Medical data processing method and device for bone nonunion healing analysis, equipment and medium

Abstract

This application discloses a medical data processing method, apparatus, equipment, and medium for nonunion healing analysis, relating to the field of medical and health information technology. The method includes: quantitatively processing CT images and biochemical indicators of the target area to obtain fracture gap parameters and microenvironment biological activity scores; inputting the gap parameters, activity scores, and implant information into a multidimensional etiology association model to calculate etiology weight vectors; extracting visual morphological feature vectors from CT images at multiple scales using a three-dimensional image network; constructing and simulating a three-dimensional finite element model based on CT images and implant information to obtain biomechanical feature vectors; weighting and fusing visual and mechanical features according to the etiology weight vectors to obtain cross-modal collaborative features; and processing the data through a multi-task prediction network to obtain scoring auxiliary data and dynamic healing prediction curves. This application can obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of nonunion and the prediction of dynamic healing trends.

Description

Technical Field

[0001] This application relates to the field of medical and health information technology, and in particular to medical data processing methods, devices, equipment and media for bone nonunion healing analysis. Background Technology

[0002] Nonunion is a serious complication after fracture surgery, and its accurate pathological classification and pathogenic mechanism attribution directly determine the choice of intervention strategy and patient prognosis. In the evolution of nonunion, pathological features usually involve the three-dimensional anatomical morphology of the fracture ends, the biomechanical stability of the internal fixation implant, as well as the patient's own microenvironment biological activity and underlying risk factors.

[0003] Currently, the processing and analysis of data related to bone nonunion mainly relies on a variety of independent technical methods. In morphological data processing, this primarily involves professional physicians manually interpreting CT (Computed Tomography) images to extract geometric features, or using single-modality deep learning models such as CNN (Convolutional Neural Network) to extract visual image features. For biomechanical data acquisition, traditional FEA (Finite Element Analysis) techniques are typically employed. These techniques involve complex mesh generation and boundary condition settings on digital bone models to simulate and calculate the stress distribution and micro-displacement field of fracture ends and implants under specific loads. For patients' clinical metabolic and disease risk data, statistical risk scoring scales are often used for separate numerical quantification and qualitative recording.

[0004] The existing processing methods described above have significant technical limitations when dealing with complex medical data on bone nonunion. Single-dimensional visual morphological analysis for feature extraction is limited, and purely imaging features often lag behind changes in the microscopic biomechanical environment, failing to capture early pathological changes caused by mechanical instability. Traditional finite element analysis is extremely cumbersome; complex mesh generation and nonlinear iterative solutions typically consume significant computational resources and time, resulting in low efficiency in extracting key mechanical feature data. More critically, existing intelligent medical data processing methods often treat imaging visual features, mechanical simulation parameters, and clinical physiological features in isolation. Therefore, how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of bone nonunion and the prediction of dynamic healing trends has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a medical data processing method, device, equipment and medium for nonunion healing analysis, aiming to solve the technical problem of how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of nonunion and the prediction of dynamic healing trends.

[0006] To achieve the above objectives, this application proposes a medical data processing method for analyzing bone nonunion healing, the method comprising: The computed tomography images of the target area and the biochemical index data of the subjects were quantified to obtain the fragment gap parameters and the biological activity score of the microenvironment. The fracture gap parameters, the microenvironment biological activity score, and the implant information are input into a pre-trained multidimensional etiology association model for evaluation and calculation to obtain an etiology weight vector. The computed tomography images are subjected to multi-scale feature extraction using a three-dimensional image feature extraction network to obtain visual morphological feature vectors. A three-dimensional finite element model is established based on the computed tomography images and the implant information, and biomechanical simulation analysis is performed to obtain biomechanical feature vectors. The visual morphological feature vector and the biomechanical feature vector are weighted and fused according to the etiological weight vector to obtain cross-modal collaborative features; The cross-modal collaborative features are processed by a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

[0007] Furthermore, to achieve the above objectives, this application also proposes a medical data processing device for analyzing bone nonunion healing, the device comprising: The data quantification module is used to quantify the computed tomography images of the target area and the biochemical index data of the subjects to obtain the fragment gap parameters and the biological activity score of the microenvironment. The etiology assessment module is used to input the fracture gap parameters, the microenvironment biological activity score and implant information into a pre-trained multidimensional etiology association model for evaluation and calculation, and to obtain the etiology weight vector. The visual extraction module is used to extract multi-scale features from the computed tomography images through a three-dimensional image feature extraction network to obtain visual morphological feature vectors. The mechanical simulation module is used to establish a three-dimensional finite element model based on the computed tomography images and the implant information, and to perform biomechanical simulation analysis to obtain biomechanical feature vectors. The feature fusion module is used to perform weighted fusion of the visual morphological feature vector and the biomechanical feature vector according to the etiology weight vector to obtain cross-modal collaborative features; The analysis and prediction module is used to process the cross-modal collaborative features through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

[0008] Furthermore, to achieve the above objectives, this application also proposes a medical data processing device for analyzing nonunion healing, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the medical data processing method for analyzing nonunion healing as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the medical data processing method for bone nonunion healing analysis as described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the medical data processing method for bone nonunion healing analysis as described above.

[0011] One or more technical solutions proposed in this application have at least the following technical effects: First, by quantifying computed tomography (CT) images of the target area and biochemical data of the subjects, parameters of the fracture gap and microenvironment biological activity scores were obtained, achieving standardized representation of medical images and physiological indicators. Subsequently, the fracture gap parameters, microenvironment biological activity scores, and implant information were input into a pre-trained multidimensional etiology association model for evaluation and calculation, resulting in etiology weight vectors that quantified the proportion of different pathogenic factors affecting nonunion. Simultaneously, a three-dimensional image feature extraction network was used to extract multi-scale features from the CT images, obtaining visual morphological feature vectors and deeply mining the visual features of the fracture spatial structure. Then, a three-dimensional model was established based on the CT images and implant information. A finite element method (FEM) model was used to perform biomechanical simulation analysis, resulting in biomechanical feature vectors and quantitative physical indicators reflecting the stability of the internal fixation system. Based on this, the visual morphological feature vectors and biomechanical feature vectors were weighted and fused according to the etiology weight vector to obtain cross-modal collaborative features. This achieved adaptive fusion of imaging and mechanical features and suppressed feature conflicts between modalities. Finally, the cross-modal collaborative features were processed through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves, simultaneously outputting pathological judgment results and temporal prognostic trends. This application can obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the subtyping assessment of bone nonunion and the prediction of dynamic healing trends. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a medical data processing method for analyzing nonunion fractures according to this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the medical data processing method for analyzing bone nonunion healing in this application. Figure 3 This is a schematic diagram of the module structure of a medical data processing device for analyzing bone nonunion healing according to an embodiment of this application. Figure 4 This is a schematic diagram of the hardware operating environment of the medical data processing method for analyzing bone nonunion healing in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0017] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0018] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or medical data processing system capable of realizing the above functions. The following uses a medical data processing system as an example to describe this embodiment and the subsequent embodiments.

[0019] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0020] Based on this, embodiments of this application provide a medical data processing method for analyzing bone nonunion healing, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the medical data processing method for analyzing bone nonunion healing according to this application.

[0021] In this embodiment, the medical data processing method for analyzing bone nonunion healing includes steps S10 to S60: Step S10: Quantify the computed tomography images of the target area and the biochemical index data of the subject to obtain the fragment gap parameters and the microenvironment biological activity score. It should be noted that the computed tomography (CT) images of the target area refer to 3D CT images including the fracture ends and surrounding tissues. The subject's biochemical data can be blood test results containing bone metabolism markers (such as type I procollagen N-terminal propeptide) and inflammatory markers (such as C-reactive protein). The fracture gap parameter refers to the geometric characteristics that quantify the degree of separation and contact morphology between the fracture ends. The microenvironment biological activity score is a comprehensive quantitative score reflecting the local blood supply and osteogenic capacity of the fracture site.

[0022] Understandably, image segmentation is performed on the computed tomography images of the target area to extract the boundary of the medullary cavity, and the minimum Euclidean distance and contact surface area ratio between the two ends of the fracture are calculated to generate the fracture gap parameters. The distance measurement range involved in the calculation is set to 0.1 mm to 10 mm based on the clinically routine bone defect situation. A multiple linear regression model is used to assign weights to type I procollagen N-terminal propeptide and C-reactive protein in the subjects' biochemical index data and then perform weighted summation to obtain a microenvironment biological activity score in the range of 0 to 1. The weights of each index are set according to their correlation coefficient with bone healing. Type I procollagen N-terminal propeptide is used as a positive index, with a weight range of 0.4 to 0.6; C-reactive protein is used as a negative index, with a weight range of -0.3 to -0.1. This step transforms unstructured image data and multidimensional clinical biochemical results into unified numerical values, providing a measurable and objective data basis for subsequent calculations.

[0023] Step S20: Input the fracture gap parameter, the microenvironment biological activity score and the implant information into a pre-trained multidimensional etiology association model for evaluation and calculation to obtain the etiology weight vector; It should be noted that implant information can include the three-dimensional geometric parameters and material elastic modulus data of intramedullary nails or plates. A multidimensional etiological association model refers to a model constructed using principal component analysis (PCA) to extract the weights of principal pathogenic features. The etiological weight vector is a one-dimensional feature vector reflecting the proportion of influence of biological and mechanical factors in bone nonunion formation.

[0024] Understandably, the fracture gap parameters, microenvironment biological activity score, and implant information are concatenated to form a comprehensive feature matrix. This comprehensive feature matrix is ​​then input into a multidimensional etiology association model to calculate the eigenvalues ​​and eigenvectors of the covariance matrix. Eigenvectors with a cumulative variance contribution rate greater than a preset threshold (empirically set at 85%) are extracted as dominant risk components. Based on these dominant risk components, the contribution ratio of each original input dimension is calculated, normalized, and then the etiology weight vector is output. This step achieves dimensionality reduction of multi-source data and extracts the dominant association factors leading to bone nonunion, quantifying the influence of different etiologies.

[0025] Step S30: Multi-scale feature extraction is performed on the computed tomography image using a three-dimensional image feature extraction network to obtain a visual morphological feature vector. It should be noted that the 3D image feature extraction network can be a deep neural network containing 3D convolutional layers (3D CNN) and pooling layers. Multi-scale feature extraction refers to the feature extraction process that captures local details and macroscopic structures under receptive fields of different sizes. Visual morphological feature vectors refer to high-dimensional feature representations that reflect the details of callus texture and the distribution of fracture line orientation.

[0026] The process involves converting computed tomography images into three-dimensional tensors with a fixed spatial resolution (size set to 128×128×128); inputting these three-dimensional tensors into a three-dimensional image feature extraction network; extracting local callus texture features using 3×3×3 three-dimensional convolutional kernels; performing downsampling through max pooling layers to expand the receptive field and extract macroscopic morphological features of the fracture line; and flattening and stitching the features at each scale using global average pooling layers to output a visual morphological feature vector. This step directly extracts the implicit image features of the fracture ends in three-dimensional space, preserving the spatial distribution information of the structure.

[0027] Step S40: Establish a three-dimensional finite element model based on the computed tomography images and the implant information, and perform biomechanical simulation analysis to obtain biomechanical feature vectors; It should be noted that a three-dimensional finite element model refers to a physical model in which the continuous geometry of the bone and implant is discretized into a finite number of mesh elements for numerical calculation. Biomechanical simulation analysis refers to the computational process of obtaining stress and displacement distribution by solving static equilibrium equations. A biomechanical eigenvector is an eigenvector containing mechanical response parameters such as peak stress tensor and fretting displacement values.

[0028] Understandably, the process involves generating a skeletal mesh model by generating tetrahedral meshes from computed tomography (CT) images; meshing the computer-aided design (CAD) geometric model of the implant information and registering it with the skeletal mesh model using spatial Boolean operations to form an assembly model; calculating bone density based on image grayscale values ​​and assigning material properties to each mesh unit according to known empirical conversion formulas between bone density and elastic modulus; applying a compressive load (600N to 800N based on standard human weight) and fixing constraints along the lower limb force line on the assembly model; solving the static equilibrium equations using a finite element method (FEM) to extract the micro-displacement values ​​and maximum Von Mises stress values ​​of the fracture end region, and integrating them to generate a biomechanical feature vector. This step transforms static anatomical images into dynamic mechanical response parameters, enabling a quantitative assessment of the mechanical stability performance of the internal fixation system.

[0029] Step S50: The visual morphological feature vector and the biomechanical feature vector are weighted and fused according to the etiological weight vector to obtain cross-modal collaborative features; It should be noted that cross-modal collaborative features refer to a comprehensive high-dimensional feature matrix that integrates imaging visual representations and mechanical simulation parameters.

[0030] Understandably, the process involves separating visual and biomechanical weight coefficients from the etiological weight vector. A dot product is then performed between the visual weight coefficient and the visual morphological feature vector, and between the biomechanical weight coefficient and the biomechanical feature vector. The two weighted feature vectors are then concatenated and fed into a mapping network composed of a multilayer perceptron (MLP) for dimensionality alignment calculation, outputting cross-modal collaborative features. This step utilizes etiological weights to guide multimodal feature fusion, highlighting features related to the primary pathogenic factors and suppressing noise interference from irrelevant modalities.

[0031] Step S60: The cross-modal collaborative features are processed by a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

[0032] It should be noted that a multi-task prediction network can be a deep learning architecture that includes a shared backbone feature layer, classification prediction branches, and time-series prediction branches. Classification auxiliary reference data refers to the probability distribution values ​​indicating pathological types such as hypertrophic nonunion or atrophic nonunion. A dynamic healing prediction curve is a time-series broken line reflecting the changing trend of fracture healing probability at multiple discrete time points in the future.

[0033] Understandably, cross-modal collaborative features are input into the shared backbone feature layer (composed of fully connected layers) of the multi-task prediction network to extract high-order abstract features. These high-order abstract features are then input into the classification prediction branch, and after passing through the Softmax activation function, they output subtyping auxiliary reference data, i.e., the probability values ​​of various nonunion pathological types. Simultaneously, these high-order abstract features are input into the temporal prediction branch, and after passing through the Sigmoid activation function, they output the nonunion risk probability corresponding to preset future discrete time blocks (the time block step size is set to 3 months based on the clinical follow-up period). Based on the nonunion risk probability of each discrete time block, the cumulative survival probability is calculated, and the lines connecting these survival probabilities are fitted to generate a dynamic healing prediction curve. This step, by sharing the underlying feature representation and simultaneously performing classification and regression tasks, improves the comprehensive discriminative ability of the overall evolution of nonunion.

[0034] This embodiment integrates heterogeneous data such as medical imaging, biochemical indicators, and internal fixation, and uses principal component analysis to quantify the weight relationships of different etiological dimensions. In the feature extraction stage, it extracts three-dimensional visual features and biomechanical features based on finite element analysis in parallel, and completes cross-modal weighted fusion under the guidance of etiological weights. This embodiment overcomes the limitations of single-modal data in the pathological analysis of complex fractures, and combines a multi-task learning mechanism to simultaneously output pathological classification results and temporal healing trends, achieving multi-dimensional objective evaluation and calculation of nonunion. It effectively uncovers the correlation between clinical manifestations and underlying mechanical mechanisms, providing quantitative reference for clinical auxiliary assessment.

[0035] As an example, the step of quantifying the computed tomography images of the target area and the subject's biochemical index data to obtain fracture gap parameters and microenvironment biological activity scores includes: performing three-dimensional voxel reconstruction on the computed tomography images of the target area to obtain a three-dimensional reconstruction model; performing morphological filling processing on the medullary cavity region in the three-dimensional reconstruction model to obtain the fracture gap region; calculating the effective bone defect volume based on the number of voxels in the fracture gap region; extracting the proximal and distal surface point clouds at the junction of the fracture gap region and the bone, and calculating the proportion of point cloud area within a preset effective contact threshold to obtain the fracture contact surface congruence; combining the effective bone defect volume and the fracture contact surface congruence and converting them into fracture gap parameters; extracting the blood perfusion score, metabolic score, and infection score from the subject's biochemical index data, and performing weighted calculations using a multiple linear regression model to obtain the microenvironment biological activity score.

[0036] It should be noted that a 3D reconstruction model refers to a spatial three-dimensional structure composed of a voxel array carrying spatial coordinate sequences and grayscale information. The medullary cavity region refers to the tubular internal space within the shaft of a long bone that originally contained bone marrow but has been morphologically destroyed or exposed after a fracture. The fracture gap region refers to the three-dimensional spatial range requiring tissue repair formed between the fracture ends due to the interruption of bone continuity. The effective bone defect volume refers to the actual physical volume of space lacking bone filling between the fracture ends. The proximal and distal surface point clouds refer to the spatial geometric point sets composed of discrete 3D coordinate points on the cross-sections of both sides of the fracture ends. The preset effective contact threshold refers to the distance limit for determining whether the spatial coordinate points on the surfaces of the two fracture ends have a mechanical support transmission effect. The fracture end contact surface congruence refers to the proportion of the area of ​​actual mechanical contact between the two fracture ends to the total cross-sectional area of ​​the fracture ends. In this example, the fracture end gap parameter refers to a one-dimensional feature vector composed of the normalized effective bone defect volume and the fracture end contact surface congruence. The blood perfusion score can be a score reflecting the local vascular distribution and blood circulation status of the fracture site, obtained based on imaging examinations. Metabolic scores can be a comprehensive score reflecting the impact of factors such as a patient's glycated hemoglobin level and smoking index on bone metabolism. Infection scores can be a score that assesses the local inflammatory status based on laboratory results such as C-reactive protein levels and erythrocyte sedimentation rate.

[0037] Understandably, the first step is to perform isotropic resampling on the acquired computed tomography image sequence, adjust the voxel spacing to 0.5 mm to eliminate spatial deformation caused by different scan slice thicknesses, and then use the traveling cube algorithm to extract isosurfaces from the resampled voxel data to construct a three-dimensional reconstruction model.

[0038] In the 3D reconstruction model, a region of interest containing the fracture ends is selected. The medullary cavity region within the region of interest is processed using 3D morphological closing operations. The structuring element is a spherical structure with a radius of 3 to 5 pixels. The tiny fracture gaps between the fracture ends are connected by an expansion operation, and the outer edge contour is restored by an erosion operation, thereby wrapping the open defect space to form a closed 3D connected domain. The fracture gap region is then extracted from this connected domain.

[0039] Next, the spatial nodes within the extracted fracture gap region are traversed, and the number of discrete voxels with gray values ​​lower than the gray value threshold for cancellous bone is counted. This gray value threshold for cancellous bone is set to 150 to 250 Henle units according to clinical practice. The counted number of voxels is multiplied by the physical volume of a single voxel (in this example, the physical volume of a single voxel is 0.125 cubic millimeters), and the effective bone defect volume is obtained by summing the results. The formula for calculating the effective bone defect volume is as follows: in, Indicates the effective bone defect volume. A collection of voxels representing the fracture gap region. Indicates the first [fracture interval region] Individual factors, The physical volume of a single three-dimensional voxel is represented by multiplying the horizontal and vertical resolution and slice thickness of a CT image.

[0040] The three-dimensional contours of the junctions between the fracture gap region and normal bone are extracted to generate proximal and distal surface point clouds. A preset effective contact threshold of 1.5 mm is set. Each coordinate point in the proximal surface point cloud is traversed, and its minimum Euclidean distance to all points in the distal surface point cloud is calculated. The number of coordinate points whose minimum Euclidean distance is less than or equal to the preset effective contact threshold is counted. This number is divided by the total number of points in the proximal surface point cloud to obtain the fracture end contact surface fit. The formula for calculating the fracture end contact surface fit is as follows: in, Indicates the fit of the contact surfaces at the fracture ends, with a value range of [value range missing]. The closer the value is to 1, the larger the proportion of effective contact area and the higher the degree of fit. This represents the set of proximal fracture surface points in the fracture gap region. This represents the set of distal fracture surface points in the fracture gap region. This represents the total number of points in the near-endpoint cloud, used as the denominator to normalize the result to a percentage. For indicator functions, when the condition When the condition is met (i.e., the shortest distance from a point on the near end to the far end surface is within the effective contact threshold), the value is 1; otherwise, the value is 0. Point With point The Euclidean distance (i.e., L2 norm) between them; For effective contact threshold, express Any three-dimensional coordinate point in the space, express Any three-dimensional spatial coordinate point in the array.

[0041] Furthermore, the minimum-maximum normalization method is used to linearly scale the effective bone defect volume and the fit of the fracture end contact surface, respectively, to map their values ​​to a dimensionless range of 0 to 1. The normalized effective bone defect volume and the fit of the fracture end contact surface are then concatenated into a one-dimensional feature vector, which is converted into the fracture end gap parameter.

[0042] Finally, the blood perfusion score, metabolic score, and infection score were extracted from the subjects' biochemical index data. A multiple linear regression model based on historical case correlation statistics was obtained. The extracted scores were used as independent variables and input into the multiple linear regression model for weighted summation. The regression coefficients of the blood perfusion score and metabolic score were set to positive values ​​(range 0.2 to 0.6), and the regression coefficient of the infection score was set to negative values ​​(range -0.5 to -0.1). The microenvironment biological activity score was then output and normalized to the interval 0 to 1.

[0043] This example transforms irregular medical images into measurable closed geometric entities through resampling and morphological operations, eliminating the volume interference of free bone fragments and reducing the dimensionality of complex spatial occlusion relationships to uniform proportional values. Combined with a multiple linear regression model, it integrates multidimensional biochemical test indicators into objective physiological repair potential indicators, eliminating dimensional differences between multi-source heterogeneous data, and providing standardized geometric features and biological basis for the comprehensive assessment of the causes of bone nonunion.

[0044] As an example, the three-dimensional image feature extraction network includes multiple three-dimensional convolutional layers, pooling layers, and a global average pooling layer. The step of extracting multi-scale features from the computed tomography image using the three-dimensional image feature extraction network to obtain a visual morphological feature vector includes: converting the computed tomography image into three-dimensional tensor data of a preset resolution; inputting the three-dimensional tensor data into the three-dimensional image feature extraction network, performing local spatial convolution calculations through the multiple three-dimensional convolutional layers to obtain local callus texture features; performing stepwise downsampling processing on the local callus texture features through the pooling layer to obtain macroscopic fracture line morphological features; and performing feature dimensionality reduction and stitching processing on the local callus texture features and the macroscopic fracture line morphological features through the global average pooling layer to obtain a visual morphological feature vector.

[0045] It should be noted that the 3D image feature extraction network refers to a deep neural network architecture that includes 3D convolution operations and downsampling mechanisms. 3D tensor data refers to a three-dimensional numerical matrix composed of three spatial dimensions: width, height, and depth. Preset resolution refers to the standardized and uniform size specifications of the 3D spatial matrix before it is input into the deep learning network. Local callus texture features can be spatial feature expressions reflecting the microscopic trabecular structure of fracture ends and subtle density changes in new bone tissue. Macroscopic fracture line morphology features refer to abstract representations reflecting the overall fracture orientation, misalignment angle of the fracture ends, and large-scale spatial distribution characteristics. Visual morphological feature vectors refer to a one-dimensional numerical sequence that integrates microscopic details and macroscopic morphology, used as input to subsequent prediction models.

[0046] Understandably, after acquiring computed tomography (CT) images, their spatial dimensions are scaled and converted into 3D tensor data at a preset resolution using a trilinear interpolation algorithm. The preset resolution is set to 128×128×128 or 64×64×64, based on the receptive field and hardware memory limitations of mainstream 3D convolutional networks, to ensure that the original medical images of different patients are aligned to the same input dimension.

[0047] Next, the 3D tensor data is input into a 3D image feature extraction network. Multiple 3D convolutional layers configured in the network are used for local spatial convolution calculations. The 3D convolutional layers use 3×3×3 kernels with a stride of 1. The convolutional kernels slide within the spatial domain of the 3D tensor data to extract local spatial gradients and gray-level co-occurrence relationships between adjacent voxels. After activation processing with Leaky ReLU units, local callus texture features are extracted.

[0048] Building upon this, local callus texture features are progressively downsampled using pooling layers. The pooling layers employ the max pooling algorithm, with a pooling window size of 2×2×2 and a stride of 2. This progressive downsampling operation multiplies the feature map spatial size while gradually expanding the receptive field of the network layers, thereby extracting macroscopic fracture line morphological features containing structural information over a larger spatial range.

[0049] Finally, a global average pooling layer is used to reduce the dimensionality of both local callus texture features and macroscopic fracture line morphology features. The global average pooling layer calculates the mean of each channel in the feature map in three-dimensional space, compressing the three-dimensional feature map into a one-dimensional vector, thus eliminating spatial dimension. The dimensionality-reduced one-dimensional microscopic feature vector is then concatenated with the one-dimensional macroscopic feature vector to output a visual morphological feature vector.

[0050] This example preserves the original three-dimensional spatial topology of medical images through trilinear interpolation resampling and a 3D convolutional network design. 3D convolutional operations capture the density variations of bone trabeculae between tiny voxels, while progressive pooling extracts large-scale fracture orientation structures, solving the problem that a single receptive field cannot simultaneously capture both microscopic texture and macroscopic morphology. Global average pooling is used for dimensionality reduction, converting high-dimensional spatial features into one-dimensional vectors while reducing the number of network parameters and suppressing overfitting risk. This provides an objective visual morphological basis that considers both macroscopic and microscopic details for subsequent cross-modal feature fusion.

[0051] As an example, the 3D image feature extraction network is trained through the following steps: acquiring a historical computed tomography (CT) image sample library and the corresponding real morphological labels for each image sample; converting the historical CT image sample library into a 3D tensor sample set; inputting the 3D tensor sample set into the 3D image feature extraction network to be trained for forward propagation convolution calculation to obtain predicted morphological feature vectors; mapping the predicted morphological feature vectors to predicted labels through a preset classifier, and calculating the network prediction error loss based on the predicted labels and the real morphological labels; with the goal of minimizing the network prediction error loss, iteratively updating the convolution kernel weight parameters in the 3D image feature extraction network to be trained based on the backpropagation algorithm until a preset convergence condition is reached, thus obtaining the 3D image feature extraction network.

[0052] It should be noted that the historical computed tomography (CT) image sample database refers to a collection of CT 3D image data containing multiple real fracture cases from the past. Real morphological labels can be objective classification results annotated by professionals, reflecting the actual morphological classification of fracture ends in the images (such as hypertrophic or atrophic morphological features). The 3D tensor sample set refers to a set of three-dimensional numerical matrices constructed by standardizing the spatial dimensions of the image samples. The 3D image feature extraction network to be trained refers to a deep learning model containing 3D convolutional layers that has not yet undergone extensive data parameter optimization and has randomly set initial weights. The predicted morphological feature vector refers to the high-dimensional numerical representation output by the network after feature extraction from the input 3D tensor. The preset classifier can be a fully connected layer structure connected to the end of the feature extraction network, used to map high-dimensional features to specific classification categories. The predicted label refers to the probability distribution judgment result of the model output sample belonging to various morphological features. The network prediction error loss refers to the numerical value that measures the degree of deviation between the model's predicted probability distribution and the objective real label results. The preset convergence condition refers to the criterion for stopping the iterative update of network parameters.

[0053] Understandably, the process begins by retrieving a historical computed tomography (CT) image sample library from a medical imaging database, simultaneously acquiring the corresponding real morphological labels for each image sample. A trilinear interpolation algorithm is then used to spatially resample each image sample in the historical CT image sample library, uniformly adjusting its size to a preset resolution (128×128×128), thereby converting the original images into a structure-aligned three-dimensional tensor sample set.

[0054] Then, the 3D tensor sample set is input into the 3D image feature extraction network to be trained according to the set batch size (Batch Size is set to 8 or 16). Inside the network, the input 3D tensor is used to extract features through multiple 3D convolutional layers containing 3×3×3 convolutional kernels, and forward propagation convolution calculations are performed sequentially to output a one-dimensional predicted morphological feature vector.

[0055] Next, the predicted morphological feature vectors are input into a pre-defined classifier. In this example, the pre-defined classifier is constructed using a single fully connected layer and a Softmax activation function. This classifier maps the feature vectors to predicted labels representing the probabilities of each morphological classification. The predicted labels and their corresponding ground truth morphological labels are obtained, and the difference between their probability distributions is calculated using the Cross-Entropy Loss function to determine the network prediction error loss.

[0056] Based on this, with the goal of minimizing the network prediction error loss, the gradient of the loss function with respect to the weight parameters of each convolutional kernel in the network is calculated using the backpropagation algorithm. The Adam optimization algorithm is then used to iteratively update the convolutional kernel weight parameters based on the calculated gradients, with the initial learning rate of the optimization algorithm set to 0.001.

[0057] Finally, the iterative training process of forward and backward propagation is continued until the preset convergence condition is met. The preset convergence condition is set as follows: the loss value of the network on the validation set no longer decreases for 10 consecutive iterations (Epochs), or the maximum number of training epochs is reached (set to 100 Epochs). When the preset convergence condition is met, the iteration of weight parameter updates is stopped, the current network weight parameters are saved, and the trained 3D image feature extraction network is obtained.

[0058] This example utilizes historical image data with objective morphological annotations to supervise the learning of a 3D convolutional network. Features are extracted through forward propagation, and prediction bias is quantified using the cross-entropy loss function. Backpropagation and an optimizer are then used to continuously adjust the network parameters. This approach allows the convolutional kernel weights to gradually adapt and accurately capture 3D spatial features such as callus texture and fracture lines. The set convergence conditions ensure sufficient model learning while avoiding overfitting, thus providing a reliable visual network foundation for subsequent cross-modal data processing.

[0059] As an example, the steps of establishing a three-dimensional finite element model based on the computed tomography (CT) image and the implant information, and performing biomechanical simulation analysis to obtain biomechanical feature vectors include: performing mesh generation on the CT image to generate a three-dimensional mesh model of the bone; performing spatial registration and topological merging processing on the virtual implant mesh model and the three-dimensional mesh model based on the three-dimensional parameters of the implant information to obtain a bone-implant assembly model; calculating the material property values ​​of each mesh element based on the grayscale values ​​of the CT image and a preset bone density-elastic modulus conversion relationship, and mapping the material property values ​​to the bone-implant assembly model; configuring preset boundary constraints and preset simulated physiological loads as simulation boundary parameters into the node matrix of the bone-implant assembly model to obtain a simulation model to be solved; performing finite element iterative solution on the simulation model to be solved based on a preset elastic static equilibrium equation to obtain the peak stress tensor and micro-displacement values ​​of the stress concentration region, and integrating them into a biomechanical feature vector.

[0060] It should be noted that a 3D mesh model refers to discretizing a continuous 3D geometric entity into a mesh structure composed of a finite number of nodes and elements. A virtual implant mesh model can be the mesh data of internal fixation devices such as intramedullary nails, plates, or screws after finite element discretization. Spatial registration and topology merging refer to the computational process of aligning different models in a unified 3D coordinate system and eliminating contact surface gaps through shared nodes. A bone-implant assembly model refers to an overall mesh structure that integrates bone tissue and internal fixation devices and has a unified mechanical transmission path. Grayscale value refers to the quantitative value reflecting the degree of X-ray absorption by tissue in medical images, usually expressed in Heinz units. The bone density-elastic modulus conversion relationship refers to an empirical mathematical expression based on biomechanical experiments that maps image density to the Young's modulus of a material. Material property values ​​can be Young's modulus and Poisson's ratio, which characterize the tissue's resistance to deformation. Preset boundary constraints refer to boundary parameters that restrict spatial translation or rotation of specific regions of the model. Preset simulated physiological loads refer to force parameters applied to specific locations of the model based on normal human activities. A node matrix is ​​a matrix that records the spatial coordinates of discrete points in a finite element model and their interconnections. The simulation model to be solved is a discrete mathematical model with complete material parameters, boundary constraints, and load conditions configured, ready for calculation by the finite element solver. The preset static equilibrium equations of elasticity describe the physical governing equations that describe the equilibrium state between internal stress and external torque of an object under external loads. The peak stress tensor refers to the maximum stress value and its directional components appearing in a local region of the model during force analysis. The fretting displacement value refers to the relative displacement distance between fracture ends under load.

[0061] Understandably, the process begins by acquiring the reconstructed 3D surface geometry of the bone, followed by mesh generation using a tetrahedral meshing algorithm (such as Delaunay triangulation combined with tetrahedral generation) to generate a 3D mesh model of the bone. To balance computational efficiency with the representation of the fracture end geometry, the mesh element size in the fracture end region is set to 1 mm, while the element size gradually transitions to 3 mm further away from the fracture end region.

[0062] Secondly, the three-dimensional geometric parameters of the implant information are obtained to generate a corresponding virtual implant mesh model. The Iterative Closest Point (ICP) algorithm is used to rigidly register and align the virtual implant mesh model with the three-dimensional mesh model of the bone in the spatial coordinate system. After alignment, spatial registration and topological merging are performed at the contact interface between the bone and the implant. Boolean operations are used to re-divide the contact surface mesh so that the bone units and implant units share nodes, resulting in a bone-implant assembly model.

[0063] Next, the grayscale values ​​corresponding to each voxel in the computed tomography (CT) images are extracted. Based on a pre-defined bone mineral density-elastic modulus conversion relationship, the material property values ​​of each mesh element in the 3D mesh model are calculated. This conversion relationship uses a piecewise empirical formula; for example, grayscale values ​​corresponding to cancellous bone regions are mapped to Young's modulus values ​​ranging from 100 MPa to 1000 MPa, and cortical bone regions are mapped to Young's modulus values ​​ranging from 10 GPa to 20 GPa. The Poisson's ratio for all bone tissues is uniformly set to 0.3. After calculation, these material property values, along with the inherent metallic material properties of the implant, are mapped one by one to the corresponding elements of the bone-implant assembly model.

[0064] Based on this, preset boundary constraints are determined to completely fix the six degrees of freedom of the nodes at the end of the bone-implant assembly model furthest from the fracture region (such as the distal section of the bone shaft). Preset simulated physiological loads are set according to standard human body weight and lower limb force distribution, for example, applying a distributed load of 600 to 800 Newtons along the direction of gravity to the proximal articular surface node region of the model. These constraints and loads are then configured and written into the node matrix of the bone-implant assembly model as simulation boundary parameters to generate the simulation model to be solved.

[0065] Finally, based on the preset static equilibrium equations of elasticity, a finite element solver (such as using the conjugate gradient method to solve the linear algebraic equations) is invoked to perform finite element iterative solutions on the simulation model to be solved. After convergence, the peak stress tensors (such as Von Mises equivalent stress) in stress concentration areas such as the bending points of the internal fixation device and the fracture ends are extracted from the model, and the relative fretting displacement values ​​of the cross-sectional nodes of the upper and lower fracture ends before and after loading are extracted. The extracted peak stress tensors and fretting displacement values ​​are then concatenated in one dimension and integrated into a biomechanical feature vector.

[0066] This example realistically reproduces the mechanical transmission process of fracture lesions containing metal implants under physiological weight-bearing conditions by transforming static medical image voxel data into a finite element mesh structure containing discrete material properties. It directly maps mechanical modulus using image grayscale values, preserving the heterogeneity of bone density differences among individual patients. The extracted peak stress tensor and microdisplacement values, among other objective physical quantities, effectively overcome the limitation of traditional pure morphological observation in assessing the mechanical stability of internal fixation, supplementing the overall assessment and prediction network with underlying mechanical characteristics reflecting the evolution of the microscopic mechanical environment.

[0067] As an example, the step of weighting and fusing the visual morphological feature vector and the biomechanical feature vector according to the etiological weight vector to obtain cross-modal collaborative features includes: parsing visual gating coefficients and mechanical gating coefficients from the etiological weight vector; weighting the visual morphological feature vector according to the visual gating coefficients to obtain weighted visual features; weighting the biomechanical feature vector according to the mechanical gating coefficients to obtain weighted mechanical features; concatenating the weighted visual features and the weighted mechanical features into vectors, and performing dimensionality mapping processing through a fully connected layer to obtain cross-modal collaborative features.

[0068] It should be noted that visual gating coefficients refer to numerical weights used to adjust the proportion of importance of imaging visual features in overall feature fusion. Mechanical gating coefficients refer to numerical weights used to adjust the proportion of importance of finite element biomechanical features in overall feature fusion. Weighted visual features refer to visual feature representations that highlight the associated pathogenic morphological information after being scaled by weighting coefficients. Weighted mechanical features refer to physical feature representations that highlight the associated pathogenic mechanical information after being scaled by weighting coefficients. A fully connected layer can be a layer structure in a deep neural network that fully maps and connects all nodes of the input feature vector to the output node through matrix multiplication. In this example, cross-modal collaborative features refer to aligned feature vectors that eliminate heterogeneous semantic gaps and integrate both morphological and mechanical dimensions.

[0069] Understandably, the first step is to obtain the etiology weight vector output from the previous processing step. The numerical weights corresponding to the internal fixation mechanical dimension are extracted from this etiology weight vector and analyzed as mechanical gating coefficients. The numerical weights corresponding to the anatomical structure dimension and the biological activity dimension are also extracted from this etiology weight vector. These two values ​​are summed to obtain the visual gating coefficients.

[0070] Secondly, the visual morphological feature vector is multiplied and weighted according to the visual gating coefficients obtained from the analysis to obtain the weighted visual features; similarly, the biomechanical feature vector is multiplied and weighted according to the mechanical gating coefficients to obtain the weighted mechanical features.

[0071] Finally, the weighted visual features and weighted mathematical features are concatenated end-to-end along the feature dimension to form a high-dimensional concatenated feature vector. This concatenated feature vector is input into a fully connected layer, where dimensionality mapping is performed through internal weight matrix multiplication and bias term addition. The number of output nodes of the fully connected layer is set to 256 or 512 depending on the input size of the downstream prediction network. The mapped data undergoes non-linear activation using a rectified linear function (ReLU) to output cross-modal collaborative features.

[0072] This example generates corresponding gating coefficients by parsing the causal weight vector, realizing a dynamic weight allocation mechanism based on prior causal factors. During feature fusion, the modal feature responses corresponding to the dominant causal factors are amplified (e.g., the amplitude of mechanical features is significantly amplified when mechanical factors are dominant), reducing information redundancy caused by secondary factors. Combined with dimensionality reduction and mapping processing of fully connected layers, heterogeneous image visual features and physical mechanical features are mapped to a unified high-dimensional feature space, eliminating feature expression differences between different modalities and providing targeted and information-aligned underlying fusion features for subsequent disease classification prediction.

[0073] As an example, the multi-task prediction network includes a shared backbone feature layer, a classification prediction branch, and a temporal prediction branch. The step of processing the cross-modal collaborative features through the multi-task prediction network to obtain subtyping auxiliary reference data and a dynamic healing prediction curve includes: dividing a preset future observation period into multiple discrete time blocks; inputting the cross-modal collaborative features into the shared backbone feature layer for nonlinear mapping to obtain higher-order collaborative features; inputting the higher-order collaborative features into the classification prediction branch to calculate the probability distribution value of each pathological type through a normalized exponential function, and using it as subtyping auxiliary reference data; inputting the higher-order collaborative features into the temporal prediction branch for calculation to obtain the non-healing risk probability in each of the discrete time blocks; calculating the cumulative survival probability based on the non-healing risk probability corresponding to each of the discrete time blocks, and fitting and generating a dynamic healing prediction curve based on the cumulative survival probability.

[0074] It should be noted that the shared backbone feature layer refers to a computational layer composed of multiple cascaded fully connected layers used to extract general abstract features. The classification prediction branch refers to a neural network substructure specifically used to determine the pathological type of nonunion. The temporal prediction branch refers to a neural network substructure used to calculate the probability of nonunion risk at different time points. The preset future observation period refers to the continuous duration for observing fracture healing progress, starting from the current assessment time. A discrete time block refers to a time interval divided by a fixed step size from the continuous observation time. Higher-order collaborative features refer to the abstract feature vector resulting from cross-modal collaborative features after nonlinear spatial transformation. The normalized exponential function refers to the Softmax function used to convert neuron output values ​​into a probability distribution. The cumulative survival probability refers to the probability value that the subject is still in a nonunion state by a specific discrete time block. The dynamic healing prediction curve refers to a trend line reflecting the fluctuation of the predicted healing probability over time.

[0075] Understandably, firstly, a preset future observation period is set based on the routine clinical follow-up cycle for nonunion, with a duration of 12 to 24 months. The preset future observation period is then divided into multiple discrete time blocks, with the step size of each time block set to 3 months based on the frequency of clinical follow-up, for example, divided into [0,3] months, [3,6] months, [6,9] months, and [9,12] months.

[0076] Then, the cross-modal collaborative features are input into the shared backbone feature layer. Nonlinear mapping is performed through fully connected layers in the shared backbone feature layer. Inter-layer processing using the Rectified Linear Unit (ReLU) function projects the input features into a higher-dimensional manifold space to obtain higher-order collaborative features.

[0077] Next, high-order collaborative features are input into the classification prediction branch in parallel. The classification prediction branch maps the features to the pathological type space through a fully connected layer, and calculates the probability distribution values ​​of pathological types such as hypertrophic nonunion, atrophic nonunion, and infectious nonunion through a normalized exponential function. These probability values ​​are used as auxiliary reference data for classification.

[0078] Furthermore, higher-order collaborative features are synchronously input into the temporal prediction branch for computation. The temporal prediction branch contains an output node equal to the number of discrete time blocks. Each node uses the sigmoid activation function to calculate the probability of unhealed risk for the subject within the corresponding discrete time block under the constraints of higher-order collaborative features.

[0079] Based on this, the cumulative survival probability is calculated according to the non-healing risk probability corresponding to each discrete time block. The calculation adopts a multiplicative logic, that is, the cumulative survival probability of a certain time block is the product of (1 - non-healing risk probability) of all discrete time blocks before that time block. The calculated cumulative survival probabilities of each time point are connected, and a dynamic healing prediction curve is generated by fitting the curve using a cubic spline interpolation algorithm.

[0080] This example utilizes a multi-task parallel computing architecture to simultaneously predict pathological classification and temporal prognosis based on shared feature representations. The shared backbone layer extracts generalizable features through nonlinear mapping, effectively improving computational efficiency. The classification branch provides statistically significant type references using a normalized exponential function, while the temporal branch, combined with survival analysis logic, transforms the complex dynamic healing process into an objective probability curve. This solves the problem that traditional assessment methods struggle to quantitatively predict future healing trends, providing multi-dimensional quantitative reference data for subsequent auxiliary analyses.

[0081] As an example, the construction steps of the multi-task prediction network include: constructing a shared backbone feature layer based on a multi-layer fully connected network; constructing a classification prediction branch based on a first fully connected layer and a normalized exponential function activation layer; constructing a temporal prediction branch based on a second fully connected layer and a multi-node logistic regression output layer; and integrating the shared backbone feature layer, the classification prediction branch, and the temporal prediction branch to obtain the multi-task prediction network.

[0082] It should be noted that, in this example, a multilayer fully connected network refers to a collection of cascaded neuron layers that map input features to a high-dimensional nonlinear space through weight matrix multiplication and bias term addition. The first fully connected layer can be a transformation layer configured at the end of a shared backbone to perform pathology type spatial mapping. In this example, the normalized exponential function activation layer refers to an output layer that performs Softmax calculations to map the raw neuron outputs to mutually exclusive probability distributions. The second fully connected layer can be a transformation layer configured at the end of a shared backbone to perform healing risk probability mapping. The multi-node logistic regression output layer refers to a probability calculation array consisting of multiple independent neuron nodes using a logistic regression activation function.

[0083] Understandably, the first step is to construct a shared backbone feature layer based on a multi-layer fully connected network. Three to five fully connected layers are then set as the basic architecture for feature extraction. The number of neurons in each layer is set to 128 to 512 depending on the dimension of the cross-modal collaborative features input, ensuring effective feature compression and spatial transformation of high-dimensional collaborative features. Rectified linear functions (ReLU) are configured between layers to perform non-linear transformations, extracting general abstract associations that can support different downstream prediction tasks.

[0084] Then, a classification prediction branch is constructed based on the first fully connected layer and the normalized exponential function activation layer. The first fully connected layer is connected to the output node of the shared backbone feature layer. The output dimension of the first fully connected layer is set to 3 to correspond to the three preset pathological types: hypertrophic nonunion, atrophic nonunion, and infectious nonunion. A normalized exponential function activation layer is configured after the first fully connected layer to convert the linear values ​​of the neuron output into a normalized distribution representing the probability of each classification result.

[0085] Next, a temporal prediction branch is constructed based on the second fully connected layer and the multi-node logistic regression output layer. The second fully connected layer is connected to the same output port of the shared backbone feature layer. The number of output neurons in the second fully connected layer is set according to the preset number of discrete time blocks, which is set to 4 to 8 nodes in this example. A multi-node logistic regression output layer is configured after the second fully connected layer, and each neuron node uses the sigmoid activation function to calculate the probability of unhealed risk at the corresponding time point.

[0086] Building upon this foundation, the shared backbone feature layer, classification prediction branch, and temporal prediction branch are integrated. The classification and temporal prediction branches are arranged in parallel spatially and jointly receive high-order feature information extracted from the shared backbone feature layer. Through this integration method of sharing underlying parameters, a multi-task prediction network with cross-domain multi-index inference capabilities is constructed.

[0087] This example extracts abstract representations with general semantics through a shared feature layer, reducing computational resource consumption during model training and significantly improving feature utilization efficiency. The branched prediction structure enhances constraints between different tasks by sharing underlying weights, suppressing prediction fluctuations that may arise from independent modeling. The collaborative work of the classification branch and the temporal branch enables the simultaneous output of pathological classification references and dynamic healing curves, solving the problem that traditional methods cannot perform multi-dimensional state assessments within a unified computational framework. This provides logically consistent and data-aligned quantitative support for clinical decision support.

[0088] As an example, the multi-task prediction network is trained through the following steps: acquiring a cross-modal collaborative feature sample set of historical nonunion cases, as well as the true classification label and follow-up healing status label corresponding to each sample; inputting the cross-modal collaborative feature sample set into the multi-task prediction network to be trained, and outputting the predicted classification probability and predicted risk probability respectively; calculating the cross-entropy loss based on the predicted classification probability and the true classification label, and calculating the log-likelihood loss based on the predicted risk probability and the follow-up healing status label; weighted summing the cross-entropy loss and the log-likelihood loss to obtain the total loss function; using the gradient descent algorithm to iteratively update the parameters of the multi-task prediction network to be trained with the goal of minimizing the total loss function, until a preset convergence condition is met, thus obtaining the multi-task prediction network.

[0089] It should be noted that the cross-modal collaborative feature sample set refers to a collection of sample data fused from multiple visual and biomechanical features. The true classification label refers to the annotation information of the pathological type of nonunion confirmed by clinical diagnosis. The follow-up healing status label refers to the annotation data recording whether historical cases have reached the clinical healing criteria at multiple follow-up time points. In this example, the multi-task prediction network to be trained refers to a deep learning model with randomly distributed initial weight parameters used to learn the mapping relationship between features and labels. In this example, the predicted classification probability refers to the numerical value of the likelihood that a sample belongs to different pathological types. In this example, the predicted risk probability refers to the numerical value of the likelihood that a subject will not experience a healing event within a preset discrete time block. Cross-entropy loss is a cost function used to measure the distributional difference between the classification prediction result and the true category label. Log-likelihood loss is a cost function used to measure the accuracy of risk prediction based on the maximum likelihood estimation logic of survival analysis. The total loss function is a comprehensive optimization objective that integrates classification error and temporal prediction error through weighted summation. Gradient descent algorithm is an optimization algorithm that updates network parameters in the opposite direction of the gradient of the objective function to seek a local optimum.

[0090] Understandably, the process begins by obtaining a cross-modal collaborative feature sample set of historical nonunion cases from a historical case database, and simultaneously obtaining the true classification label and follow-up healing status label corresponding to each sample. The follow-up healing status label contains healing event indicators for the patient at each follow-up period, such as using 1 to represent healed and 0 to represent not healed.

[0091] Then, the cross-modal collaborative feature sample set is input into the multi-task prediction network to be trained according to the preset batch size. The preset batch size is set to an integer between 16 and 64 according to the computing resources, and the initial learning rate is set to 0.0001 to 0.001. The multi-task prediction network to be trained outputs the predicted classification probability and the predicted risk probability respectively.

[0092] Next, the predicted subtyping probability and the true subtyping label are obtained and substituted into the cross-entropy function. The cross-entropy loss is obtained by calculating the difference in classification distribution. Simultaneously, the predicted risk probability and the follow-up healing status label are obtained and substituted into the negative log-likelihood function. The log-likelihood loss is obtained by calculating the fitting deviation of the model to the true healing event.

[0093] Based on this, a balance coefficient is introduced to weight and sum the cross-entropy loss and log-likelihood loss to obtain the total loss function. The value range of this balance coefficient is set from 0.3 to 0.7, which is used to adjust the contribution ratio of different prediction tasks in the parameter learning process according to clinical needs.

[0094] Finally, the Adam optimization algorithm is used as the gradient descent algorithm. With the goal of minimizing the total loss function, the gradient of the total loss function with respect to the network parameters is calculated. The parameters in the multi-task prediction network to be trained are iteratively updated along the inverse direction of the gradient until a preset convergence condition is met. In this example, the preset convergence condition is set as follows: the validation set loss value no longer decreases within 20 consecutive training epochs, or the total number of iteration epochs reaches 150. At this point, updates are stopped, the weights are saved, and the trained multi-task prediction network is obtained.

[0095] This example demonstrates how a model can simultaneously learn the static pathological classification features and dynamic healing evolution patterns of nonunion by constructing a multi-objective loss function that incorporates both classification error and survival risk error. A weighted summation loss function balances the optimization weights across different clinical assessment tasks, enhancing the generalization ability of feature representations in multi-dimensional analysis scenarios. This ensures that the trained multi-task prediction network can output logically consistent and clinically reliable auxiliary classification data and healing risk prediction results.

[0096] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the medical data processing method for bone nonunion healing analysis according to this application. The multidimensional etiological association model includes a data standardization processing layer, a principal component analysis layer, and a weight allocation evaluation layer. Step S20 of the medical data processing method for bone nonunion healing analysis includes steps S21 to S26: Step S21: Extract anatomical features, biological activity features, and internal fixation mechanical features from the fracture end gap parameters, the microenvironment biological activity score, and the implant information, respectively. Step S22: Obtain the subject's basic medical history data and extract risk exposure characteristics from it; Step S23: The anatomical structural features, biological activity features, internal fixation mechanical features, and risk exposure features are combined into a comprehensive etiological feature matrix; Step S24: The comprehensive etiological feature matrix is ​​subjected to dimensional unification processing through the data standardization processing layer to obtain a standardized feature matrix; Step S25: Perform feature decomposition on the standardized feature matrix through the principal component analysis layer to extract the dominant risk components; Step S26: Through the weight allocation evaluation layer, the contribution ratio of each feature dimension to nonunion is calculated based on the dominant risk component, and the contribution ratios are normalized to obtain the etiology weight vector.

[0097] It should be noted that the data standardization layer refers to the computational logic unit that performs mean centering and standard deviation scaling. The principal component analysis layer refers to the computational module that extracts the main variance directions of high-dimensional data using covariance matrix decomposition. The weight allocation evaluation layer refers to the evaluation module that calculates the influence of the weights of the original input features based on the variance contribution rate of the principal components. Anatomical structural features can refer to a set of numerical values ​​reflecting fracture end displacement, angulation, and bone defect volume. Biological activity features can refer to quantitative data including blood supply fraction, metabolic rate, and osteogenic factor levels. Internal fixation mechanical features refer to mechanical parameters characterizing the Young's modulus, fixation configuration, and force line distribution of the implant. Risk exposure features refer to the quantitative risk values ​​calculated based on the subject's age, underlying disease records, and smoking and alcohol consumption history. The standardized feature matrix refers to a numerical matrix that has eliminated dimensional differences and conforms to a standard normal distribution. The dominant risk component refers to the set of feature vectors whose cumulative variance contribution rate reaches a preset proportion.

[0098] Understandably, firstly, the displacement distance and defect volume of the fracture ends are extracted from the fracture gap parameters as anatomical structural features; metabolic rate and blood perfusion score are extracted from the microenvironment biological activity score as biological activity features; and Young's modulus values ​​corresponding to the material and fixation configuration parameters are extracted from the implant information as internal fixation mechanical features. These features together constitute the basic feature set reflecting the local condition of the fracture.

[0099] Next, basic medical history data of the subjects were obtained. This data included age, whether there was a history of diabetes (e.g., 1 for yes, 0 for no), and smoking index (e.g., by annual smoking amount). Risk exposure characteristics were extracted from these records to measure the potential negative interference of systemic factors on the healing process.

[0100] Based on this, a one-dimensional first-to-last stitching operation is performed on the anatomical structural features, biological activity features, internal fixation mechanical features, and risk exposure features in the feature dimension to construct a comprehensive etiological feature matrix containing all input dimensions.

[0101] Then, the comprehensive etiological feature matrix is ​​input into the data standardization layer. The data standardization layer calculates the mean and standard deviation of each dimension in the input matrix and performs Z-Score standardization on the comprehensive etiological feature matrix to eliminate dimensional differences between indicators with different physical meanings. The resulting standardized feature matrix output has elements that conform to a distribution with a mean of 0 and a variance of 1, thus eliminating the order-of-magnitude differences between indicators with different physical meanings.

[0102] Furthermore, the standardized feature matrix is ​​input into the principal component analysis (PCA) layer for eigenvalue decomposition. The PCA layer first calculates the covariance matrix of the standardized feature matrix, and then calculates the eigenvalues ​​and eigenvectors of this covariance matrix. The eigenvectors are sorted in descending order of eigenvalues, and the eigenvectors whose cumulative variance contribution rate reaches a preset threshold are selected as the dominant risk components. The preset threshold is set between 80% and 90% according to general statistical principles; in this example, it is set to 85% to ensure that the extracted components represent the vast majority of key information in the original data.

[0103] Finally, the dominant risk components are input into the weight allocation evaluation layer. The weight allocation evaluation layer uses the variance contribution rate corresponding to each principal component as a scaling factor to map the feature vectors in the dominant risk components back to the original input feature space. By summing the absolute values ​​of the contributions of each feature dimension to all selected principal components, the contribution ratio of each feature dimension to the formation of nonunion is obtained. These contribution ratios are then normalized so that their sum equals 1, ultimately outputting the etiology weight vector.

[0104] This embodiment constructs a multidimensional etiological association model and utilizes principal component analysis (PCA) to uncover deep correlations among anatomical, biological, mechanical, and risk factors. The data standardization layer eliminates the dimensional gap between heterogeneous data, while the PCA layer effectively filters redundant noise from the multidimensional data and extracts the core components that play a decisive role in bone healing outcomes. The weight allocation and evaluation layer transforms complex statistical analysis results into intuitive etiological weight vectors, quantifying the proportion of different pathogenic factors' influence on nonunion and providing a scientific pathological basis for the subsequent differentiated weighted fusion of visual and mechanical features, significantly enhancing the interpretability of the model in assessing complex clinical cases.

[0105] As an example, the multidimensional etiology association model is trained through the following steps: acquiring historical statistical data on bone nonunion and extracting historical multidimensional etiology feature data from it as training samples; inputting the training samples into the multidimensional etiology association model to be trained, and performing mean centering and variance normalization through the data standardization processing layer to obtain a standardized feature matrix; calculating the covariance matrix of the standardized feature matrix, and performing eigenvalue decomposition on the covariance matrix to extract eigenvectors and corresponding eigenvalues; calculating the variance contribution rate corresponding to each eigenvector based on the eigenvalues, and selecting a preset number of eigenvectors as dominant risk components in descending order of variance contribution rate; updating and configuring the mapping matrix parameters of the multidimensional etiology association model to be trained based on the selected dominant risk components to obtain the multidimensional etiology association model.

[0106] It should be noted that historical statistics on nonunion of bones refer to a collection of clinical records integrating anatomical, biological, mechanical, and risk dimensions from past cases. Historical multidimensional etiological feature data refers to a set of numerical features extracted from historical statistical data to quantify pathogenic factors. Training samples refer to one or more sets of numerical vectors carrying etiological features as input to the model. Mean centering refers to the process of subtracting the sample mean to make the mean of the data sequence zero. Variance normalization refers to the process of dividing by the sample standard deviation to make the variance of the data sequence equal to one. The covariance matrix is ​​a symmetric matrix used to characterize the degree of correlation between multidimensional random variables. Eigenvalue decomposition refers to the spatial transformation process of decomposing a matrix into eigenvalues ​​and corresponding eigenvectors. An eigenvalue is a scalar value reflecting the amount of information carried in the direction of the corresponding eigenvector in eigenvalue decomposition. Variance contribution rate refers to the proportion of a single eigenvalue in the sum of all eigenvalues. The preset number can be the number of eigenvectors included when the cumulative variance contribution rate reaches a preset threshold. The mapping matrix parameters refer to the weight coefficient matrix used in the multidimensional etiology association model to be trained to project high-dimensional original features onto a low-dimensional principal component space. In this example, the data standardization layer refers to the computational logic unit that performs mean shifting and standard deviation scaling to align feature dimensions.

[0107] Understandably, the first step is to retrieve historical statistical data on bone nonunion from medical imaging and medical record databases, extract quantitative feature values ​​related to four dimensions: anatomical structure, biological activity, internal fixation mechanics, and risk exposure, and combine them into historical multidimensional etiological feature data as training samples.

[0108] Subsequently, the training samples are input into the multidimensional etiology association model to be trained. The data standardization layer calculates the arithmetic mean of all training samples in each dimension, and performs mean centering by subtracting the arithmetic mean from the original values. Next, the standard deviation of the samples in each dimension is calculated, and the variance is normalized by dividing the centered values ​​by the standard deviation, resulting in a standardized feature matrix that eliminates the order-of-magnitude differences between indicators.

[0109] Then, the covariance matrix of the standardized eigenvalue matrix is ​​calculated. The covariance matrix is ​​then decomposed using the eigenvalue decomposition algorithm in numerical analysis to extract a set of mutually orthogonal eigenvectors and the corresponding eigenvalues ​​for each eigenvector.

[0110] Based on this, the variance contribution rate of each eigenvector is calculated according to the magnitude of the eigenvalues. The eigenvectors are sorted in descending order of variance contribution rate, and the contribution rate of each vector is accumulated. When the accumulated contribution rate reaches a preset threshold, a preset number of eigenvectors is determined. The preset threshold is set between 80% and 95% based on the requirement to retain original data information; in this example, it is set to 85%. The eigenvectors that meet this condition are selected as the dominant risk components.

[0111] Finally, the feature vectors from the selected dominant risk components are used as column vectors to update the mapping matrix parameters. The updated mapping matrix parameters establish a linear projection relationship from the multidimensional original etiology space to the dominant risk space, thus obtaining the trained multidimensional etiology association model.

[0112] This example eliminates dimensional interference between different clinical indicators by performing mean centering and variance normalization on historical etiology data. Combined with eigenvalue decomposition of the covariance matrix and variance contribution rate calculation, it extracts the dominant components from high-dimensional pathogenic factors. By updating the mapping matrix parameters, the model can objectively configure etiology association weights based on the statistical distribution characteristics of historical samples. This not only achieves feature dimensionality reduction and data denoising but also ensures the statistical rigor of weight allocation in subsequent evaluation processes, providing a stable computational foundation for accurately analyzing the dominant causes of bone nonunion.

[0113] This application also provides a medical data processing device for analyzing bone nonunion healing; please refer to [reference needed]. Figure 3 The medical data processing device for analyzing bone nonunion healing includes: Data quantification module 10 is used to quantify the computed tomography images of the target area and the biochemical index data of the subjects to obtain the fragment gap parameters and the microenvironment biological activity score. The etiology assessment module 20 is used to input the fracture gap parameter, the microenvironment biological activity score and implant information into a pre-trained multidimensional etiology association model for assessment and calculation to obtain an etiology weight vector. The visual extraction module 30 is used to extract multi-scale features from the computed tomography image through a three-dimensional image feature extraction network to obtain a visual morphological feature vector. The mechanical simulation module 40 is used to establish a three-dimensional finite element model based on the computed tomography images and the implant information, and to perform biomechanical simulation analysis to obtain biomechanical feature vectors. The feature fusion module 50 is used to perform weighted fusion of the visual morphological feature vector and the biomechanical feature vector according to the etiology weight vector to obtain cross-modal collaborative features; The analysis and prediction module 60 is used to process the cross-modal collaborative features through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

[0114] The medical data processing device for nonunion healing analysis provided in this application employs the medical data processing method for nonunion healing analysis described in the above embodiments. It solves the technical problem of how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of nonunion and the prediction of dynamic healing trends. Compared with the prior art, the beneficial effects of the medical data processing device for nonunion healing analysis provided in this application are the same as those of the medical data processing method for nonunion healing analysis provided in the above embodiments. Furthermore, other technical features in the medical data processing device for nonunion healing analysis are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0115] This application provides a medical data processing device for analyzing bone nonunion healing. The medical data processing device for analyzing bone nonunion healing includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the medical data processing method for analyzing bone nonunion healing described in Embodiment 1 above.

[0116] The following is for reference. Figure 4This document illustrates a schematic diagram of a medical data processing device suitable for implementing bone nonunion healing analysis according to embodiments of this application. The medical data processing device for bone nonunion healing analysis in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Android Devices), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The medical data processing device shown for analyzing bone nonunion healing is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0117] like Figure 4 As shown, a medical data processing device for analyzing bone nonunion healing may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes based on programs stored in ROM 1002 (Read Only Memory) or programs loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the medical data processing device for analyzing bone nonunion healing. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. An I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the medical data processing device for analyzing nonunion healing to exchange data wirelessly or via wired communication with other devices. Although the figure shows a medical data processing device for analyzing nonunion healing with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0118] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application 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 a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0119] The medical data processing device for nonunion healing analysis provided in this application, employing the medical data processing method for nonunion healing analysis described in the above embodiments, can solve the technical problem of how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of nonunion and the prediction of dynamic healing trends. Compared with the prior art, the beneficial effects of the medical data processing device for nonunion healing analysis provided in this application are the same as those of the medical data processing method for nonunion healing analysis provided in the above embodiments, and other technical features in this medical data processing device for nonunion healing analysis are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0120] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0122] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the medical data processing method for bone nonunion healing analysis in the above embodiments.

[0123] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the 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, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0124] The aforementioned computer-readable storage medium may be included in a medical data processing device for analyzing bone nonunion healing; or it may exist independently and not be assembled into a medical data processing device for analyzing bone nonunion healing.

[0125] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a medical data processing device for analyzing bone nonunion healing, the medical data processing device for analyzing bone nonunion healing performs the following: quantification processing of computed tomography images of the target area and biochemical index data of the subject to obtain fracture gap parameters and microenvironment biological activity scores; inputting the fracture gap parameters, the microenvironment biological activity scores, and implant information into a pre-trained multidimensional etiology association model for evaluation and calculation to obtain etiology weight vectors; performing multi-scale feature extraction on the computed tomography images through a three-dimensional image feature extraction network to obtain visual morphological feature vectors; establishing a three-dimensional finite element model based on the computed tomography images and the implant information, and performing biomechanical simulation analysis to obtain biomechanical feature vectors; weighted fusion of the visual morphological feature vectors and the biomechanical feature vectors based on the etiology weight vectors to obtain cross-modal collaborative features; and processing the cross-modal collaborative features through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

[0126] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as 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 LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0127] 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 application. 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 the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can 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.

[0128] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0129] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described medical data processing method for analyzing bone nonunion healing. This method can solve the technical problem of how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of bone nonunion and the prediction of dynamic healing trends. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the medical data processing method for analyzing bone nonunion healing provided in the above embodiments, and will not be repeated here.

[0130] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the medical data processing method for bone nonunion healing analysis as described above.

[0131] The computer program product provided in this application can solve the technical problem of how to obtain accurate quantitative analysis reference data based on multi-source heterogeneous medical data to assist in the classification assessment of bone nonunion and the prediction of dynamic healing trends. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the medical data processing method for bone nonunion healing analysis provided in the above embodiments, and will not be repeated here.

[0132] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A medical data processing method for nonunion healing analysis, characterized by, The method includes: The computed tomography images of the target area and the biochemical index data of the subjects were quantified to obtain the fragment gap parameters and the biological activity score of the microenvironment. The fracture gap parameters, the microenvironment biological activity score, and the implant information are input into a pre-trained multidimensional etiology association model for evaluation and calculation to obtain an etiology weight vector. The computed tomography images are subjected to multi-scale feature extraction using a three-dimensional image feature extraction network to obtain visual morphological feature vectors. A three-dimensional finite element model is established based on the computed tomography images and the implant information, and biomechanical simulation analysis is performed to obtain biomechanical feature vectors. The visual morphological feature vector and the biomechanical feature vector are weighted and fused according to the etiological weight vector to obtain cross-modal collaborative features; The cross-modal collaborative features are processed by a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

2. The method of claim 1, wherein, The multidimensional etiology association model includes a data standardization processing layer, a principal component analysis layer, and a weight allocation evaluation layer. The step of inputting the fracture gap parameter, the microenvironment biological activity score, and implant information into a pre-trained multidimensional etiology association model for evaluation and calculation to obtain the etiology weight vector includes: Anatomical structural features, biological activity features, and internal fixation mechanical features are extracted from the fracture gap parameters, the microenvironment biological activity score, and the implant information, respectively. Obtain the subjects' basic medical history data and extract risk exposure characteristics from it; The anatomical structural features, biological activity features, internal fixation mechanical features, and risk exposure features are combined into a comprehensive etiological feature matrix; The data standardization processing layer performs dimensional unification processing on the comprehensive etiological feature matrix to obtain a standardized feature matrix. The principal component analysis layer is used to perform feature decomposition on the standardized feature matrix to extract the dominant risk components. The weight allocation evaluation layer calculates the contribution ratio of each feature dimension to nonunion based on the dominant risk component, and normalizes each contribution ratio to obtain the etiology weight vector.

3. The method as described in claim 1, characterized in that, The three-dimensional image feature extraction network includes multiple three-dimensional convolutional layers, pooling layers, and global average pooling layers; The step of extracting multi-scale features from the computed tomography image using a three-dimensional image feature extraction network to obtain a visual morphological feature vector includes: The computed tomography images are converted into three-dimensional tensor data of a preset resolution; The three-dimensional tensor data is input into a three-dimensional image feature extraction network, and local spatial convolution calculation is performed through the multiple three-dimensional convolutional layers to obtain local callus texture features; The local callus texture features are downsampled stepwise through the pooling layer to obtain the macroscopic fracture line morphology features. The local callus texture features and the macroscopic fracture line morphology features are processed by the global average pooling layer to perform feature dimensionality reduction and splicing, resulting in a visual morphology feature vector.

4. The method as described in claim 1, characterized in that, The steps of establishing a three-dimensional finite element model based on the computed tomography images and the implant information, and performing biomechanical simulation analysis to obtain biomechanical feature vectors include: The computed tomography images are processed into a mesh to generate a three-dimensional mesh model of the skeleton; Based on the three-dimensional parameters of the implant information, the virtual implant mesh model and the three-dimensional mesh model are spatially registered and topologically merged to obtain the bone-implant assembly model. Based on the grayscale values ​​of the computed tomography images and the preset bone density-elastic modulus conversion relationship, the material property values ​​of each grid unit are calculated, and the material property values ​​are mapped to the bone-implant assembly model. The preset boundary constraints and preset simulated physiological loads are configured as simulation boundary parameters into the node matrix of the bone-implant assembly model to obtain the simulation model to be solved. The simulation model to be solved is solved by finite element iterative solution based on the preset static equilibrium equation of elasticity to obtain the peak stress tensor and micro-displacement value of the stress concentration region, and integrated into a biomechanical feature vector.

5. The method as described in claim 1, characterized in that, The multi-task prediction network includes a shared backbone feature layer, a classification prediction branch, and a temporal prediction branch. The step of processing the cross-modal collaborative features through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves includes: The preset future observation period is divided into multiple discrete time blocks; The cross-modal collaborative features are input into the shared backbone feature layer for nonlinear mapping to obtain higher-order collaborative features. The higher-order collaborative features are input into the classification prediction branch, and the probability distribution values ​​of each pathological type are calculated through the normalized exponential function, which are used as auxiliary reference data for classification. The higher-order collaborative features are input into the temporal prediction branch for calculation to obtain the probability of unhealed risk in each discrete time block; The cumulative survival probability is calculated based on the non-healing risk probability corresponding to each discrete time block, and a dynamic healing prediction curve is generated by fitting the cumulative survival probability.

6. The method as described in claim 1, characterized in that, The step of quantifying the computed tomography images of the target area and the biochemical index data of the subject to obtain the fragment gap parameters and the microenvironment biological activity score includes: Three-dimensional voxel reconstruction is performed on the computed tomography images of the target area to obtain a three-dimensional reconstruction model; Morphological filling processing is performed on the medullary cavity region in the three-dimensional reconstruction model to obtain the fracture gap region; The effective bone defect volume is calculated based on the number of voxels within the fracture gap region. Extract the proximal and distal surface point clouds of the fracture gap region at the junction with the bone, and calculate the proportion of point cloud area within a preset effective contact threshold to obtain the contact surface fit of the fracture ends. The effective bone defect volume and the degree of fit of the fracture end contact surface are combined and converted into a fracture end gap parameter. Blood perfusion score, metabolic score, and infection score were extracted from the subjects' biochemical data and weighted using a multiple linear regression model to obtain the microenvironment biological activity score.

7. The method according to any one of claims 1 to 6, characterized in that, The multi-task prediction network is trained through the following steps: Obtain a cross-modal collaborative feature sample set of historical nonunion cases, as well as the true subtype label and follow-up healing status label for each sample; The cross-modal collaborative feature sample set is input into the multi-task prediction network to be trained, and the predicted classification probability and predicted risk probability are output respectively. The cross-entropy loss is calculated based on the predicted subtyping probability and the true subtyping label, and the log-likelihood loss is calculated based on the predicted risk probability and the follow-up healing status label. The total loss function is obtained by weighted summing of the cross-entropy loss and the log-likelihood loss. With the goal of minimizing the total loss function, the gradient descent algorithm is used to iteratively update the parameters of the multi-task prediction network to be trained until the preset convergence condition is met, thus obtaining the multi-task prediction network.

8. A medical data processing device for analyzing bone nonunion healing, characterized in that, The device employs the medical data processing method for analyzing bone nonunion healing as described in any one of claims 1 to 7, and the device comprises: The data quantification module is used to quantify the computed tomography images of the target area and the biochemical index data of the subjects to obtain the fragment gap parameters and the biological activity score of the microenvironment. The etiology assessment module is used to input the fracture gap parameters, the microenvironment biological activity score and implant information into a pre-trained multidimensional etiology association model for evaluation and calculation, and to obtain the etiology weight vector. The visual extraction module is used to extract multi-scale features from the computed tomography images through a three-dimensional image feature extraction network to obtain visual morphological feature vectors. The mechanical simulation module is used to establish a three-dimensional finite element model based on the computed tomography images and the implant information, and to perform biomechanical simulation analysis to obtain biomechanical feature vectors. The feature fusion module is used to perform weighted fusion of the visual morphological feature vector and the biomechanical feature vector according to the etiology weight vector to obtain cross-modal collaborative features; The analysis and prediction module is used to process the cross-modal collaborative features through a multi-task prediction network to obtain subtyping auxiliary reference data and dynamic healing prediction curves.

9. A medical data processing device for analyzing bone nonunion healing, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the medical data processing method for analyzing nonunion healing as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the medical data processing method for bone nonunion healing analysis as described in any one of claims 1 to 7.

Images