by processing device 201, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0121] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, 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, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this 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, apparatus, or device. In some embodiments of this 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. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0122] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0123] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: segment the region of interest (ROI) of the original proximal femoral image set to obtain a three-dimensional proximal femoral image set; perform multi-condition finite element modeling on the aforementioned three-dimensional proximal femoral image set to obtain a label dataset under multiple conditions, where each condition is represented by adjusting the loading direction, the loading direction being a directional parameter for applying external force to the femur; generate a first training set and a first test set corresponding to the aforementioned three-dimensional proximal femoral image set; 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; Based on the aforementioned radiomics feature training set, the aforementioned deep learning feature training set, the clinical feature information set corresponding to the aforementioned first training set, and the corresponding loading direction set, an initial operating condition parameter prediction model is trained to obtain the operating condition parameter prediction model; based on the aforementioned radiomics feature test set, the aforementioned deep learning feature test set, the clinical feature information set corresponding to the aforementioned first test set, and the corresponding loading direction set, the operating condition parameter prediction model is used to generate a multi-operating condition parameter prediction result set.

[0124] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone 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 remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0125] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or a combination of blocks in a block diagram and / or a flowchart may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0126] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0127] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for generating femoral biomechanical parameters under multiple working conditions 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 also includes a model training end and a model application end. The image preprocessing unit is configured to perform multi-condition finite element modeling on the three-dimensional proximal femur image set to obtain a labeled dataset under multiple conditions. Each condition is represented by adjusting the loading direction, which is a directional parameter for applying external force to the femur. The preprocessing includes: generating a mesh model set corresponding to the three-dimensional proximal femur image set; obtaining a grayscale image set corresponding to the mesh model set; establishing a correspondence between the mesh model set and the grayscale image set to obtain an enhanced mesh dataset with grayscale attributes; inputting the enhanced mesh dataset into a pre-trained key region recognition model to obtain a stress heat map; and using the stress heat map to analyze the mesh model. The lattice model set undergoes dynamic mesh optimization to obtain an optimized non-uniform lattice model set; this optimized non-uniform lattice model set is then defined as the lattice model set corresponding to the three-dimensional proximal femoral image set; material values ​​are assigned to each lattice model in the lattice model set to obtain an assigned three-dimensional model set; boundary conditions for each working condition in the assigned three-dimensional model set are set using finite element analysis software to obtain a working condition configuration file; a load-displacement curve set is generated based on the working condition configuration file using the target yield criterion; femoral mechanical parameters are extracted from the load-displacement curve set to generate femoral mechanical parameters corresponding to each working condition, thus obtaining a labeled dataset; The image preprocessing end is configured to generate the three-dimensional proximal femur image set corresponding to the first training set and the first test 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, thereby obtaining the working condition parameter prediction model. The model application is configured to generate a multi-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, using the condition parameter prediction model.

2. The method according to claim 1, characterized in that, The generation of the three-dimensional proximal femoral image set corresponding to the first training set and the first test set includes: The three-dimensional proximal femur image set is preprocessed to obtain a preprocessed image set; The preprocessed image set is subjected to two-dimensional projection to obtain a two-dimensional projected 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.

3. The method according to claim 1, characterized in that, The generation of the image omics feature training set and deep learning feature training set corresponding to the first training set, and the generation of the image omics feature test set and deep learning feature test set corresponding to the first test set, include: Image omics features are extracted from the preprocessed image subset corresponding to the first training set to obtain the image omics feature training set. Image omics features are extracted from the preprocessed image subset corresponding to the first test set to obtain the image omics feature test set. The two-dimensional projected image set corresponding to the first training set is divided to obtain a two-dimensional training set and a two-dimensional test set. The two-dimensional training set is normalized and enhanced, and the two-dimensional test set is normalized to obtain a normalized and enhanced two-dimensional training set and a normalized two-dimensional test set. The pre-set initial depth feature extraction model is fine-tuned using the normalized and enhanced two-dimensional training set, the objective loss function, and the objective optimizer to obtain the depth feature extraction model. Input the normalized and enhanced image training set corresponding to the first training set into the deep feature information extraction model to obtain the deep learning feature training set; The two-dimensional projected image set corresponding to the first test set is normalized to obtain the normalized image test set. The normalized image test set is input into the deep feature information extraction model to obtain the deep learning feature test set.

4. The method according to claim 1, characterized in that, The step of training an initial load 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 load condition parameter prediction model includes: The image omics feature training set and the deep learning feature training set are preprocessed to obtain the processed image omics feature training set and the processed deep learning feature training set. 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 are fused to obtain the target feature training set. The standardized feature training set and the standardized label dataset corresponding to the target feature training set and the corresponding label dataset are randomly divided into mutually exclusive subsets of the number of targets. Determine the parameter range group corresponding to each hyperparameter of the hyperparameter set corresponding to the initial working condition parameter prediction model to obtain the parameter range group set; Parameter ranges are selected from each parameter range group in the parameter range group set to generate parameter range combinations, thus obtaining a parameter range combination set; The mutual exclusion subset is used to perform cross-validation on each parameter range combination in the parameter range combination set to obtain the target hyperparameter combination; Based on the standardized feature training set and the standardized label dataset, the initial working condition parameter prediction model corresponding to the target hyperparameter combination is trained using the target optimization algorithm and the target loss function to obtain the working condition parameter prediction model.

5. The method according to claim 1, characterized in that, The step of generating a multi-condition parameter prediction result set using the 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 includes: The image omics feature test set and the deep learning feature test set are preprocessed to obtain the processed image omics feature test set and the processed deep learning feature test set. 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 are fused to obtain the target feature test set. The target feature test set is input into the working condition parameter prediction model to obtain a multi-working condition parameter prediction result set.

6. The method according to claim 5, 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 number of targets. The shared feature extraction module includes: a residual connection layer, a hierarchical normalization layer, and a dynamic deactivation layer. The independent prediction module for the number of targets includes: an ultimate strength prediction layer, a yield strength prediction layer, and a failure energy prediction layer. The step of inputting the target feature test set into the working condition parameter prediction model to obtain a multi-working condition parameter prediction result set includes: The target feature test set is input into the shared feature extraction module to obtain the deep shared feature set; The deep shared features are input into the ultimate intensity prediction layer to obtain the ultimate intensity prediction parameter set; The deep shared features are input into the yield strength prediction layer to obtain the yield strength prediction parameter set; The deep shared features are input into the failure energy prediction layer to obtain the 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-condition parameter prediction result set.

7. An electronic device, characterized in that, include: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

8. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images