Method for creating image recognition model based on clinical disease data integration

Abstract

The present invention relates to the field of model construction technology, and in particular to a method for creating an image recognition model based on clinical disease data integration. The method comprises the following steps: acquiring a medical clinical orthopedic image set; extracting bone density and bone morphology from the medical clinical orthopedic image set and performing bone structure segmentation on the medical clinical orthopedic image set to generate a bone structure segmentation image, wherein the bone structure segmentation image includes a large bone structure image and a small bone structure image; analyzing the bone layer thickness of the large bone structure image and performing physical layer thickness compensation reconstruction on the large bone structure image to generate large bone intermediate virtual slice data; performing detail-guided reconstruction on the small bone structure image to generate small bone fracture texture reconstruction data. The present invention improves the accuracy of dynamic bone image recognition by enhancing image resolution, integrating multimodal data, performing motion simulation, and optimizing model training.

Description

Technical Field

[0001] The present invention relates to the field of model construction technology, and in particular to a method for creating an image recognition model based on clinical disease data integration. Background Art

[0002] With the deep integration of artificial intelligence and healthcare, image recognition-based diagnostic technologies have become a research hotspot. Early medical image recognition models primarily relied on traditional image processing methods and shallow machine learning algorithms, which struggled to cope with the complex and ever-changing characteristics of clinical lesions. In recent years, the widespread application of deep learning, particularly models such as convolutional neural networks (CNNs) and Transformers, has significantly improved the accuracy and automation of medical image processing. However, single image data often fails to fully reflect a patient's condition, limiting the adaptability of these models in real-world clinical settings. To address this, researchers have begun exploring methods for integrating structured electronic medical record data (such as age, gender, medical history, and laboratory test results) with image information, gradually developing a research direction for "clinical-image data integration" models. These methods typically utilize advanced technologies such as multimodal learning, feature alignment, attention mechanisms, and graph neural networks to achieve cross-data source information collaboration and semantic enhancement. However, traditional orthopedic image analysis is often limited by image resolution, failing to capture subtle changes and details in bone structure. Furthermore, they often rely solely on image data or a single feature type, overlooking the interaction and collaboration of other relevant information, resulting in low accuracy in dynamic bone image recognition. Summary of the Invention

[0003] Based on this, it is necessary to provide a method for creating an image recognition model based on clinical disease data integration to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a method for creating an image recognition model based on clinical disease data integration is provided, the method comprising the following steps:

[0005] Step S1: obtaining a medical clinical orthopedic image set; extracting bone density and bone morphology from the medical clinical orthopedic image set and performing bone structure segmentation on the medical clinical orthopedic image set to generate a bone structure segmentation image, wherein the bone structure segmentation image includes a large bone structure image and a small bone structure image;

[0006] Step S2: Analyze the bone layer thickness of the large bone structure image and perform physical layer thickness compensation reconstruction on the large bone structure image to generate large bone middle virtual slice data; perform detail-guided reconstruction on the small bone structure image to generate small bone fracture texture reconstruction data; use the large bone middle virtual slice data and small bone fracture texture reconstruction data to enhance the image resolution of the bone structure segmentation image to generate an enhanced medical clinical orthopedic image set;

[0007] Step S3: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set; performing graph convolution dynamic interaction modeling on the medical clinical orthopedic image set using the decoder to generate a bone-muscle-ligament spatial topology map; performing motion simulation on the medical clinical orthopedic image set based on the bone-muscle-ligament spatial topology map to generate bone motion simulation data;

[0008] Step S4: Perform model training on the medical clinical orthopedic image set using bone motion simulation data to generate an orthopedic dynamic image recognition model.

[0009] This invention significantly improves image clarity and detail preservation by reconstructing large bone structures using physical slice thickness compensation and small bone structures using fracture texture-guided reconstruction, providing high-quality input data for subsequent analysis. Image resolution enhancement, combining large bone intermediate virtual slice data with small bone texture reconstruction data, effectively improves image spatial resolution, particularly for the accurate representation of small bones and fine structures. Based on a dual-channel image decoder design, it enables joint perception and decoding of bone structure and soft tissue (such as muscle and ligaments), providing comprehensive data support for subsequent 3D modeling and motion simulation. Dynamic interactive graph convolution modeling constructs a bone-muscle-ligament spatial topology map, enhancing the ability to model associations between anatomical structures and facilitating accurate simulation of real-world physiological motion states. The resulting orthopedic dynamic image recognition model, trained on skeletal motion simulation data, possesses enhanced structural understanding and dynamic perception capabilities, making it particularly suitable for clinical applications such as preoperative planning, functional assessment, and lesion prediction. The overall process integrates multiple technologies, including image enhancement, structural reconstruction, topological modeling, and dynamic simulation, advancing the transition of medical image analysis from static recognition to dynamic understanding, improving the system's intelligent decision-making capabilities and clinical adaptability. Therefore, the present invention improves the accuracy of dynamic recognition of skeletal images by enhancing image resolution, multimodal data integration, motion simulation and optimization model training.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: obtaining a medical clinical orthopedic image set, wherein the medical clinical orthopedic image set includes CT images and MRI images; performing image normalization on the CT images and the MRI images to generate a normalized orthopedic image set;

[0012] Step S12: analyzing the grayscale variation rate of the bone region in the orthopedic image set to obtain regional bone density distribution data;

[0013] Step S13: extracting bone morphological features from the orthopedic image data to generate bone morphological feature data;

[0014] Step S14: performing bone structure clustering on the medical clinical orthopedic image set according to the regional bone density distribution data and the bone morphological feature data to generate a bone structure clustering graph;

[0015] Step S15: using the bone structure clustering graph to perform structural label division on the medical clinical orthopedic image set to generate a bone structure division image, wherein the bone structure division image includes a large bone structure image and a small bone structure image.

[0016] The present invention normalizes CT images and MRI images, eliminating the differences between different image sources and scanning methods, allowing different types of images to be compared and analyzed under the same standard, thereby improving the consistency and stability of the data. By analyzing the grayscale variation rate of the bone area in the orthopedic image, regional bone density distribution data is obtained. This analysis helps to more accurately reflect the density distribution changes of the bone structure and can provide accurate bone density features for subsequent image refinement processing and deep learning models. By extracting bone morphological features in orthopedic images, detailed morphological data can be provided for subsequent image reconstruction and modeling. These feature data can effectively distinguish different types of bone structures and provide fine-grained bone morphological information for subsequent analysis. Based on the analysis of regional bone density distribution and bone morphological feature data, the image is clustered and divided into bone structures. This process can automatically identify and distinguish large and small bone areas, effectively reducing manual intervention and improving the efficiency and accuracy of data processing. By dividing the bone structure cluster graph into structural labels, large bone structure images and small bone structure images are generated, so that subsequent data analysis and processing can be carried out based on a more detailed image hierarchy. For example, in multidimensional data reconstruction or image analysis, unnecessary complexity can be reduced and analytical accuracy improved. The entire process significantly improves data processing efficiency through automated image normalization, feature extraction, clustering, and label assignment. This reduces manual intervention and accelerates data preparation, making it particularly suitable for processing and analyzing large-scale image datasets. Accurate structural segmentation and feature extraction provide more refined input data for subsequent tasks such as image recognition and deep learning model training, thereby improving the accuracy of overall image analysis, especially when dealing with complex structures or minute details.

[0017] Preferably, analyzing the grayscale variation rate of the bone region in the orthopedic image set in step S12 includes:

[0018] When the grayscale standard deviation of the bone area in the orthopedic image set is less than 10 gray levels and the grayscale mean is higher than 180 gray levels, the bone density in the area is judged to be abnormally elevated, corresponding to a bone density greater than 1.4 g / cm 3 , obtain high-density bone area data;

[0019] When the grayscale standard deviation of the bone area in the orthopedic image set is between 10-30 grayscale levels and the grayscale mean is between 120-180 grayscale levels, the area is judged to be a normal bone density area, corresponding to a bone density range of 0.9-1.4 g / cm 3 , obtain normal density bone area data;

[0020] When the grayscale standard deviation of the bone area in the orthopedic image set is greater than 30 grayscale levels, or the grayscale mean is less than 120 grayscale levels, and the regional gradient direction texture variation rate exceeds 20%, the bone density of the area is judged to be abnormally reduced, corresponding to a bone density of less than 0.9g / cm 3 , obtain low-density bone area data;

[0021] The high-density bone area data, normal-density bone area data and low-density bone area data are integrated into the grayscale variation rate of the bone area.

[0022] This invention achieves precise bone density assessment by accurately classifying bone regions in orthopedic images (e.g., high-density, normal-density, and low-density regions) based on grayscale standard deviation and mean. This classification facilitates a more precise understanding of the spatial distribution of bone density, making it suitable for further analysis and modeling. Identifying high-density bone regions can effectively detect abnormalities such as bone hyperplasia. This classification method, by setting specific thresholds for the grayscale standard deviation and mean, enables more precise detection of these regions, further enhancing the reliability of image analysis. Identifying regions of normal bone density provides standardized reference data, facilitating comparative analysis with other density regions. This analysis clearly distinguishes between normal and abnormal bone density regions, helping doctors and researchers focus on areas with greater potential for problems. Detecting low-density bone regions clearly identifies areas of potential osteoporosis or other bone thinning within an image. By comprehensively considering the grayscale standard deviation, mean, and gradient-directed texture variation, low-density regions can be effectively identified, providing a highly accurate method for detecting abnormal bone density. By analyzing the grayscale variation rate in orthopedic images, misjudgments caused by image noise and bias can be reduced, especially when the grayscale standard deviation in low-density bone areas is large. This method helps eliminate unnecessary noise interference and improves the accuracy of overall image analysis. By integrating high-density, normal-density, and low-density bone areas, the grayscale variation rate of the generated bone areas provides clear density hierarchy data for subsequent image reconstruction, bone structure analysis, and three-dimensional modeling. This dataset can effectively support complex analysis tasks and provide reliable input for model training and decision-making. After generating high-density, normal-density, and low-density area data, this data can be used as key input for deep learning model training, further improving the performance and accuracy of orthopedic image recognition models, especially in areas with large density variations.

[0023] Preferably, in step S2, analyzing the bone layer thickness of the large bone structure image and performing physical layer thickness compensation reconstruction on the large bone structure image includes:

[0024] Perform three-dimensional block division on the large bone structure image to obtain large bone structure image block data at low resolution;

[0025] Perform super-resolution reconstruction on large bone structure image block data to generate high-resolution large bone structure images;

[0026] Extract the thickness variation of interlayer structure in sagittal and coronal directions in the large bone structure image under high resolution to obtain the bone layer thickness of the large bone structure image;

[0027] Performing slice thickness interpolation on large bone slice thickness parameter image data to generate slice thickness interpolation control parameter data;

[0028] Multi-neighborhood context-aware interpolation is performed on the high-resolution large bone structure image through slice thickness interpolation control parameter data to generate large bone middle virtual slice data.

[0029] The present invention can effectively improve the level of detail of the image by performing super-resolution reconstruction on the large bone structure image, making the tiny features and complex structures of the bone structure clearer. This process improves the image resolution and enhances subsequent analysis and accuracy. The interlayer structure thickness changes in the sagittal and coronal directions in the large bone structure image are extracted to provide more accurate data for the quantitative analysis of bone layer thickness. This analysis is crucial for evaluating bone density, bone morphology and bone health. Performing layer thickness interpolation on the large bone layer thickness parameter image data and generating layer thickness interpolation control parameter data can effectively compensate for the lack of layer thickness information caused by insufficient scanning resolution or other reasons. This compensation method helps to more accurately simulate the bone layer structure during the reconstruction process, thereby reducing errors. Multi-neighborhood context-aware interpolation of the high-resolution large bone structure image through layer thickness interpolation control parameter data not only improves the quality of image reconstruction, but also ensures the continuity and consistency of the bone structure between different regions. This process enhances the smoothness and transition effect of the image, helping to form a more natural and coherent image. By generating virtual slice data in the middle of the large bone, the cross-section of the large bone structure can be effectively simulated and reconstructed. Virtual slices provide more accurate slice data for subsequent 3D reconstruction, bone structure analysis, and deep learning model training, improving the image's operability and applicability. Super-resolution reconstruction and slice thickness interpolation can effectively reduce information loss caused by variations in image resolution or slice thickness, ensuring richer details in large bone structure images, thereby making bone analysis more accurate. Through automated image segmentation, super-resolution reconstruction, thickness extraction, and interpolation, data processing efficiency is greatly improved, manual intervention is reduced, and significant advantages are demonstrated when processing large datasets.

[0030] Preferably, performing detail-guided reconstruction on the ossicular structure image in step S2 includes:

[0031] Detecting the structural edge of the ossicular structure image to obtain initial edge gradient data of the ossicular structure;

[0032] The initial edge gradient data of the small bone structure is enhanced in the periosteal layer and fissure edge area to generate enhanced feature data of key detail areas;

[0033] Perform structural gradient fusion on the small bone structure image and the key detail area enhanced feature data to generate a detail gradient perception feature map;

[0034] The microstructure of the small bone structure image is restored using the detail gradient perception feature map to generate candidate data of the small bone micro fracture texture;

[0035] Texture consistency correction and noise suppression are performed on the candidate data of small bone micro-fracture texture to generate small bone fracture texture reconstruction data.

[0036] The present invention extracts significant boundary features of the periosteum and fissure areas in the small bone structure image through initial gradient detection of the structural edge, providing key positioning information for subsequent detail restoration and improving the accuracy of microstructure recognition. The periosteum layer and the fissure edge area are enhanced to effectively highlight pathological details such as fractures, cracks, and wear, thereby improving the recognition of these areas in the image and enhancing the learning effect of the reconstruction model on lesion-sensitive areas. The structural gradient fusion process combines the original image with the enhanced feature data to generate a detail gradient perception feature map, which significantly enhances the expression of details while maintaining the overall structural information and improves the overall layering of the image. Microstructure restoration of small bone images based on the detail gradient perception feature map can effectively reconstruct imperceptible structures such as tiny cracks and slight fractures, enhance the value of the image, and is of great significance in early lesion detection. After consistency correction and noise suppression processing, the candidate texture data can effectively eliminate texture discontinuity or artifact problems caused by image acquisition, compression or low resolution, ensuring that the generated small bone fracture texture reconstruction data has good continuity and realism.

[0037] Preferably, in step S2, performing image resolution enhancement on the bone structure segmentation image using the large bone middle virtual slice data and the small bone fracture texture reconstruction data includes:

[0038] Performing structural alignment on the bone structure segmentation image to generate a registered bone structure image;

[0039] Resample the scale of the virtual slice data in the middle of the large bone to generate a reference image of the bone structure scale;

[0040] Extracting texture features of small bone fracture texture reconstruction data to generate small bone crack texture feature images;

[0041] The registered bone structure images are fused by region based on the bone structure scale reference image and the small bone crack texture feature image to generate an enhanced medical clinical orthopedic image set.

[0042] The present invention realizes the effective coordination of the middle virtual slice of the large bone and the texture reconstruction data of the small bone fracture through structural alignment and multi-source image registration, ensuring the accurate spatial position and good tissue continuity of various bone structures in the enhanced image. Using the high-resolution virtual slice of the large bone and the texture reconstruction data of the small bone as a reference, the original bone structure image is enhanced at multiple scales, which significantly improves the resolution of the bone structure image and makes the detailed structures such as bone layers, bone sutures, and cracks more clearly visible. The regional feature fusion strategy is based on the characteristic differences between large bones and small bones, and adopts customized enhancement methods for different regions respectively, effectively retaining the overall morphology of the large bone while highlighting the micro-texture features of the small bones, thereby enhancing the difference and pertinence of the structural expression. The introduction of the small bone crack texture feature image improves the expression ability of edge details, solves the problems of blurred boundaries and missing cracks in the original image, and provides a reliable basis for micro-damage detection and fractures. A high-precision registration strategy is adopted in the structural alignment stage to ensure the accurate superposition of data from different sources in spatial coordinates, significantly improving the structural consistency and geometric fidelity in the image fusion process. By fusing and enhancing bone structure segmentation images from CT and MRI, we effectively integrate the structural advantages of different modalities, achieve complementary utilization of image information, and improve the usability of the final image. The enhanced medical clinical orthopedic image set has higher image quality and structural integrity, and can be used as high-quality input for training orthopedic dynamic image recognition models, improving the model's recognition accuracy and stability.

[0043] Preferably, step S3 includes the following steps:

[0044] Step S31: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set;

[0045] Step S32: performing tissue region segmentation processing on the medical clinical orthopedic image set through the decoder to generate mask data of bone tissue, muscle tissue and ligament regions;

[0046] Step S33: performing spatial position encoding on the mask data of the bone tissue, muscle tissue, and ligament regions to generate initial data of structured spatial nodes; performing graph convolution adjacency modeling on the initial data of structured spatial nodes to generate initial bone-muscle-ligament node connection atlas data;

[0047] Step S34: performing dynamic interaction modeling on the initial bone-muscle-ligament node connection map data to capture the interaction and spatial dependency between tissues and generate dynamic interaction relationship feature map data; performing spatial structure optimization and map topology constraint reconstruction on the initial bone-muscle-ligament node connection map data based on the dynamic interaction relationship feature map data to generate a bone-muscle-ligament spatial topology map;

[0048] Step S35: performing motion simulation on the medical clinical orthopedic image set according to the bone-muscle-ligament spatial topology map to generate bone motion simulation data.

[0049] This invention utilizes a dual-channel image decoder to decode bone and soft tissue (muscle and ligament) channels separately, significantly improving segmentation accuracy and effectively addressing the issues of blurred and overlapping tissue boundaries encountered in traditional methods. Through the tissue segmentation process, mask data for multiple tissue types, including bone, muscle, and ligament, is automatically generated, providing clear, structured basic region annotations for subsequent structural modeling and motion analysis. Spatial position encoding is performed on the segmented tissue mask data to extract spatial node information for each tissue. Combined with graph convolutional adjacency modeling, this enables efficient construction of an initial bone-muscle-ligament node connectivity map. A dynamic interaction modeling mechanism is introduced to effectively explore the coupling mechanisms and motion synergies between bones, muscles, and ligaments in different spatial structures, enhancing the model's spatiotemporal expressiveness. Using an interaction feature map to guide structural optimization and reconstruction of the topological map, a bone-muscle-ligament spatial topological map that more closely reflects the actual physiological structure is obtained, enhancing the model's adaptability to complex anatomical structures. The introduction of the topological map establishes spatial connectivity and dependencies between tissues, providing fundamental support for further functional modeling (such as stress propagation and injury prediction).

[0050] Preferably, step S31 includes the following steps:

[0051] Step S311: Construct a bone / soft tissue dual-channel encoding output interface to receive the bone tissue feature map and the soft tissue feature map output from the backbone network encoder respectively. The size is set to H / 16×W / 16×256, where H and W are the height and width of the input image respectively, in pixels, and 256 is the number of channels.

[0052] Step S312: Design a decoder with a symmetrical structure, which consists of 5 layers of deconvolution modules. Each layer of decoding module includes:

[0053] a dual-channel input fusion unit;

[0054] A two-channel parallel deconvolution operation with a kernel size of 3×3, a stride of 2, and padding of "same". The number of output channels is: 256 in the first layer, 128 in the second layer, 64 in the third layer, 32 in the fourth layer, and 16 in the fifth layer.

[0055] Each layer is followed by a batch normalization and ReLU activation function;

[0056] Step S313: At the decoding end, the reconstruction output layers for bone tissue and soft tissue are set respectively, and 1×1 convolution is used to compress the channels to 1 channel. A Sigmoid activation function is then applied to output a normalized image, and the output image size is restored to H×W×1, in pixels.

[0057] Step S314: Introduce a skip connection mechanism to input the feature map of each level in the encoder into the corresponding layer of the decoder through a layer-by-layer connection method, thereby obtaining a bone / soft tissue dual-channel image decoder.

[0058] The present invention realizes structured diversion of the feature map output by the backbone network encoder through the bone / soft tissue dual-channel encoding output interface, performs exclusive processing on bone tissue and soft tissue respectively, and effectively improves the discrimination of feature expression. A five-layer cascaded symmetrical decoding structure is adopted, and the image resolution is gradually restored through layer-by-layer deconvolution operations. Combined with the step-by-step channel number reduction strategy, the feature restoration process is made more hierarchical and information complete. Each layer of the decoding module contains a dual-channel input fusion unit and a parallel deconvolution structure, so that the bone and soft tissue can maintain high-fidelity processing of independent paths during the decoding process, and can also fuse shared information to enhance the expression ability of the model. 1×1 convolution is used for channel compression and combined with the Sigmoid activation function to achieve normalized output, which not only maintains the dimensional accuracy of the output image, but also ensures that the image intensity value is within the effective range, which is convenient for subsequent processing or visualization. The feature maps of each layer of the encoder are input into the decoder layer by layer through skip connections, which not only enhances the spatial feature transmission during the decoding process, but also effectively preserves the structural details and edge information of the original image, reducing the erosion of high-level semantic information on low-level spatial information. The dual-channel decoder, through its parallel architecture and scale synergy, is suitable for image reconstruction at varying resolutions, particularly for detailed restoration of structures where bone and soft tissue coexist. The decoder's layered design and clearly defined interfaces offer excellent scalability, allowing for easy integration of other structural analysis modules or image generation networks in subsequent workflows, adapting to diverse image processing scenarios.

[0059] Preferably, step S35 includes the following steps:

[0060] Step S351: extracting anatomical structure motion parameters from the bone-muscle-ligament spatial topology map to generate bone-muscle-ligament motion control factor data;

[0061] Step S352: performing dynamic driving modeling on the medical clinical orthopedic image set and the bone-muscle-ligament motion control factor data to generate initial bone displacement response data;

[0062] Step S353: performing physiological constraint modeling on the initial bone displacement response data, combining anatomical boundaries and tendon tension limits to generate constrained bone displacement field data;

[0063] Step S354: reconstructing the constrained skeletal displacement field data in time series, constructing a continuous inter-frame motion trajectory, and generating skeletal dynamic posture evolution data;

[0064] Step S355: Perform visual rendering and biomechanical consistency verification on the skeletal dynamic posture evolution data to generate skeletal motion simulation data.

[0065] The present invention accurately captures the linkage mechanisms and control parameters between anatomical structures through bone-muscle-ligament motion control factors extracted from spatial topology, providing a precise driving basis for subsequent motion modeling. By fusing image sets with motion control factors for dynamic modeling, the skeletal displacement response data more closely matches actual motion patterns, enhancing the spatial adaptability of the model-driven approach. By combining anatomical boundary constraints and tendon tension factors within physiological conditions, the skeletal motion path is effectively constrained, and the generated displacement field is more consistent with the structural physiological characteristics, avoiding unreasonable motion deformation. The time series reconstruction step enables continuous connection of skeletal dynamic postures across multiple frames, enabling the simulation process to evolve smoothly over time and be applied to construct visual scenes for dynamic processes. Visual rendering of the posture evolution data intuitively displays the dynamic behavior of the bone structure in different states, providing intuitive support for subsequent three-dimensional display and analysis. After generation, the skeletal motion simulation data undergoes biomechanical consistency verification, ensuring that the simulation results maintain consistency in structural coherence and physical and mechanical aspects, enhancing the data's versatility and credibility in various analytical tasks. Adding a dynamic modeling dimension on the basis of static topological graphs expands the application boundaries of graph structures under space-time coupling and provides a well-structured data-driven foundation for simulation and emulation systems.

[0066] Preferably, step S4 includes the following steps:

[0067] Step S41: labeling the medical clinical orthopedic image set with the bone motion simulation data to generate orthopedic image training label images; dividing the orthopedic image training label images into data sets to generate a model training set, a model verification set, and a model test set;

[0068] Step S42: Using the CNN neural network algorithm to perform model training on the model training set, and using the model validation set to perform accuracy evaluation on the trained model to obtain model performance evaluation index data;

[0069] Step S43: Optimize the model parameters of the model performance evaluation index data, import the optimized model parameters into the trained model for optimization training, and perform model testing iteration on the optimized trained model according to the model test set to generate an orthopedic dynamic image recognition model.

[0070] The present invention labels the original image set by combining it with bone motion simulation data, so that the generated training label images have dynamic structural semantic support, significantly improving the anatomical accuracy and boundary consistency of the labels, and providing a high-confidence labeling basis for supervised training. By scientifically dividing the training set, validation set, and test set, each subset has good representativeness and independence in terms of data distribution and structural performance, which is conducive to the robust training and generalization ability evaluation of the model. The convolutional neural network is used to deeply extract structural features, significantly enhancing the model's perception of complex bone structure morphology and change patterns, and adapting to the input features of diverse clinical images. Performance evaluation indicators are introduced during the model training process to monitor key indicators such as the model's accuracy, recall rate, and loss value in real time, providing reliable performance feedback for subsequent optimization. Parameter adjustment and model structure optimization are performed based on the evaluation indicator data, improving the model's fitting effect and convergence speed during the training process, while also enhancing its recognition stability in practical application scenarios. The model not only supports static orthopedic image recognition, but is also suitable for dynamic image sequence analysis. It has the ability to parse time series information and can be applied to scenarios such as bone evolution monitoring and rehabilitation behavior analysis. Step S4 established a complete AI training closed-loop process from data construction, model development, performance feedback to optimization iteration, promoting the efficient construction and continuous optimization of the clinical orthopedic image intelligent processing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A flowchart illustrating the steps of a method for creating an image recognition model based on the integration of clinical disease data;

[0072] Figure 2 for Figure 1 Detailed implementation steps of step S1 in FIG.

[0073] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0074] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0075] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0076] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0077] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0078] To achieve this, please refer to Figures 1 to 3 A method for creating an image recognition model based on clinical disease data integration, the method comprising the following steps:

[0079] Step S1: obtaining a medical clinical orthopedic image set; extracting bone density and bone morphology from the medical clinical orthopedic image set and performing bone structure segmentation on the medical clinical orthopedic image set to generate a bone structure segmentation image, wherein the bone structure segmentation image includes a large bone structure image and a small bone structure image;

[0080] Step S2: Analyze the bone layer thickness of the large bone structure image and perform physical layer thickness compensation reconstruction on the large bone structure image to generate large bone middle virtual slice data; perform detail-guided reconstruction on the small bone structure image to generate small bone fracture texture reconstruction data; use the large bone middle virtual slice data and small bone fracture texture reconstruction data to enhance the image resolution of the bone structure segmentation image to generate an enhanced medical clinical orthopedic image set;

[0081] Step S3: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set; performing graph convolution dynamic interaction modeling on the medical clinical orthopedic image set using the decoder to generate a bone-muscle-ligament spatial topology map; performing motion simulation on the medical clinical orthopedic image set based on the bone-muscle-ligament spatial topology map to generate bone motion simulation data;

[0082] Step S4: Perform model training on the medical clinical orthopedic image set using bone motion simulation data to generate an orthopedic dynamic image recognition model.

[0083] This invention significantly improves image clarity and detail preservation by reconstructing large bone structures using physical slice thickness compensation and small bone structures using fracture texture-guided reconstruction, providing high-quality input data for subsequent analysis. Image resolution enhancement, combining large bone intermediate virtual slice data with small bone texture reconstruction data, effectively improves image spatial resolution, particularly for the accurate representation of small bones and fine structures. Based on a dual-channel image decoder design, it enables joint perception and decoding of bone structure and soft tissue (such as muscle and ligaments), providing comprehensive data support for subsequent 3D modeling and motion simulation. Dynamic interactive graph convolution modeling constructs a bone-muscle-ligament spatial topology map, enhancing the ability to model associations between anatomical structures and facilitating accurate simulation of real-world physiological motion states. The resulting orthopedic dynamic image recognition model, trained on skeletal motion simulation data, possesses enhanced structural understanding and dynamic perception capabilities, making it particularly suitable for clinical applications such as preoperative planning, functional assessment, and lesion prediction. The overall process integrates multiple technologies, including image enhancement, structural reconstruction, topological modeling, and dynamic simulation, advancing the transition of medical image analysis from static recognition to dynamic understanding, improving the system's intelligent decision-making capabilities and clinical adaptability. Therefore, the present invention improves the accuracy of dynamic recognition of skeletal images by enhancing image resolution, multimodal data integration, motion simulation and optimization model training.

[0084] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a flow chart showing the steps of a method for creating an image recognition model based on clinical disease data integration according to the present invention. In this example, the method for creating an image recognition model based on clinical disease data integration includes the following steps:

[0085] Step S1: obtaining a medical clinical orthopedic image set; extracting bone density and bone morphology from the medical clinical orthopedic image set and performing bone structure segmentation on the medical clinical orthopedic image set to generate a bone structure segmentation image, wherein the bone structure segmentation image includes a large bone structure image and a small bone structure image;

[0086] Step S2: Analyze the bone layer thickness of the large bone structure image and perform physical layer thickness compensation reconstruction on the large bone structure image to generate large bone middle virtual slice data; perform detail-guided reconstruction on the small bone structure image to generate small bone fracture texture reconstruction data; use the large bone middle virtual slice data and small bone fracture texture reconstruction data to enhance the image resolution of the bone structure segmentation image to generate an enhanced medical clinical orthopedic image set;

[0087] Step S3: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set; performing graph convolution dynamic interaction modeling on the medical clinical orthopedic image set using the decoder to generate a bone-muscle-ligament spatial topology map; performing motion simulation on the medical clinical orthopedic image set based on the bone-muscle-ligament spatial topology map to generate bone motion simulation data;

[0088] Step S4: Perform model training on the medical clinical orthopedic image set using bone motion simulation data to generate an orthopedic dynamic image recognition model.

[0089] In this embodiment of the present invention, CT scans or X-ray images are obtained from a hospital or imaging database with a resolution of 512x512 pixels and a slice thickness of 1 mm for each image. Bone density is extracted using a threshold segmentation method, with a threshold of 300 HU (Hounsfield units). Values ​​above this threshold are considered bone tissue, while values ​​below this threshold are considered soft tissue. Morphological operations, such as erosion and dilation, are used to segment the images into large bone structures (e.g., femur, spine) and small bone structures (e.g., fingers, toes), ensuring that images of large and small bone structures are separated. Specifically, the average bone layer thickness for large bones is 10 mm, with a range of 8 mm to 12 mm. Based on this information, geometric algorithms (e.g., interpolation) are used to estimate the specific distribution of bone layers. Based on the known bone layer thickness, the intermediate layers of large bone structures are virtually compensated through interpolation or digital reconstruction. Specifically, each compensated slice thickness is 2 mm. For small bone structures (e.g., joints or fracture areas), image sharpening techniques (e.g., Laplacian filtering) are used for texture enhancement to increase image detail clarity. Super-resolution techniques (such as bilinear interpolation or bicubic interpolation) were used to upscale the image, specifically increasing the image resolution from 512x512 pixels to 1024x1024 pixels to enhance image detail. An image decoder was designed to process bone and soft tissue image data separately. Images were sized at 256x256 pixels, and bone and soft tissue were segmented using manually adjusted thresholds to ensure they were processed separately. Key spatial topological information was manually extracted based on the anatomical structures in the images (such as the boundaries of bones, muscles, and ligaments). Each bone was assumed to have a range of motion of 0° to 60°, and joint motion was predicted using geometric simulations. Joint motion simulations were performed with a 5-degree stride. The relative positions of bones and soft tissue were manually adjusted, and relative joint motion was calculated to generate motion data (e.g., joint angle changes). Image recognition techniques (such as template matching or feature point detection) were used to train a recognition model using the simulated bone motion data. The training batch size was 16 images, and the training cycle was 50 times. The image accuracy requirement was 90%. After training, the model is used to analyze new images, identifying and labeling abnormalities such as fractures and joint dislocations. Assume that the recognition accuracy reaches 95% and the recall rate is 92%.

[0090] As an example of the present invention, refer to Figure 2 As shown, in this example, step S1 includes:

[0091] Step S11: obtaining a medical clinical orthopedic image set, wherein the medical clinical orthopedic image set includes CT images and MRI images; performing image normalization on the CT images and the MRI images to generate a normalized orthopedic image set;

[0092] Step S12: Analyze the grayscale variation rate of the bone area in the orthopedic image set to obtain regional bone density distribution data;

[0093] Step S13: extracting bone morphological features from the orthopedic image data to generate bone morphological feature data;

[0094] Step S14: performing bone structure clustering on the medical clinical orthopedic image set according to the regional bone density distribution data and the bone morphological feature data to generate a bone structure clustering graph;

[0095] Step S15: using the bone structure clustering graph to perform structural label division on the medical clinical orthopedic image set to generate a bone structure division image, wherein the bone structure division image includes a large bone structure image and a small bone structure image.

[0096] In this embodiment of the present invention, CT scan images and MRI images are obtained from a hospital or imaging database. CT images have a resolution of 512x512 pixels, a slice thickness of 1mm, and are 16-bit grayscale images. MRI images have a resolution of 256x256 pixels, a slice thickness of 5mm, and are 12-bit grayscale images. The Hounsfield unit (HU) range of the CT images is normalized. The HU value range of CT images is -1000 to +3000. The images are normalized to a range of 0 to 1 using the formula: Normalized value = Original value - (-1000) ÷ 3000 - (-1000). MRI images, whose grayscale values ​​range from 0 to 4095, are similarly normalized to a range of 0 to 1 using the formula: Normalized value = Original value ÷ 4095. This normalized image has a uniform scale, facilitating subsequent analysis and processing. Edge detection is performed on the images using the Sobel operator, and the rate of change of grayscale values ​​in the orthopedic images is calculated. The grayscale variation rate is obtained by calculating the local gradient of the grayscale values ​​around each pixel in the image. For areas with large grayscale variations, it is indicated that the area is a bone area. By setting the grayscale variation rate threshold to 0.3, it is determined which areas belong to bones and which belong to soft tissues. Based on the analysis results of step 1, the average grayscale value of each bone area is calculated to represent the bone density of the area. For example, the grayscale value of a bone area is 0.75, while that of the soft tissue area is 0.2. This corresponds to a bone area with a density range of 700HU to 1500HU and a soft tissue density range of 0HU to 300HU in the CT image. Edge detection and morphological operations (such as erosion and dilation) are used to extract the morphological features of the bone to obtain the bone contour, area, volume, etc. For large bone structures (such as femur and spine), the length-to-width ratio of the bone is calculated to be 3:1 and the surface area is 1500cm. 2 , with a volume of 1200cm 3 For small bone structures (such as fingers and toes), the length-to-width ratio of the bone is calculated to be 1:1, and the surface area is 100 cm 2, with a volume of 30cm 3 . According to the regional bone density distribution data in step S12 and the bone morphological feature data in step S13, the K-means clustering algorithm (K-means) is used to divide the areas in the image into large bone structures and small bone structures. K-means clustering is used for grouping, and the K value is set to 2, and the image is divided into two categories: large bone structures and small bone structures. After clustering is completed, large bone structures (red) and small bone structures (green) are identified respectively by color mapping, and the generated bone structure clustering map can clearly show the areas of different bone structures. Use the clustering results for label division to identify the specific areas of large bone structures and small bone structures. By marking each area, the type of bone structure in the image can be obtained. For example, the area of ​​the large bone structure is 500cm 2 , the area of ​​the small bone structure is 50cm 2 , the area of ​​the soft tissue area is 150cm 2 Based on the label segmentation results, a bone structure segmentation image is generated. The image of the large bone structure area is marked in red, the small bone structure area is marked in green, and the soft tissue area is marked in blue. In this way, the bone structure segmentation image can clearly distinguish different types of bone structure areas.

[0097] Preferably, analyzing the grayscale variation rate of the bone region in the orthopedic image set in step S12 includes:

[0098] When the grayscale standard deviation of the bone area in the orthopedic image set is less than 10 gray levels and the grayscale mean is higher than 180 gray levels, the bone density in the area is judged to be abnormally elevated, corresponding to a bone density greater than 1.4 g / cm 3 , obtain high-density bone area data;

[0099] When the grayscale standard deviation of the bone area in the orthopedic image set is between 10-30 grayscale levels and the grayscale mean is between 120-180 grayscale levels, the area is judged to be a normal bone density area, corresponding to a bone density range of 0.9-1.4 g / cm 3 , obtain normal density bone area data;

[0100] When the grayscale standard deviation of the bone area in the orthopedic image set is greater than 30 grayscale levels, or the grayscale mean is less than 120 grayscale levels, and the regional gradient direction texture variation rate exceeds 20%, the bone density of the area is judged to be abnormally reduced, corresponding to a bone density of less than 0.9g / cm 3 , obtain low-density bone area data;

[0101] The high-density bone area data, normal-density bone area data and low-density bone area data are integrated into the grayscale variation rate of the bone area.

[0102] In an embodiment of the present invention, for each bone area in an orthopedic image, the grayscale value of the local area is first calculated. The block can be divided into 3x3 or 5x5 pixels, and then the grayscale mean and grayscale standard deviation of each block are calculated. The grayscale mean is the average brightness value of all pixels in the area, which is used to reflect the overall density of the bone. The grayscale standard deviation reflects the degree of grayscale change in the area and is used to analyze the uniformity of bone density. If the grayscale standard deviation of an area is less than 10 grayscale levels and the grayscale mean is higher than 180 grayscale levels, the bone density of the area is judged to be abnormally elevated. Usually, the bone density value corresponding to such an area is greater than 1.4g / cm 3 This area is usually displayed as a part with higher bone density, such as hard bone structure, and usually appears as a bright white area. When the grayscale standard deviation of the area is within the range of 10-30 grayscale levels and the grayscale mean is between 120-180 grayscale levels, it is judged as a normal bone density area. These areas usually correspond to bone density values ​​between 0.9-1.4g / cm 3 Such areas are generally presented as standard bone areas, displayed as medium gray areas. If the gray standard deviation of an area is greater than 30 gray levels, or the gray mean is less than 120 gray levels, and the gradient direction texture variation rate of the area exceeds 20%, the bone density of the area is judged to be abnormally low, usually with a bone density value of less than 0.9g / cm 3 These areas typically correspond to fragile bones and appear as dark gray or even black areas, representing the risk of osteoporosis or fractures. The data from the above three types of areas are integrated to generate a bone density distribution map. In this map, three different bone density areas can be marked with different colors: high-density bone areas are marked in red. Normal bone density bone areas are marked in green. Low-density bone areas are marked in blue. Combining the data from high-density, normal-density, and low-density areas generates an overall bone area grayscale variation rate map.

[0103] Preferably, in step S2, analyzing the bone layer thickness of the large bone structure image and performing physical layer thickness compensation reconstruction on the large bone structure image includes:

[0104] Perform three-dimensional block division on the large bone structure image to obtain large bone structure image block data at low resolution;

[0105] Perform super-resolution reconstruction on large bone structure image block data to generate high-resolution large bone structure images;

[0106] Extract the thickness variation of interlayer structure in sagittal and coronal directions in the large bone structure image under high resolution to obtain the bone layer thickness of the large bone structure image;

[0107] Performing slice thickness interpolation on large bone slice thickness parameter image data to generate slice thickness interpolation control parameter data;

[0108] Multi-neighborhood context-aware interpolation is performed on the high-resolution large bone structure image through slice thickness interpolation control parameter data to generate large bone middle virtual slice data.

[0109] In an embodiment of the present invention, a large bone structure image is divided into multiple small 3D image blocks through 3D segmentation. Each image block can be adjusted based on the image size and resolution, typically using 3x3x3 or 5x5x5 voxel blocks to ensure that each block represents a local area. For each large bone structure image block, super-resolution reconstruction techniques are used to improve the image resolution. High-resolution large bone structure images can be generated by employing deep learning or interpolation methods. Common super-resolution methods include convolutional neural networks (CNNs) or interpolation-based techniques such as bicubic interpolation or edge-enhanced interpolation. Inter-layer thickness variations in the sagittal and coronal planes of the high-resolution large bone structure image are extracted. For the sagittal plane, layer thickness measurements are performed along the anterior-posterior direction (left-right cutting direction); for the coronal plane, layer thickness measurements are performed along the superior-inferior direction (top-to-bottom cutting direction). The thickness variation between adjacent slices is calculated to obtain the bone layer thickness of each layer. For each slice, contour detection algorithms (such as edge detection or gradient detection) can be used to identify the boundaries of bone structures. The distance between two layers is then calculated to obtain the bone thickness. The extracted bone thickness data is interpolated to generate a smooth bone thickness image. Common interpolation methods include linear interpolation, spline interpolation, and cubic interpolation, and the interpolation algorithm can be selected as needed. Slice thickness interpolation control parameter data is generated from the interpolated bone thickness image. This data serves as input to control the interpolation process of the high-resolution image, enabling compensation and reconstruction based on variations in bone thickness. The control parameter data includes information such as the range and rate of variation of bone thickness, providing geometric rules to be followed during reconstruction. Finally, based on the slice thickness interpolation control parameter data, multi-neighborhood context-aware interpolation is performed on the high-resolution large bone structure image. This step primarily adjusts each pixel by considering information from neighboring regions. Multi-neighborhood context-aware interpolation can effectively improve image quality and maintain the coherence and integrity of image details. This method optimizes based on both local features and global structure of the image, thereby reducing artifacts and distortion caused by interpolation. After this interpolation process, the middle virtual slice data of the large bone is generated.

[0110] Preferably, performing detail-guided reconstruction on the ossicular structure image in step S2 includes:

[0111] Detecting the structural edge of the ossicular structure image to obtain initial edge gradient data of the ossicular structure;

[0112] The initial edge gradient data of the small bone structure is enhanced in the periosteal layer and fissure edge area to generate enhanced feature data of key detail areas;

[0113] Perform structural gradient fusion on the small bone structure image and the key detail area enhanced feature data to generate a detail gradient perception feature map;

[0114] The microstructure of the small bone structure image is restored using the detail gradient perception feature map to generate candidate data of the small bone micro fracture texture;

[0115] Texture consistency correction and noise suppression are performed on the candidate data of small bone micro-fracture texture to generate small bone fracture texture reconstruction data.

[0116] In embodiments of the present invention, images of ossicular structures are processed using image edge detection algorithms (such as Canny edge detection or the Sobel operator) to obtain preliminary information about structural edges within the image. This edge information includes the outline of the ossicular structure, small fissures, and the edges of the periosteum. The resulting edge data forms a set of gradient information representing the strength and direction of the structural edges within the image. Image enhancement techniques, such as local enhancement or adaptive filtering, focus on the periosteum and fissure regions of the ossicular structure. These regions are crucial for the details and morphological characteristics of the ossicles and therefore require particular attention. Different enhancement strategies are employed for the periosteum and fissure edge regions, such as enhancing edge sharpness and texture detail, to highlight key regions within the ossicular structure. The generated key detail region enhanced feature data can highlight these detailed regions and provide support for subsequent image reconstruction. Image fusion techniques are then used to combine the original ossicular structure image with the enhanced key detail region feature data. This process involves multi-scale image fusion, gradient fusion, or deep learning feature fusion. The resulting detail gradient-aware feature map incorporates both structural and detail information from the image. By enhancing gradient awareness of detail regions, the image reconstruction process ensures better restoration of the ossicular microstructure. Based on the detail gradient-aware feature map, an image reconstruction algorithm (such as a deep learning super-resolution network or a local texture restoration algorithm) is used to restore the microstructure of the small bone structure image. Through the restoration process, the fine fracture texture of the small bone (such as fine cracks and microscopic bone damage) is extracted from the image, generating candidate data for the small bone fine fracture texture. Texture consistency correction is performed on the restored fine fracture texture candidate data to ensure that the texture features remain spatially consistent and conform to the true representation of the small bone structure. Simultaneously, noise in the image, particularly high-frequency noise associated with the fracture texture, is removed using noise suppression algorithms (such as median filtering, total variation denoising, or deep denoising techniques). Ultimately, the resulting reconstructed small bone fracture texture data clearly displays the microcracks, fractures, and detailed texture of the small bone structure.

[0117] Of particular importance is the use of detail gradient-aware feature maps to perform microstructure restoration on small bone structure images, which also includes:

[0118] Extract the detail gradient features of the small bone structure image;

[0119] Calculate the gradient direction and amplitude of the small bone structure image based on the detail gradient features;

[0120] Localize microstructure regions based on gradient magnitude and direction;

[0121] Utilize the detail gradient feature map to locate the microstructure area and perform microstructure restoration;

[0122] Based on the results of microstructure restoration, candidate data of small bone microfracture texture are generated.

[0123] In this embodiment of the present invention, edge detection is used to extract detail gradient features from the original ossicular structure image. Image gradients are calculated using gradient operators such as Sobel and Canny, thereby extracting detailed features from the ossicular structure image, particularly surface contours, crack regions, and minor structural variations. Next, the image's gradient direction and magnitude are calculated based on these detail gradient features. The gradient direction reflects the direction of image variation, while the gradient magnitude indicates the intensity of structural variation, particularly the significance of surface cracks and minor irregularities. This information is used to determine the location of microstructural regions within the ossicular structure image. Based on the calculated gradient magnitude and direction, a threshold is set to filter out regions with significant variation. These regions typically correspond to cracks or subtle fracture textures in the ossicular structure. The gradient direction information is then used to precisely locate these regions. Next, these microstructural regions are restored using the detail gradient feature map. Image inpainting techniques, such as texture synthesis and image filtering, are used to restore the continuity of the crack or fracture regions and enhance the visual detail of the microstructural regions, thereby improving the accuracy and realism of the ossicular structure image. Finally, candidate data for ossicular microfracture textures is generated from the restored image.

[0124] Preferably, in step S2, performing image resolution enhancement on the bone structure segmentation image using the large bone middle virtual slice data and the small bone fracture texture reconstruction data includes:

[0125] Performing structural alignment on the bone structure segmentation image to generate a registered bone structure image;

[0126] Resample the scale of the virtual slice data in the middle of the large bone to generate a reference image of the bone structure scale;

[0127] Extracting texture features of small bone fracture texture reconstruction data to generate small bone crack texture feature images;

[0128] The registered bone structure images are fused by region based on the bone structure scale reference image and the small bone crack texture feature image to generate an enhanced medical clinical orthopedic image set.

[0129] In an embodiment of the present invention, image registration technology is used to structurally align the bone structure segmentation image with other bone structure images (such as small bone fracture texture reconstruction data or large bone intermediate virtual slice data). Common registration methods include feature-based registration (such as SIFT, SURF, and ORB feature matching) or intensity-based registration (such as mutual information or mean square error minimization). After registration, the resulting registered bone structure image ensures spatial consistency among all images, making subsequent regional feature fusion more accurate. The large bone intermediate virtual slice data provides information about the middle cross-section of the large bone, but this data differs from the target image in resolution. Therefore, rescaling is required to adjust its resolution to meet image enhancement requirements. Interpolation methods (such as bilinear interpolation, cubic spline interpolation, or convolutional neural network (CNN)) are used to resample the large bone intermediate virtual slice data to generate large bone structure scale reference images with a reference scale. These images serve as reference scales in subsequent enhancement processes, helping to improve image detail. Texture features are extracted from the small bone fracture texture reconstruction data, primarily including texture direction, frequency, and roughness. Small bone crack texture feature images can be generated using statistically based texture descriptors (such as gray-level co-occurrence matrices and local binary patterns (LBP)) or deep learning methods (such as high-level texture features extracted by convolutional neural networks) to demonstrate the detailed features of cracks in small bone structures. These feature images are then used to enhance the detail resolution of bone structures. The registered bone structure images are then subjected to region-by-region feature fusion using a bone structure scale reference image and a small bone crack texture feature image. The specific steps include: dividing the registered bone structure image into different regions (such as large bone regions, small bone regions, and crack regions). For each region, the details and texture features are enhanced using the bone structure scale reference image and the small bone crack texture feature image, respectively, to improve the image resolution and level of detail. Weighted fusion of regional features ensures that the enhanced images have high quality in terms of texture consistency and structural accuracy. Ultimately, the resulting enhanced medical clinical orthopedic image set exhibits higher resolution and clearer detail.

[0130] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0131] Step S31: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set;

[0132] Step S32: performing tissue region segmentation processing on the medical clinical orthopedic image set through the decoder to generate mask data of bone tissue, muscle tissue and ligament regions;

[0133] Step S33: performing spatial position encoding on the mask data of the bone tissue, muscle tissue, and ligament regions to generate initial data of structured spatial nodes; performing graph convolution adjacency modeling on the initial data of structured spatial nodes to generate initial bone-muscle-ligament node connection atlas data;

[0134] Step S34: performing dynamic interaction modeling on the initial bone-muscle-ligament node connection map data to capture the interaction and spatial dependency between tissues and generate dynamic interaction relationship feature map data; performing spatial structure optimization and map topology constraint reconstruction on the initial bone-muscle-ligament node connection map data based on the dynamic interaction relationship feature map data to generate a bone-muscle-ligament spatial topology map;

[0135] Step S35: performing motion simulation on the medical clinical orthopedic image set according to the bone-muscle-ligament spatial topology map to generate bone motion simulation data.

[0136] In the embodiment of the present invention, a dual-channel image decoder architecture based on a deep neural network is constructed to process the bone tissue channel and the soft tissue (muscle and ligament) channel separately. The decoder consists of the following modules: Encoder part: ResNet or EfficientNet is used to extract image features. Decoder part: Two decoding branches are designed respectively: Bone tissue channel: focuses on extracting high-density structures. Soft tissue channel: focuses on analyzing low-contrast muscle and ligament areas. Multi-scale fusion modules and attention mechanisms (such as SE modules or CBAM) are used to enhance the ability to perceive information between channels, and the output size is the same as the original. Figure 1Consistent bone and soft tissue images are obtained. The enhanced medical image set is fed into a dual-channel decoder to obtain feature probability maps for bone, muscle, and ligaments, respectively. A multi-label semantic segmentation strategy (such as a combined Softmax and Dice Loss) is used for training and inference to generate binary mask data for three regions: bone, muscle, and ligament. Morphological operations (such as opening and closing operations) are used to optimize mask boundaries and remove artifacts and isolated noise. Key points such as the geometric center, boundary points, and spatial contours of each tissue region are extracted. These points are encoded into vector representations using positional embedding methods (such as sinusoidal encoding or 3D coordinate normalization), resulting in an initial dataset of structured spatial nodes (bone-muscle-ligament node sets). Based on Euclidean distance, tissue adjacency probability, and common medical connectivity knowledge (such as muscle attachment to bone points), an adjacency matrix is ​​constructed and used in graph convolutional network (GCN) modeling to generate an initial bone-muscle-ligament node connectivity map. Dynamic graph neural networks (such as Dynamic-GCN and ST-GCN) are used to model the spatiotemporal relationships between nodes. Attention mechanisms are introduced to extract the interaction strength and direction between bones, muscles, and ligaments, generating dynamic interaction feature graph data. Based on the interaction features, the connection strength and directionality of nodes in the graph are adjusted. Topological constraints (e.g., a bone point must connect to two tendons) are introduced to optimize the graph structure. Graph pruning and structure completion algorithms are used to refine the connection graph, ultimately generating a bone-muscle-ligament spatial topology map. This topology map is then mapped back to the image coordinate system to serve as the skeleton structure for motion simulation. Physical modeling and deep learning methods are combined to model skeletal motion: multi-body dynamics and finite element simulations are applied to calculate skeletal motion, and graph neural network-based motion prediction models (such as GraphMotionNet) are used to infer dynamic behavior. Motion priors (e.g., bending, rotation, and tension) are applied to simulate coupled motion responses between tissues. The resulting skeletal motion simulation data is output, including images of skeletal deformation after motion; changes in key joint angles; and changes in muscle and ligament tension or relaxation.

[0137] Preferably, step S31 includes the following steps:

[0138] Step S311: Construct a bone / soft tissue dual-channel encoding output interface to receive the bone tissue feature map and the soft tissue feature map output from the backbone network encoder respectively. The size is set to H / 16×W / 16×256, where H and W are the height and width of the input image respectively, in pixels, and 256 is the number of channels.

[0139] Step S312: Design a decoder with a symmetrical structure, which consists of 5 layers of deconvolution modules. Each layer of decoding module includes:

[0140] a dual-channel input fusion unit;

[0141] A two-channel parallel deconvolution operation with a kernel size of 3×3, a stride of 2, and padding of "same". The number of output channels is: 256 in the first layer, 128 in the second layer, 64 in the third layer, 32 in the fourth layer, and 16 in the fifth layer.

[0142] Each layer is followed by a batch normalization and ReLU activation function;

[0143] Step S313: At the decoding end, the reconstruction output layers for bone tissue and soft tissue are set respectively, and 1×1 convolution is used to compress the channels to 1 channel. A Sigmoid activation function is then applied to output a normalized image, and the output image size is restored to H×W×1, in pixels.

[0144] Step S314: Introduce a skip connection mechanism to input the feature map of each level in the encoder into the corresponding layer of the decoder through a layer-by-layer connection method, thereby obtaining a bone / soft tissue dual-channel image decoder.

[0145] In this embodiment of the present invention, a branching structure is implemented at the end of a backbone network (such as ResNet-50 or EfficientNet) to divide the output features into two parts: a bone tissue feature map channel and a soft tissue feature map channel. The corresponding interface output format is uniformly set to H / 16×W / 16×256, where H is the input image height (in pixels); W is the input image width; and 256 is the number of channels, shared by the two types of tissue features. Logical separation or channel splitting (128+128) is used to produce a dual-channel output. This interface is responsible for providing the decoder with clearly structured and organized input features. A five-layer symmetric deconvolutional decoding architecture is constructed, with each layer comprising the following modules: receiving bone and soft tissue feature maps; performing weighted fusion of the two channels using a channel attention mechanism (such as SE or CBAM); and preserving channel independence to avoid feature interference. Each layer uses a 3×3 deconvolution kernel with a stride of 2 and "same" padding to ensure gradual upsampling. The number of output channels is set as follows: 256 for the first layer, 128 for the second layer, 64 for the third layer, 32 for the fourth layer, and 16 for the fifth layer. After each deconvolution operation, batch normalization (BN) and ReLU activation are performed on both channels independently to improve network convergence and prevent gradient vanishing. A bone tissue output layer and a soft tissue output layer are respectively provided at the end of the decoder. Each output layer consists of a 1×1 convolution operation to reduce the number of channels from 16 to 1, a sigmoid activation function to normalize the output to the range [0, 1], and an output image of size H×W×1 (pixels). The outputs represent bone tissue probability maps and soft tissue (muscle, ligament) probability maps, respectively. Feature maps (e.g., H / 2, H / 4, H / 8, H / 16) are extracted from the output of each stage of the backbone encoder. These feature maps are fed into the decoder's deconvolution module at the corresponding scale via layer-by-layer skip connections. 1×1 convolution or convolutional fusion units are used to perform channel compression on the skip-connected features. The encoder features are then fused with the current decoder layer output using the concat or add method. The skip connection mechanism improves the ability to recover local structures such as edges, textures, and details, and is a key improvement to the U-Net architecture, resulting in a dual-channel bone / soft tissue image decoder.

[0146] Preferably, step S35 includes the following steps:

[0147] Step S351: extracting anatomical structure motion parameters from the bone-muscle-ligament spatial topology map to generate bone-muscle-ligament motion control factor data;

[0148] Step S352: performing dynamic driving modeling on the medical clinical orthopedic image set and the bone-muscle-ligament motion control factor data to generate initial bone displacement response data;

[0149] Step S353: performing physiological constraint modeling on the initial bone displacement response data, combining anatomical boundaries and tendon tension limits to generate constrained bone displacement field data;

[0150] Step S354: reconstructing the constrained skeletal displacement field data in time series, constructing a continuous inter-frame motion trajectory, and generating skeletal dynamic posture evolution data;

[0151] Step S355: Perform visual rendering and biomechanical consistency verification on the skeletal dynamic posture evolution data to generate skeletal motion simulation data.

[0152] In this embodiment of the present invention, key anatomical nodes (bone anchor points, joint mid-axes, tendon attachment points, and ligament tension nodes) are identified based on a constructed bone-muscle-ligament spatial topology map. Edge attributes (length, direction, tension value, etc.) are analyzed and, combined with the spatial topology, the following types of motion control factors are extracted: rigid transformation factor for the skeleton; muscle activation parameter; and ligament elastic constraint. These factors are standardized and encoded, outputting bone-muscle-ligament motion control factor data as a set of driving input parameters. Bone tissue masks and tendon and ligament regions from a medical clinical orthopedic image set are mapped to the spatial topology and voxel-level registration is performed. A driving network based on finite element modeling (FEM) or a motion graph model is constructed. Using the bone-muscle-ligament motion control factor data as driving input, external forces, internal forces, and active muscle traction are applied to the anatomical structure to simulate motion. The simulation output is initial skeletal displacement response data: a three-dimensional rigid body node displacement vector field in the format V(x, y, z, t). Based on this initial response data, physiological constraint models are introduced: anatomical boundary constraints (such as joint motion angle constraints and bone contact surface constraints); and tendon tension constraints (based on muscle traction force models and maximum contraction length). Optimization solvers (such as L-BFGS and ADMM) are used to correct the displacement field to ensure that the simulation does not violate physiological structures. Constrained skeletal displacement field data is generated that meets the physical properties of real human anatomy. The constrained displacement field is reconstructed by interpolation along the time dimension (using Bezier curve interpolation, B-splines, etc.), achieving the conversion from static response to dynamic sequence. A bone tissue motion trajectory map is constructed for consecutive frames T1, T2, ..., Tn, and the dynamic skeletal posture evolution data is output: recording the 3D position, angle changes, and structural interaction trajectory of the bone tissue at each time frame. Use 3D visualization engines (such as VTK, Blender, and Unity3D) to reconstruct skeletal animation from dynamic evolution data. These tools support various angles and rendering modes (such as transparent skeletons and force field color overlays), motion path overlays, and real-time node interaction trajectory labeling. Biomechanical consistency verification is performed by comparing with real clinical motion capture data (such as XROMM and MOcap), verifying the accuracy of musculoskeletal coordination, and analyzing joint stability. Finally, skeletal motion simulation data is output.

[0153] It is particularly important that step S352 further includes the following steps:

[0154] Step S3521: performing three-dimensional bone reconstruction on the medical clinical orthopedic image set to generate three-dimensional bone structure reconstruction data;

[0155] Step S3522: extracting power source parameters from the three-dimensional reconstruction data of the bone structure, wherein the power source parameters include muscle force distribution, contraction direction, and joint angle change curve;

[0156] Step S3523: performing finite element meshing on the three-dimensional reconstruction data of the bone structure based on the muscle force distribution, contraction direction, and joint angle change curve, dividing the three-dimensional bone structure into regular / irregular units, and generating finite element analysis data of the bone structure;

[0157] Step S3524: Apply load boundaries according to the power source parameters and the bone structure finite element analysis data to obtain bone structure dynamic boundary data; perform time step simulation on the bone structure dynamic boundary data to generate initial bone displacement response data.

[0158] In this embodiment of the present invention, in step S3521, a medical clinical orthopedic image set is subjected to three-dimensional bone reconstruction. This process uses high-resolution CT or MRI images as input, extracts bone tissue regions using a deep learning-based image segmentation network (such as 3D U-Net), and then uses mesh reconstruction algorithms such as Marching Cubes or Poisson Surface Reconstruction to convert the bone structure surface contour into a three-dimensional mesh model. The output format can be STL, OBJ, or PLY, obtaining anatomically accurate three-dimensional reconstruction data of the bone structure. In step S3522, power source parameters are extracted from the three-dimensional bone structure reconstruction data. This parameter set includes muscle force distribution, muscle contraction direction, and joint angle change curves. Muscle force distribution is calculated based on muscle physiological parameters and motion databases using the Hill model or Zajac muscle model, and fitted in conjunction with musculoskeletal simulation platforms such as OpenSim. The direction of muscle contraction is determined by the position vectors of the tendon origin and insertion and the bone attachment point, forming a three-dimensional direction vector matrix. The joint angle change curve is spline fitted using physiological joint change trajectories extracted from clinical gait acquisition systems, motion capture systems, or existing databases to obtain the angle-time function relationship. Next, in step S3523, based on the aforementioned dynamic source parameters, the 3D bone model undergoes finite element meshing. Using engineering modeling tools such as Gmsh, HyperMesh, or Abaqus CAE, the bone is divided into regular hexahedral or irregular tetrahedral meshes. In conjunction with the bone density image, spatial variational meshing strategies are applied to different anatomical regions (e.g., mesh refinement in cortical bone regions and moderate simplification in cancellous bone regions). Element side lengths are controlled between 0.5 and 2 mm. A material property table is generated based on clinically measured or literature-provided biomechanical parameters such as bone elastic modulus and Poisson's ratio. Finite element analysis data for the bone structure is constructed, including node information, element topology, and material parameters. The process then proceeds to step S3524, where dynamic boundary conditions are applied to the finite element bone model. First, a load vector is applied to the tendon attachment point. Based on the aforementioned muscle force distribution, the vector is decomposed and concentrated or distributed forces are applied to the corresponding nodes. Second, based on the joint angle change function, angle-driven boundary conditions are set within the joint motion region to simulate physiological motion processes such as flexion, extension, and rotation. Third, fixed, hinged, or sliding boundaries are set in conjunction with motion function constraints and reference structures such as the pelvis and spine to ensure the stability and physiological rationality of the simulation system, thereby forming complete dynamic boundary data for the bone structure. After the boundary setting is completed, a dynamic simulation algorithm is used for time-series simulation processing. Explicit dynamics algorithms (such as the Euler explicit method) can be used to handle high-speed motion responses, or more stable implicit algorithms (such as the Newmark-β method) can be used to simulate slowly varying joint motion.The simulation time step Δt is generally controlled between 0.01 and 0.1 seconds, and the total simulation time is set to 1 to 5 seconds depending on the length of the action cycle. By solving the node displacement, rotation angle and bone stress and strain response step by step, the initial bone displacement response data is finally output. The data structure includes the three-dimensional coordinate displacement vector of each node of the bone structure in each time frame, angle change information and force response indicators of key parts, providing basic simulation data support with high temporal and spatial resolution for subsequent physiological constraint modeling, continuous frame motion trajectory reconstruction and mechanical consistency analysis, while being highly realistic and interpretable.

[0159] Preferably, step S4 includes the following steps:

[0160] Step S41: labeling the medical clinical orthopedic image set with the bone motion simulation data to generate orthopedic image training label images; dividing the orthopedic image training label images into data sets to generate a model training set, a model verification set, and a model test set;

[0161] Step S42: Using the CNN neural network algorithm to perform model training on the model training set, and using the model validation set to perform accuracy evaluation on the trained model to obtain model performance evaluation index data;

[0162] Step S43: Optimize the model parameters of the model performance evaluation index data, import the optimized model parameters into the trained model for optimization training, and perform model testing iteration on the optimized trained model according to the model test set to generate an orthopedic dynamic image recognition model.

[0163] In this embodiment of the present invention, the skeletal motion simulation data generated in the previous step (including key skeletal node trajectories and tendon dynamic deformation) is automatically or semi-automatically annotated with structural semantics for the corresponding medical clinical orthopedic image set. Annotation targets include bone tissue contours, motion trend arrows, muscle region annotations, and kinematic labels. The output is a labeled orthopedic image training set in a unified format consisting of a registered RGB image and a label mask (e.g., NIfTI or PNG+JSON). The image set is divided into three categories: a training set (70%), a validation set (15%), and a test set (15%). Stratified sampling ensures representation of different anatomical structures and motion patterns. A multi-input dynamic convolutional neural network (CNN) architecture is employed, with a time series image perception network (e.g., a ConvLSTM+3D-CNN fusion architecture) as the backbone. Continuous frame images are input to automatically learn dynamic bone patterns. Loss function: Dice loss + Focal Loss weighted combination; Optimizer: Adam with learning rate decay; Training epochs: 100-200, batch size: 48. Evaluate the model at the end of each epoch using the validation set, calculating the following performance metrics: Accuracy; Intersection over Union (IoU); Dice coefficient; and Hausdorff distance (HD), used for bone edge fitting. Based on these metrics, analyze bottlenecks (e.g., low accuracy in identifying small bone cracks) and adjust network structure or parameters: add an attention mechanism module (e.g., SE module, CBAM); adjust the convolution kernel size or increase the channel width; and perform transfer learning fine-tuning. Import the optimized parameters into the model for optimization training. A complete test is conducted on the model test set, and the recognition results are output for comparison and analysis with the label image. Visualization techniques (such as Grad-CAM and dynamic heat maps) are used to assist in verifying the anatomical consistency of the model's focus area. If the test indicators still do not meet the set standards, the system can be looped back to the optimization stage for further fine-tuning to generate an orthopedic dynamic image recognition model.

[0164] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0165] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for creating an image recognition model based on clinical disease data integration, characterized in that: The following steps are involved: Step S1: obtaining a medical clinical orthopedic image set; extracting bone density and bone morphology from the medical clinical orthopedic image set and performing bone structure segmentation on the medical clinical orthopedic image set to generate a bone structure segmentation image, wherein the bone structure segmentation image includes a large bone structure image and a small bone structure image; Step S2: Analyze the bone layer thickness of the large bone structure image and perform physical layer thickness compensation reconstruction on the large bone structure image to generate large bone middle virtual slice data; perform detail-guided reconstruction on the small bone structure image to generate small bone fracture texture reconstruction data; use the large bone middle virtual slice data and small bone fracture texture reconstruction data to enhance the image resolution of the bone structure segmentation image to generate an enhanced medical clinical orthopedic image set; Step S3: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set; performing graph convolution dynamic interaction modeling on the medical clinical orthopedic image set through the decoder to generate a bone-muscle-ligament spatial topology map; Perform motion simulation on a medical clinical orthopedic image set based on the bone-muscle-ligament spatial topology map to generate bone motion simulation data; Step S4: Perform model training on the medical clinical orthopedic image set using bone motion simulation data to generate an orthopedic dynamic image recognition model.

2. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining a medical clinical orthopedic image set, wherein the medical clinical orthopedic image set includes CT images and MRI images; performing image normalization on the CT images and the MRI images to generate a normalized orthopedic image set; Step S12: Analyze the grayscale variation rate of the bone area in the orthopedic image set to obtain regional bone density distribution data; Step S13: extracting bone morphological features from the orthopedic image data to generate bone morphological feature data; Step S14: performing bone structure clustering on the medical clinical orthopedic image set according to the regional bone density distribution data and the bone morphological feature data to generate a bone structure clustering graph; Step S15: using the bone structure clustering graph to perform structural label division on the medical clinical orthopedic image set to generate a bone structure division image, wherein the bone structure division image includes a large bone structure image and a small bone structure image.

3. The method for creating an image recognition model based on clinical disease data integration according to claim 2, characterized in that: Analyzing the grayscale variation rate of the bone region in the orthopedic image set in step S12 includes: When the grayscale standard deviation of the bone area in the orthopedic image set is less than 10 gray levels and the grayscale mean is higher than 180 gray levels, the bone density in the area is judged to be abnormally elevated, corresponding to a bone density greater than 1.4 g / cm 3 , obtain high-density bone area data; When the grayscale standard deviation of the bone area in the orthopedic image set is between 10-30 grayscale levels and the grayscale mean is between 120-180 grayscale levels, the area is judged to be a normal bone density area, corresponding to a bone density range of 0.9-1.4 g / cm 3 , obtain normal density bone area data; When the grayscale standard deviation of the bone area in the orthopedic image set is greater than 30 grayscale levels, or the grayscale mean is less than 120 grayscale levels, and the regional gradient direction texture variation rate exceeds 20%, the bone density of the area is judged to be abnormally reduced, corresponding to a bone density of less than 0.9g / cm 3 , obtain low-density bone area data; The high-density bone area data, normal-density bone area data and low-density bone area data are integrated into the grayscale variation rate of the bone area.

4. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: Analyzing the bone layer thickness of the large bone structure image and performing physical layer thickness compensation reconstruction on the large bone structure image in step S2 includes: Perform three-dimensional block division on the large bone structure image to obtain large bone structure image block data at low resolution; Perform super-resolution reconstruction on large bone structure image block data to generate high-resolution large bone structure images; Extract the thickness variation of interlayer structure in sagittal and coronal directions in the large bone structure image under high resolution to obtain the bone layer thickness of the large bone structure image; Performing slice thickness interpolation on large bone slice thickness parameter image data to generate slice thickness interpolation control parameter data; Multi-neighborhood context-aware interpolation is performed on the high-resolution large bone structure image through slice thickness interpolation control parameter data to generate large bone middle virtual slice data.

5. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: The detail-guided reconstruction of the ossicular structure image in step S2 includes: Detecting the structural edge of the ossicular structure image to obtain initial edge gradient data of the ossicular structure; The initial edge gradient data of the small bone structure is enhanced in the periosteal layer and fissure edge area to generate enhanced feature data of key detail areas; Perform structural gradient fusion on the small bone structure image and the key detail area enhanced feature data to generate a detail gradient perception feature map; The microstructure of the small bone structure image is restored using the detail gradient perception feature map to generate candidate data of the small bone micro fracture texture; Texture consistency correction and noise suppression are performed on the candidate data of small bone micro-fracture texture to generate small bone fracture texture reconstruction data.

6. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: In step S2, the image resolution enhancement of the bone structure segmentation image using the large bone middle virtual slice data and the small bone fracture texture reconstruction data includes: Performing structural alignment on the bone structure segmentation image to generate a registered bone structure image; Resample the scale of the virtual slice data in the middle of the large bone to generate a reference image of the bone structure scale; Extracting texture features of small bone fracture texture reconstruction data to generate small bone crack texture feature images; The registered bone structure images are fused by region based on the bone structure scale reference image and the small bone crack texture feature image to generate an enhanced medical clinical orthopedic image set.

7. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: designing a bone / soft tissue dual-channel image decoder based on the enhanced medical clinical orthopedic image set; Step S32: performing tissue region segmentation processing on the medical clinical orthopedic image set through the decoder to generate mask data of bone tissue, muscle tissue and ligament regions; Step S33: performing spatial position encoding on the mask data of the bone tissue, muscle tissue, and ligament regions to generate initial data of structured spatial nodes; performing graph convolution adjacency modeling on the initial data of structured spatial nodes to generate initial bone-muscle-ligament node connection atlas data; Step S34: performing dynamic interaction modeling on the initial bone-muscle-ligament node connection map data to capture the interaction and spatial dependency between tissues and generate dynamic interaction relationship feature map data; performing spatial structure optimization and map topology constraint reconstruction on the initial bone-muscle-ligament node connection map data based on the dynamic interaction relationship feature map data to generate a bone-muscle-ligament spatial topology map; Step S35: performing motion simulation on the medical clinical orthopedic image set according to the bone-muscle-ligament spatial topology map to generate bone motion simulation data.

8. The method for creating an image recognition model based on clinical disease data integration according to claim 7, characterized in that: Step S31 includes the following steps: Step S311: Construct a bone / soft tissue dual-channel encoding output interface to receive the bone tissue feature map and the soft tissue feature map output from the backbone network encoder respectively. The size is set to H / 16×W / 16×256, where H and W are the height and width of the input image respectively, in pixels, and 256 is the number of channels. Step S312: Design a decoder with a symmetrical structure, which consists of 5 layers of deconvolution modules. Each layer of decoding module includes: a dual-channel input fusion unit; A two-channel parallel deconvolution operation with a kernel size of 3×3, a stride of 2, and padding of "same". The number of output channels is: 256 in the first layer, 128 in the second layer, 64 in the third layer, 32 in the fourth layer, and 16 in the fifth layer. Each layer is followed by a batch normalization and ReLU activation function; Step S313: At the decoding end, the reconstruction output layers for bone tissue and soft tissue are set respectively, and 1×1 convolution is used to compress the channels to 1 channel. A Sigmoid activation function is then applied to output a normalized image, and the output image size is restored to H×W×1, in pixels. Step S314: Introduce a skip connection mechanism to input the feature map of each level in the encoder into the corresponding layer of the decoder through a layer-by-layer connection method, thereby obtaining a bone / soft tissue dual-channel image decoder.

9. The method for creating an image recognition model based on clinical disease data integration according to claim 7, characterized in that: Step S35 includes the following steps: Step S351: extracting anatomical structure motion parameters from the bone-muscle-ligament spatial topology map to generate bone-muscle-ligament motion control factor data; Step S352: performing dynamic driving modeling on the medical clinical orthopedic image set and the bone-muscle-ligament motion control factor data to generate initial bone displacement response data; Step S353: performing physiological constraint modeling on the initial bone displacement response data, combining anatomical boundaries and tendon tension limits to generate constrained bone displacement field data; Step S354: reconstructing the constrained skeletal displacement field data in time series, constructing a continuous inter-frame motion trajectory, and generating skeletal dynamic posture evolution data; Step S355: Perform visual rendering and biomechanical consistency verification on the skeletal dynamic posture evolution data to generate skeletal motion simulation data.

10. The method for creating an image recognition model based on clinical disease data integration according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: labeling the medical clinical orthopedic image set with the bone motion simulation data to generate orthopedic image training label images; dividing the orthopedic image training label images into data sets to generate a model training set, a model verification set, and a model test set; Step S42: Using the CNN neural network algorithm to perform model training on the model training set, and using the model validation set to perform accuracy evaluation on the trained model to obtain model performance evaluation index data; Step S43: Optimize the model parameters of the model performance evaluation index data, import the optimized model parameters into the trained model for optimization training, and perform model testing iteration on the optimized trained model according to the model test set to generate an orthopedic dynamic image recognition model.

Images