Knee joint prosthesis evaluation method and system

Abstract

The invention discloses a knee joint prosthesis evaluation method and system, and the method comprises the steps: carrying out the multi-modal data collection according to the biomechanical performance test data of a knee joint prosthesis, the medical image data and the postoperative motion capture data of a patient, and carrying out the data fusion through a multi-modal alignment network based on deep learning, a fused multi-modal data set is obtained; inputting the multi-modal data set into an adaptive convolutional neural network, and extracting mechanical characteristics of the knee joint prosthesis under a dynamic load to obtain a dynamic mechanical characteristic matrix; and inputting the dynamic mechanical characteristic matrix into a performance evaluation model based on a multi-scale attention mechanism, performing prosthesis performance quantitative analysis, and generating a performance evaluation report. By utilizing the embodiment of the invention, comprehensive quantitative analysis of the prosthesis performance can be realized, and scientificity and accuracy of prosthesis design are improved.

Description

Technical Field

[0001] This invention belongs to the field of joint prosthesis technology, and in particular to a method and system for evaluating knee joint prostheses. Background Technology

[0002] Knee replacement surgery has become an effective treatment for severe knee joint diseases such as osteoarthritis and rheumatoid arthritis, aiming to restore patients' motor function and quality of life. However, with the continuous advancement of prosthetic materials and design technologies, how to comprehensively and scientifically evaluate knee prostheses to ensure their superior performance and safety has become an important issue in current medical research and clinical applications.

[0003] Existing methods for evaluating knee prostheses often rely on a single assessment metric, typically focusing on biomechanical performance testing and postoperative clinical outcomes, while lacking comprehensive analysis of data from different sources. This approach struggles to fully reflect the prosthesis's performance in actual use, particularly stress distribution and motion under dynamic loads. Furthermore, the effective integration of medical imaging data with postoperative motion capture data remains a significant technological challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for evaluating knee joint prostheses, in order to overcome the shortcomings of the prior art, enable comprehensive quantitative analysis of prosthesis performance, and improve the scientificity and accuracy of prosthesis design.

[0005] One embodiment of this application provides a method for evaluating knee joint prostheses, the method comprising: Based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, multimodal data were collected, and a deep learning-based multimodal alignment network was used for data fusion to map data from different sources into a unified spatiotemporal coordinate system, resulting in a fused multimodal dataset. The multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. The dynamic mechanical feature matrix is ​​input into a performance evaluation model based on a multi-scale attention mechanism to perform quantitative analysis of prosthesis performance. The performance evaluation model evaluates the stability, wear resistance, and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0006] Optionally, the process involves collecting multimodal data based on biomechanical performance test data of the knee prosthesis, medical imaging data, and postoperative motion capture data of the patient. A deep learning-based multimodal alignment network is then used for data fusion, mapping data from different sources to a unified spatiotemporal coordinate system to obtain a fused multimodal dataset, including: Based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, an edge computing-based data acquisition framework is adopted to acquire multimodal data in real time. A lightweight data caching mechanism is used to ensure the real-time and continuous nature of data acquisition. For multimodal data, a deep learning-based format recognition model is used to automatically identify the format types of different data sources. An adaptive data cleaning algorithm is used to filter noise and impute missing values ​​in the data to generate a preliminary standardized dataset. For the initial standardized dataset, a deep learning-based multimodal alignment network is used to map data from different sources to a unified spatiotemporal coordinate system. Through timestamp alignment algorithms and spatial registration techniques, the temporal and spatial differences between data are eliminated, generating an initial aligned multimodal dataset. For the initially aligned multimodal dataset, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse the features of biomechanical performance data, medical imaging data and motion capture data. Through cross-modal attention mechanism, the correlation between different data sources is captured to generate the fused multimodal dataset.

[0007] Optionally, the step of inputting the multimodal dataset into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load, wherein the adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix, including: For multimodal datasets, a feature extraction method based on adaptive convolutional neural networks is used to extract stress-strain features from biomechanical performance data, structural features from medical imaging data, and motion trajectory features from motion capture data. Through multi-scale convolutional kernels, mechanical features at different scales are captured to generate preliminary feature representations. For the initial feature representation, a dynamic modeling method based on temporal convolution modules is adopted. By combining the load changes of the prosthesis at different motion stages, the temporal features of stress distribution and strain change are captured. Through the sliding window mechanism, short-term fluctuation and long-term trend features are extracted to generate dynamic mechanical feature representation. For the representation of dynamic mechanical features, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse stress distribution features, strain change features and motion trajectory features. Through cross-modal attention mechanism, the correlation between different features is captured to generate a fused dynamic mechanical feature matrix. For the fused dynamic mechanical feature matrix, a Bayesian optimization-based calibration method is adopted. Combining the material properties of the prosthesis and the actual use scenario, the feature weights are dynamically adjusted. Through regularization constraints, overfitting is prevented, and the final dynamic mechanical feature matrix is ​​generated.

[0008] Optionally, the dynamic mechanical feature matrix is ​​input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance. The performance evaluation model, by combining local and global features, evaluates the stability, wear resistance, and biocompatibility of the prosthesis and generates a performance evaluation report, including: For the dynamic mechanical feature matrix, a feature extraction method based on multi-scale convolutional neural networks is adopted to extract the local features of the prosthesis at the microscale. Through adaptive convolution kernels, local mechanical properties at different scales are captured to generate local feature representations. For the dynamic mechanical feature matrix, a global feature extraction method based on graph convolutional networks is adopted to abstract the overall mechanical behavior of the prosthesis into a graph structure, extract global features, capture global mechanical properties through multi-layer graph convolution operations, and generate global feature representations. For local and global feature representations, a feature fusion method based on multi-scale attention mechanism is adopted to perform weighted fusion of local and global features. Through cross-scale attention mechanism, the correlation between local and global features is captured to generate fused multi-scale feature representations. The fused multi-scale feature representation is used to perform quantitative analysis by adopting a performance evaluation model based on deep neural networks, combined with the stability, wear resistance and biocompatibility indicators of the prosthesis. Through the interpretability module in the performance evaluation model, a performance evaluation report containing quantitative results and evaluation basis is generated.

[0009] Another embodiment of this application provides a knee joint prosthesis evaluation system, the system comprising: The acquisition module is used to collect multimodal data based on biomechanical performance test data of knee joint prostheses, medical imaging data and postoperative motion capture data of patients, and to perform data fusion using a deep learning-based multimodal alignment network to map data from different sources into a unified spatiotemporal coordinate system to obtain a fused multimodal dataset. The extraction module is used to input the multimodal dataset into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. The analysis module is used to input the dynamic mechanical feature matrix into a performance evaluation model based on a multi-scale attention mechanism to perform quantitative analysis of prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0010] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0011] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0012] Compared with existing technologies, the present invention provides a method for evaluating knee joint prostheses. This method involves multimodal data acquisition based on biomechanical performance test data of the knee joint prosthesis, medical imaging data, and postoperative motion capture data of the patient. Data fusion is performed using a deep learning-based multimodal alignment network to obtain a fused multimodal dataset. This multimodal dataset is then input into an adaptive convolutional neural network to extract the mechanical characteristics of the knee joint prosthesis under dynamic load, resulting in a dynamic mechanical feature matrix. This dynamic mechanical feature matrix is ​​then input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance, generating a performance evaluation report. This comprehensive quantitative analysis of prosthesis performance enhances the scientific rigor and accuracy of prosthesis design. Attached Figure Description

[0013] Figure 1 A hardware structure block diagram of a computer terminal for a knee joint prosthesis evaluation method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a knee joint prosthesis evaluation method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a knee joint prosthesis evaluation system provided in an embodiment of the present invention. Detailed Implementation

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] This invention first provides a method for evaluating knee joint prostheses, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0016] The following detailed explanation uses a computer terminal as an example. Figure 1This is a hardware structure block diagram of a computer terminal for a knee joint prosthesis evaluation method provided in an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0017] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions that, when executed, cause the processor to perform any knee prosthesis evaluation method.

[0018] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0019] Internal memory provides an environment for the execution of computer programs in non-volatile storage media, which, when executed by a processor, enable the processor to perform any knee joint prosthesis evaluation method.

[0020] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0021] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0022] See Figure 2 The present invention provides a method for evaluating knee joint prostheses, which may include the following steps: S201. Based on the biomechanical performance test data of the knee joint prosthesis, medical imaging data and postoperative motion capture data of the patient, multimodal data is collected, and a deep learning-based multimodal alignment network is used to fuse the data, mapping the data from different sources to a unified spatiotemporal coordinate system to obtain the fused multimodal dataset. This step describes how to fuse biomechanical performance test data, medical imaging data, and postoperative motion capture data of knee prostheses through multimodal data acquisition and a deep learning-based multimodal alignment network. First, multimodal data acquisition gathers data from various sources, including biomechanical performance test data (such as stress and strain), medical imaging data (such as CT and MRI), and postoperative motion capture data (such as joint angles and gait analysis). Next, a deep learning-based multimodal alignment network maps these different data sources to a unified spatiotemporal coordinate system. This process involves timestamp alignment and spatial registration techniques to ensure data consistency in time and space. Finally, a fused multimodal dataset is generated, providing a foundation for subsequent mechanical feature extraction and performance evaluation.

[0023] The core function of this step is to address the temporal and spatial inconsistencies of multimodal data, ensuring data integrity and consistency. Through a multimodal alignment network, data from different sources can be effectively fused, providing a high-quality data foundation for subsequent mechanical feature extraction and performance evaluation. This fusion not only improves data utilization but also more comprehensively reflects the performance of knee prostheses in actual use, providing a scientific basis for prosthesis design and optimization. Furthermore, the fusion of multimodal data helps physicians and researchers better understand the interaction between the prosthesis and the patient, thus supporting the development of personalized medicine and rehabilitation programs.

[0024] Specifically, based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, an edge computing-based data acquisition framework can be used to acquire multimodal data in real time. Through a lightweight data caching mechanism, the real-time and continuous nature of data acquisition can be ensured. This step describes how to acquire multimodal data of a knee prosthesis in real time using an edge computing-based data acquisition framework. The edge computing framework performs preliminary processing at the source of data generation, reducing data transmission latency and ensuring data real-time performance and continuity. A lightweight data caching mechanism is used to temporarily store data, preventing data loss and ensuring uninterrupted data transmission.

[0025] This step ensures the real-time and continuous acquisition of multimodal data, preventing data loss or delays. Edge computing frameworks enable preliminary processing at the data's source, reducing reliance on central servers and improving data acquisition efficiency. This is particularly important for knee prosthesis performance evaluation requiring real-time monitoring, providing timely and accurate data support for subsequent processing and analysis.

[0026] In practical applications, edge computing-based data acquisition frameworks can be implemented using distributed computing nodes. These nodes can be deployed in operating rooms, rehabilitation centers, or patients' homes to collect multimodal data about knee prostheses in real time. For example, biomechanical performance test data can be acquired in real time using implanted or external sensors, which can measure mechanical parameters such as stress, strain, and pressure of the prosthesis during movement. Medical imaging data can be acquired using portable CT or MRI equipment, which can perform regular postoperative scans to record structural changes in the prosthesis and surrounding tissues. Motion capture data can be captured using wearable devices (such as smart joint braces) or cameras to record the patient's gait, joint angles, and movement trajectory.

[0027] To ensure the real-time and continuous nature of data acquisition, edge computing nodes need to possess efficient data processing capabilities. For example, lightweight data caching mechanisms, such as circular buffers or in-memory databases (like Redis), can be deployed on each node to temporarily store collected data. Circular buffers ensure that the latest data is stored first when the data volume is large, preventing data loss. In-memory databases can temporarily store data when network conditions are poor, transmitting the data to the central server once network recovery is achieved. Furthermore, edge computing nodes can employ data compression techniques to reduce data transmission bandwidth requirements, further improving data acquisition efficiency.

[0028] For example, suppose a patient is undergoing post-operative rehabilitation training at a rehabilitation center. Wearable devices capture the movement trajectory of their knee joint in real time, while implanted sensors record the stress distribution of the prosthesis. Edge computing nodes can receive this data in real time and temporarily store it using a lightweight caching mechanism. When network conditions permit, the nodes compress the data and transmit it to the central server. In this way, the real-time and continuous nature of data acquisition can be ensured, providing a high-quality data source for subsequent multimodal data fusion.

[0029] For multimodal data, a deep learning-based format recognition model is used to automatically identify the format types of different data sources. An adaptive data cleaning algorithm is used to filter noise and impute missing values ​​in the data to generate a preliminary standardized dataset. This step describes how a deep learning-based format recognition model can automatically identify the format types of different data sources and perform data cleaning and standardization. Since multimodal data may originate from different devices and systems, data formats and structures may vary. The format recognition model can automatically identify the data format types and employ adaptive data cleaning algorithms to filter noise and impute missing values, generating a preliminary standardized dataset.

[0030] This step aims to resolve inconsistencies in multimodal data formats and data quality issues. Automated format identification and data cleaning reduce manual intervention and improve data processing efficiency. The standardized dataset provides a unified foundation for subsequent multimodal alignment and feature extraction, ensuring the accuracy and reliability of subsequent analyses.

[0031] Multimodal data typically originates from diverse devices and systems, and its format and structure can vary significantly. For example, biomechanical performance test data might be stored in CSV or JSON format, medical imaging data in DICOM format, and motion capture data in binary or text format. To automatically identify these formats, deep learning-based format recognition models, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), can be employed. These models can learn the characteristics of different data formats through training and automatically identify the format type of the input data. For instance, CNNs can be used to identify image data formats (such as DICOM), while RNNs can be used to identify time series data formats (such as CSV or JSON).

[0032] After identifying the data format, the data needs to be cleaned and standardized. Since multimodal data may contain noise or missing values, adaptive data cleaning algorithms can be used. For example, for noise in biomechanical performance test data, wavelet transform or Kalman filtering can be used; for missing values ​​in medical imaging data, interpolation algorithms or imputation methods based on Generative Adversarial Networks (GANs) can be used. Specifically, wavelet transform can effectively separate high-frequency noise in the signal, while Kalman filtering can predict the current value based on historical data, thereby reducing the impact of noise. For missing value imputation, GANs can learn the distribution characteristics of the data to generate reasonable imputation values.

[0033] For example, if biomechanical performance test data contains missing values ​​due to sensor malfunction, GANs can be used to generate appropriate imputation values. Simultaneously, noise in medical imaging data can be filtered using wavelet transform to generate clearer image data. Finally, after format recognition and data cleaning, a preliminary standardized dataset is generated, providing a foundation for subsequent multimodal alignment.

[0034] For the initial standardized dataset, a deep learning-based multimodal alignment network is used to map data from different sources to a unified spatiotemporal coordinate system. Through timestamp alignment algorithms and spatial registration techniques, the temporal and spatial differences between data are eliminated, generating an initial aligned multimodal dataset. This step describes how to map data from different sources to a unified spatiotemporal coordinate system using a deep learning-based multimodal alignment network. Since multimodal data may be collected under different temporal and spatial conditions, there may be temporal and spatial differences between the data. Through timestamp alignment algorithms and spatial registration techniques, these differences can be eliminated, generating a preliminarily aligned multimodal dataset.

[0035] This step addresses the temporal and spatial inconsistencies of multimodal data, ensuring data fusion within a unified spatiotemporal coordinate system. Through a multimodal alignment network, data from different sources can be effectively aligned, providing a consistent data foundation for subsequent feature extraction and performance evaluation. This is crucial for accurately evaluating the performance of knee prostheses, preventing errors caused by data inconsistencies.

[0036] Multimodal data is typically collected under different temporal and spatial conditions, and there may be temporal and spatial differences between the data. To eliminate these differences, deep learning-based multimodal alignment networks, such as spatiotemporal convolutional neural networks (ST-CNN) or graph neural networks (GNN), can be used. These networks can map data from different sources to a unified spatiotemporal coordinate system. For example, ST-CNN can use temporal convolution and spatial convolution to process the temporal and spatial dimensions of the data respectively, thereby achieving temporal and spatial alignment.

[0037] For timestamp alignment, Dynamic Time Warping (DTW) algorithms or RNN-based time alignment models can be used. DTW algorithms can effectively align time series data of different lengths, while RNN models can achieve precise time alignment by learning the dependencies between time series. For example, assuming that biomechanical performance test data and motion capture data have inconsistent timestamps, DTW algorithms can be used to align their time series, ensuring temporal consistency of the data.

[0038] For spatial registration, feature-point-based registration algorithms or deep learning registration models can be used. For example, for medical image data and motion capture data, key points can be extracted using feature point detection algorithms (such as SIFT or SURF), and then aligned using affine or perspective transformations. Deep learning registration models, on the other hand, can learn the spatial features of image data to achieve automated spatial registration. For example, aligning the knee joint structure in CT images with the joint angles in motion capture data ensures spatial consistency between the two.

[0039] For example, suppose medical imaging data and motion capture data record the structure and motion trajectory of a patient's knee joint, respectively. A deep learning-based multimodal alignment network can align the temporal and spatial information of both, generating a preliminary aligned multimodal dataset. For instance, it can align the knee joint structure in CT images with the joint angles in motion capture data, ensuring spatial consistency between the two.

[0040] For the initially aligned multimodal dataset, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse the features of biomechanical performance data, medical imaging data and motion capture data. Through cross-modal attention mechanism, the correlation between different data sources is captured to generate the fused multimodal dataset.

[0041] This step describes how to use a feature fusion method based on a multi-head attention mechanism to weightedly fuse features from biomechanical performance data, medical imaging data, and motion capture data. Through cross-modal attention, the correlations between different data sources can be captured, generating a fused multimodal dataset.

[0042] This step aims to achieve deep fusion of multimodal data, capturing the correlations between different data sources to more comprehensively reflect the performance of knee prostheses. Through a multi-head attention mechanism, the weights of different data sources are automatically learned, ensuring that the fused dataset fully utilizes information from each source. This is crucial for subsequent mechanical feature extraction and performance evaluation, improving the accuracy and reliability of the analysis.

[0043] Based on the initially aligned multimodal datasets, feature fusion methods based on multi-head attention mechanisms, such as the Transformer model, can be used to weightedly fuse features from biomechanical performance data, medical imaging data, and motion capture data. Multi-head attention mechanisms can automatically learn the weights of different data sources and capture the correlations between them through cross-modal attention. For example, for stress-strain features in biomechanical performance data, structural features in medical imaging data, and motion trajectory features in motion capture data, the attention weights of each feature can be calculated separately using multi-head attention mechanisms, and then weighted and fused.

[0044] Specifically, multi-head attention mechanisms can process different features in parallel using multiple attention heads, each focusing on capturing a specific type of feature relationship. For example, one attention head can focus on capturing the relationship between stress distribution and motion trajectory, while another attention head can focus on capturing the relationship between structural features and mechanical properties. In this way, information from various data sources can be fully utilized to generate a fused multimodal dataset.

[0045] For example, suppose biomechanical performance data records the stress distribution of the prosthesis during movement, medical imaging data records the structural changes of the prosthesis and surrounding tissues, and motion capture data records the patient's joint angles and gait information. Through a multi-head attention mechanism, the relationships between these features can be automatically learned, such as the relationship between stress distribution and joint angles, or the relationship between structural changes and gait information. Ultimately, a fused multimodal dataset is generated, providing a foundation for subsequent mechanical feature extraction and performance evaluation.

[0046] This approach enables deep fusion of multimodal data, capturing the correlation between different data sources, and thus more comprehensively reflecting the performance of knee prostheses.

[0047] S202, The multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. This step describes how to input the fused multimodal dataset into an adaptive convolutional neural network (ACNN) to extract the mechanical features of the knee prosthesis under dynamic loads. The ACNN uses a temporal convolution module to capture the stress distribution and strain changes of the prosthesis at different stages of motion. Specifically, ACNN utilizes multi-scale convolutional kernels to extract mechanical features at different scales and analyzes the changes of these features over time using a temporal convolution module. For example, the stress distribution and strain of the prosthesis change during different stages of motion such as walking, running, or climbing stairs. ACNN can capture these temporal changes and generate a dynamic mechanical feature matrix. This matrix not only contains the mechanical response of the prosthesis in the short term but also reflects changes in mechanical properties during long-term use.

[0048] The core function of this step is to extract the mechanical characteristics of the knee prosthesis under dynamic loads using an adaptive convolutional neural network, providing crucial input for subsequent performance evaluation. The temporal convolution module captures the stress distribution and strain changes of the prosthesis at different stages of motion, thus comprehensively reflecting its mechanical behavior in actual use. This is significant for evaluating the durability, stability, and adaptability of the prosthesis. For example, by analyzing stress changes in the prosthesis over long-term use, its wear can be predicted; by capturing strain characteristics at different stages of motion, the performance of the prosthesis in various activities can be evaluated. Ultimately, these dynamic mechanical characteristics provide a scientific basis for the optimized design of the prosthesis and personalized medicine.

[0049] Specifically, for multimodal datasets, a feature extraction method based on adaptive convolutional neural networks can be used to extract stress-strain features from biomechanical performance data, structural features from medical imaging data, and motion trajectory features from motion capture data. Through multi-scale convolutional kernels, mechanical features at different scales can be captured to generate preliminary feature representations. This step describes how to extract features from different data sources using an Adaptive Convolutional Neural Network (ACNN). ACNN utilizes multi-scale convolutional kernels to extract stress-strain features from biomechanical performance data, structural features from medical imaging data, and motion trajectory features from motion capture data. Multi-scale convolutional kernels can capture mechanical features at different scales, such as localized stress concentration areas or overall structural deformation.

[0050] This step aims to extract key features from multimodal data, providing a foundation for subsequent dynamic modeling and feature fusion. Through multi-scale convolutional kernels, the mechanical properties of the prosthesis can be comprehensively captured, including both local and global features, thus providing rich information for subsequent performance evaluation.

[0051] In practical applications, Adaptive Convolutional Neural Networks (ACNNs) can extract features from multimodal datasets through multiple convolutional and pooling layers. First, for biomechanical performance data, ACNNs use small-scale convolutional kernels (e.g., 3x3) to capture localized stress concentration areas, such as stress changes at a specific point in a knee prosthesis during movement. Simultaneously, large-scale convolutional kernels (e.g., 7x7) capture the overall stress distribution, such as the stress variation trend of the prosthesis throughout the gait cycle. This multi-scale kernel design comprehensively captures the mechanical properties of the prosthesis, focusing on both local details and overall trends.

[0052] For medical imaging data, ACNN extracts structural features of the prosthesis and surrounding tissues using 3D convolutional kernels. For example, CT or MRI images can show the contact surfaces between the prosthesis and bone and soft tissues, and 3D convolutional kernels can capture the geometric features and density distribution of these contact surfaces. Through multi-scale convolutional kernels, ACNN can simultaneously extract local structural features (such as the contact area between the prosthesis and bone) and global structural features (such as the position and orientation of the prosthesis throughout the knee joint).

[0053] For motion capture data, ACNN uses temporal convolutional kernels to extract temporal features of motion trajectories. For example, joint angle and gait data captured by wearable devices or cameras can be analyzed for their variation patterns using temporal convolutional kernels. Through multi-scale convolutional kernels, ACNN can capture short-term fluctuations (such as changes in joint angles during a single step) and long-term trends (such as movement patterns throughout the entire gait cycle). Finally, ACNN initially integrates the extracted stress-strain features, structural features, and motion trajectory features to generate preliminary feature representations, providing input for subsequent dynamic modeling.

[0054] For example, assuming biomechanical performance data records the stress distribution of a prosthesis during walking, ACNN can capture the stress concentration area on the medial side of the knee joint using small-scale convolutional kernels, while capturing the overall stress distribution using large-scale convolutional kernels. For medical imaging data, 3D convolutional kernels can extract the contact surface features between the prosthesis and surrounding bone tissue. For motion capture data, temporal convolutional kernels can extract joint angle changes during the gait cycle. Through multi-scale convolutional kernels, preliminary feature representations can be generated, providing a foundation for subsequent dynamic modeling.

[0055] For the initial feature representation, a dynamic modeling method based on temporal convolution modules is adopted. By combining the load changes of the prosthesis at different motion stages, the temporal features of stress distribution and strain change are captured. Through the sliding window mechanism, short-term fluctuation and long-term trend features are extracted to generate dynamic mechanical feature representation. This step describes how to dynamically model the initial feature representation using a temporal convolution module. The temporal convolution module captures load changes in the prosthesis at different stages of motion, analyzing the temporal characteristics of stress distribution and strain changes. Through a sliding window mechanism, short-term fluctuations (such as stress changes during a single step) and long-term trends (such as wear trends during prolonged use) can be extracted.

[0056] This step aims to capture the mechanical behavior of the prosthesis under dynamic loads, providing temporal characteristics for performance evaluation. Through the temporal convolution module, the performance of the prosthesis at different stages of motion can be comprehensively analyzed, thereby assessing its stability and durability in practical use.

[0057] In practical applications, the temporal convolution module dynamically models the initial feature representation using a sliding window mechanism. First, the temporal convolution module divides the initial feature representation into multiple time segments, each corresponding to the mechanical properties of the prosthesis during a specific movement phase (such as walking, running, or climbing stairs). For example, for gait cycles in motion capture data, the temporal convolution module can use a sliding window to divide the data into multiple time segments, extracting the stress distribution and strain changes for each segment. In this way, it is possible to capture the short-term fluctuation characteristics of the prosthesis at different movement phases, such as the stress peak and strain changes during a single step.

[0058] For long-term trend analysis, the temporal convolution module uses multiple stacked temporal convolutional layers to capture changes in the mechanical properties of the prosthesis over months or years of use. For example, by analyzing stress changes in the prosthesis during stair climbing and running, its performance under high-load conditions can be evaluated. The temporal convolution module can also dynamically adjust feature extraction strategies based on changes in the prosthesis's load. For instance, under high-load conditions, the temporal convolution module can increase focus on stress concentration areas, capturing the fatigue characteristics of the prosthesis.

[0059] For example, assuming motion capture data records changes in joint angles during different stages of movement (such as walking, running, and climbing stairs), the temporal convolution module can divide the data into multiple time segments using a sliding window, extracting the stress distribution and strain changes for each segment. For long-term trend analysis, multiple temporal convolutional layers can be stacked to capture the stress change trend of the prosthesis over a year. For instance, by analyzing stress changes in the prosthesis during long-term use, its wear and tear can be predicted. Ultimately, the temporal convolution module generates a dynamic mechanical feature representation, providing input for subsequent feature fusion.

[0060] For the representation of dynamic mechanical features, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse stress distribution features, strain change features and motion trajectory features. Through cross-modal attention mechanism, the correlation between different features is captured to generate a fused dynamic mechanical feature matrix. This step describes how to fuse different features using a multi-head attention mechanism. The multi-head attention mechanism can automatically learn the weights of stress distribution features, strain change features, and motion trajectory features, and capture the correlations between these features through a cross-modal attention mechanism. For example, the relationship between stress distribution and motion trajectory can be modeled using an attention mechanism.

[0061] The purpose of this step is to achieve deep fusion of multimodal features, capture the correlation between different features, and thus more comprehensively reflect the mechanical behavior of the prosthesis. Through a multi-head attention mechanism, a fused dynamic mechanical feature matrix can be generated, providing high-quality input for subsequent performance evaluation.

[0062] In practical applications, the multi-head attention mechanism achieves deep fusion of multimodal features by processing different features in parallel using multiple attention heads. First, each attention head focuses on capturing a specific type of feature relationship. For example, one attention head can focus on capturing the relationship between stress distribution and motion trajectory, analyzing the correlation between stress concentration areas and joint angle changes during prosthesis movement. Another attention head can capture the relationship between strain changes and structural features, analyzing the deformation characteristics of the prosthesis under different load conditions.

[0063] Through a cross-modal attention mechanism, the multi-head attention mechanism can automatically learn the weights of different features and perform weighted fusion. For example, for stress distribution features, the attention mechanism can dynamically adjust the weights based on their correlation with the motion trajectory; for strain change features, it can adjust the weights based on their correlation with structural features. In this way, a fused dynamic mechanical feature matrix can be generated, comprehensively reflecting the mechanical behavior of the prosthesis.

[0064] For example, suppose stress distribution features record the stress concentration areas of the prosthesis during running, and motion trajectory features record the knee joint angle changes. A multi-head attention mechanism can capture the relationship between the two through a cross-modal attention mechanism. For instance, when the knee joint angle reaches its maximum, the stress distribution features may show changes in the stress concentration areas. Through weighted fusion, a fused dynamic mechanical feature matrix is ​​generated, providing input for subsequent performance evaluation.

[0065] For the fused dynamic mechanical feature matrix, a Bayesian optimization-based calibration method is adopted. Combining the material properties of the prosthesis and the actual use scenario, the feature weights are dynamically adjusted. Through regularization constraints, overfitting is prevented, and the final dynamic mechanical feature matrix is ​​generated.

[0066] This step describes how to calibrate the fused dynamic mechanical feature matrix using Bayesian optimization. Bayesian optimization dynamically adjusts feature weights by combining the material properties of the prosthesis with the actual usage scenario, and prevents overfitting through regularization constraints to generate the final dynamic mechanical feature matrix.

[0067] This step aims to optimize the weights of the feature matrix to ensure it accurately reflects the mechanical behavior of the prosthesis. Through Bayesian optimization, the generalization ability of the feature matrix can be improved by considering real-world application scenarios, providing reliable input for performance evaluation.

[0068] In practical applications, Bayesian optimization generates an optimal dynamic mechanical feature matrix by iteratively adjusting feature weights and considering the material properties of the prosthesis and the actual usage scenario. First, Bayesian optimization dynamically adjusts the feature weights based on the material properties of the prosthesis (such as elastic modulus and wear resistance). For example, for highly elastic materials, the weight of stress distribution features can be increased to more accurately reflect the mechanical behavior of the prosthesis under high load conditions. For highly wear-resistant materials, the weight of strain change features can be increased to capture the deformation characteristics of the prosthesis during long-term use.

[0069] Secondly, Bayesian optimization further refines feature weights by incorporating real-world usage scenarios (such as motion frequency and load conditions). For example, for patients who frequently engage in high-intensity exercise, greater emphasis can be placed on stress distribution characteristics under high load conditions; for patients who engage in low-intensity exercise, greater emphasis can be placed on strain change characteristics under low load conditions. In this way, a dynamic mechanical feature matrix that highly matches the actual usage scenario can be generated.

[0070] Finally, Bayesian optimization prevents overfitting through regularization constraints, ensuring the generalization ability of the feature matrix. For example, by adding an L2 regularization term to the loss function, excessive growth of feature weights is limited, preventing the model from overfitting the training data. Ultimately, Bayesian optimization generates the final dynamic mechanical feature matrix, providing reliable input for performance evaluation.

[0071] For example, assuming the prosthesis material has a high elastic modulus, Bayesian optimization can increase the weight of stress distribution features to more accurately reflect the mechanical behavior of the prosthesis under high load conditions. Simultaneously, regularization constraints can prevent overfitting and ensure the generalization ability of the feature matrix in real-world applications. Finally, a final dynamic mechanical feature matrix is ​​generated, providing reliable input for performance evaluation.

[0072] S203, the dynamic mechanical feature matrix is ​​input into the performance evaluation model based on the multi-scale attention mechanism to perform quantitative analysis of the prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0073] This step describes how to input the dynamic mechanical feature matrix into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance. This performance evaluation model comprehensively assesses the stability, wear resistance, and biocompatibility of the prosthesis by combining local and global features. Specifically, local features are extracted using a multi-scale convolutional neural network to capture the mechanical properties of the prosthesis at the microscale, such as deformation and wear in stress concentration areas; global features are extracted using a graph convolutional network to abstract the overall mechanical behavior of the prosthesis into a graph structure, capturing its macroscopic mechanical properties, such as overall stress distribution and structural stability. Through the multi-scale attention mechanism, the model can dynamically adjust the weights of local and global features, capturing the correlation between them, thereby generating a comprehensive performance evaluation report.

[0074] The core function of this step is to comprehensively and quantitatively analyze the performance of the prosthesis through a multi-scale attention mechanism performance evaluation model. By combining local and global features, the stability, wear resistance, and biocompatibility of the prosthesis can be evaluated from both microscopic and macroscopic levels. For example, local features can reflect the wear of the prosthesis under high load conditions, while global features can assess the structural stability of the prosthesis during long-term use. This multi-scale evaluation method not only improves the accuracy of the analysis but also provides a scientific basis for the optimized design of the prosthesis and personalized medicine. The final performance evaluation report provides doctors and researchers with detailed quantitative results and evaluation evidence, which helps to develop more precise treatment and rehabilitation plans.

[0075] Specifically, a feature extraction method based on multi-scale convolutional neural networks can be used to extract local features of the prosthesis at the microscale from the dynamic mechanical feature matrix. Through adaptive convolution kernels, local mechanical properties at different scales can be captured to generate local feature representations. This step describes how to extract local features of a prosthesis at the microscale from a dynamic mechanical feature matrix using a multi-scale convolutional neural network (MS-CNN). MS-CNN captures local mechanical properties at different scales, such as deformation and wear in stress concentration areas, through adaptive convolutional kernels.

[0076] This step aims to extract the local mechanical properties of the prosthesis at the microscale, providing detailed information for subsequent performance evaluation. Through multi-scale convolution kernels, the local performance of the prosthesis under high-load conditions can be comprehensively captured, thereby assessing its wear resistance and stability.

[0077] In practical applications, multi-scale convolutional neural networks (MS-CNN) extract local features from the dynamic mechanical feature matrix through multiple convolutional and pooling layers. First, MS-CNN uses small-scale convolutional kernels (e.g., 3x3) to capture deformation features in stress concentration areas on the prosthesis surface. For example, during running, stress concentration may occur on the medial side of the knee joint prosthesis; small-scale convolutional kernels can capture the local deformation in this area. Simultaneously, MS-CNN uses medium-scale convolutional kernels (e.g., 5x5) to capture the contact surface features between the prosthesis and surrounding tissues. For example, the contact surface between the prosthesis and bone may experience minute displacements; medium-scale convolutional kernels can capture the details of these displacements.

[0078] By using adaptive convolutional kernels, MS-CNN can dynamically adjust the size and shape of the kernels to adapt to local mechanical properties at different scales. For example, for stress concentration areas of a prosthesis moving up and down stairs, MS-CNN can capture local deformation using small-scale convolutional kernels while capturing the response of surrounding tissue using medium-scale convolutional kernels. Ultimately, MS-CNN generates local feature representations, providing a foundation for subsequent global feature extraction. For example, local feature representations can include deformation maps of stress concentration areas on the prosthesis surface and displacement maps of the contact surfaces.

[0079] For the dynamic mechanical feature matrix, a global feature extraction method based on graph convolutional networks is adopted to abstract the overall mechanical behavior of the prosthesis into a graph structure, extract global features, capture global mechanical properties through multi-layer graph convolution operations, and generate global feature representations. This step describes how to extract global features of a prosthesis from a dynamic mechanical feature matrix using a Graph Convolutional Network (GCN). GCN abstracts the overall mechanical behavior of the prosthesis into a graph structure, where nodes represent key parts of the prosthesis and edges represent mechanical interactions. Through multi-layer graph convolution operations, global mechanical properties are captured.

[0080] This step aims to extract the global mechanical properties of the prosthesis at a macroscopic scale, providing holistic information for subsequent performance evaluation. Graph convolutional networks can comprehensively capture the structural stability and overall mechanical behavior of the prosthesis during long-term use.

[0081] In practical applications, Graph Convolutional Networks (GCNs) extract global features from the dynamic mechanical feature matrix through multi-layer graph convolution operations. First, GCNs abstract the overall mechanical behavior of the prosthesis as a graph structure. For example, key parts of the prosthesis (such as the contact surface and support structure of the knee joint) are abstracted as nodes in the graph, and mechanical interactions (such as stress transfer and deformation response) are abstracted as edges. Through this abstraction, GCNs can capture the overall mechanical properties of the prosthesis at different stages of motion.

[0082] For example, to capture the overall stress distribution of a prosthesis while moving up and down stairs, GCN can use graph convolution operations to capture the stress transfer path and overall deformation trend. Specifically, GCN first captures the stress distribution of the prosthesis during a single step using a single layer of graph convolution operations, and then captures the stress change trend of the prosthesis throughout the entire gait cycle using multiple layers of graph convolution operations. Finally, GCN generates a global feature representation, providing input for subsequent feature fusion. For example, the global feature representation can include the overall stress distribution map and deformation trend map of the prosthesis at different stages of motion.

[0083] For local and global feature representations, a feature fusion method based on multi-scale attention mechanism is adopted to perform weighted fusion of local and global features. Through cross-scale attention mechanism, the correlation between local and global features is captured to generate fused multi-scale feature representations. This step describes how to fuse local and global features using a multi-scale attention mechanism. The multi-scale attention mechanism dynamically adjusts the weights of local and global features and captures the correlation between them through cross-scale attention.

[0084] This step aims to achieve a deep fusion of local and global features, comprehensively reflecting the mechanical behavior of the prosthesis. Through a multi-scale attention mechanism, the correlation between local and global features can be captured, thereby generating a more comprehensive multi-scale feature representation.

[0085] In practical applications, the multi-scale attention mechanism processes local and global features in parallel using multiple attention heads. First, each attention head focuses on capturing a specific type of feature relationship. For example, one attention head can focus on capturing the relationship between local stress concentration areas and the overall stress distribution, analyzing the changing trends of stress concentration areas during the prosthesis's movement. Another attention head can capture the relationship between local deformation and overall structural stability, analyzing the deformation characteristics of the prosthesis under different load conditions.

[0086] Through a cross-scale attention mechanism, a multi-scale attention mechanism can dynamically adjust the weights of local and global features and perform weighted fusion. For example, for local stress concentration areas of the prosthesis during running, the multi-scale attention mechanism can capture the relationship between these areas and the overall stress distribution through a cross-scale attention mechanism. When the knee joint angle reaches its maximum value, the local stress concentration area may change, and the multi-scale attention mechanism can capture this trend. Finally, a fused multi-scale feature representation is generated, providing input for subsequent performance evaluation. For example, the fused multi-scale feature representation can include a comprehensive map of the local stress concentration areas of the prosthesis during running and the overall stress distribution.

[0087] The fused multi-scale feature representation is used to perform quantitative analysis by adopting a performance evaluation model based on deep neural networks, combined with the stability, wear resistance and biocompatibility indicators of the prosthesis. Through the interpretability module in the performance evaluation model, a performance evaluation report containing quantitative results and evaluation basis is generated.

[0088] This step describes how to evaluate the performance of the fused multi-scale feature representation using a deep neural network (DNN). The DNN combines indicators of the prosthesis's stability, abrasion resistance, and biocompatibility for quantitative analysis, and generates a performance evaluation report through an interpretability module.

[0089] This step aims to generate a detailed performance evaluation report, providing a scientific basis for the optimized design of prostheses and personalized medicine. The interpretability module provides quantitative results and evaluation evidence, helping doctors and researchers better understand the performance of the prostheses.

[0090] In practical applications, deep neural networks (DNNs) use multiple fully connected layers and activation functions to quantify and analyze the fused multi-scale feature representations. First, the DNN combines the stability indices of the prosthesis to analyze the stress distribution and deformation characteristics of the prosthesis at different stages of motion. For example, to assess the stability of the prosthesis while moving up and down stairs, the DNN can evaluate its stability under high load conditions by analyzing stress distribution maps and deformation trend maps.

[0091] Secondly, DNN incorporates abrasion resistance metrics of the prosthesis to assess its wear and tear over long-term use. For example, to evaluate the abrasion resistance of the prosthesis during running, DNN can assess its wear trend by analyzing deformation maps and wear prediction models of local stress concentration areas. Finally, DNN incorporates biocompatibility metrics of the prosthesis to analyze the interaction between the prosthesis and surrounding tissues. For example, regarding the contact surface between the prosthesis and bone, DNN can assess its biocompatibility performance by analyzing contact surface displacement maps and biocompatibility scores.

[0092] Through its interpretability module, the DNN can generate performance evaluation reports that include quantitative results and assessment criteria. For example, the report can display the stress distribution of the prosthesis under high-load conditions, wear predictions, and biocompatibility scores. Ultimately, the performance evaluation report provides physicians and researchers with detailed quantitative results and assessment criteria, helping to develop more precise treatment and rehabilitation plans.

[0093] As can be seen, multimodal data is collected based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients. A deep learning-based multimodal alignment network is then used for data fusion to obtain a fused multimodal dataset. This multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical characteristics of the knee joint prosthesis under dynamic load, resulting in a dynamic mechanical feature matrix. This dynamic mechanical feature matrix is ​​then input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance, generating a performance evaluation report. This enables a comprehensive quantitative analysis of prosthesis performance, improving the scientific rigor and accuracy of prosthesis design.

[0094] Another embodiment of the present invention provides a knee joint prosthesis evaluation system, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to acquire multimodal data based on the biomechanical performance test data of the knee joint prosthesis, medical imaging data and postoperative motion capture data of the patient, and to perform data fusion using a deep learning-based multimodal alignment network to map data from different sources into a unified spatiotemporal coordinate system to obtain a fused multimodal dataset. Extraction module 302 is used to input the multimodal dataset into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. Analysis module 303 is used to input the dynamic mechanical feature matrix into a performance evaluation model based on a multi-scale attention mechanism to perform quantitative analysis of prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0095] As can be seen, multimodal data is collected based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients. A deep learning-based multimodal alignment network is then used for data fusion to obtain a fused multimodal dataset. This multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical characteristics of the knee joint prosthesis under dynamic load, resulting in a dynamic mechanical feature matrix. This dynamic mechanical feature matrix is ​​then input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance, generating a performance evaluation report. This enables a comprehensive quantitative analysis of prosthesis performance, improving the scientific rigor and accuracy of prosthesis design.

[0096] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0097] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps: S201. Based on the biomechanical performance test data of the knee joint prosthesis, medical imaging data and postoperative motion capture data of the patient, multimodal data is collected, and a deep learning-based multimodal alignment network is used to fuse the data, mapping the data from different sources to a unified spatiotemporal coordinate system to obtain the fused multimodal dataset. S202, The multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. S203, the dynamic mechanical feature matrix is ​​input into the performance evaluation model based on the multi-scale attention mechanism to perform quantitative analysis of the prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0098] As can be seen, multimodal data is collected based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients. A deep learning-based multimodal alignment network is then used for data fusion to obtain a fused multimodal dataset. This multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical characteristics of the knee joint prosthesis under dynamic load, resulting in a dynamic mechanical feature matrix. This dynamic mechanical feature matrix is ​​then input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance, generating a performance evaluation report. This enables a comprehensive quantitative analysis of prosthesis performance, improving the scientific rigor and accuracy of prosthesis design.

[0099] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0100] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0101] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program: S201. Based on the biomechanical performance test data of the knee joint prosthesis, medical imaging data and postoperative motion capture data of the patient, multimodal data is collected, and a deep learning-based multimodal alignment network is used to fuse the data, mapping the data from different sources to a unified spatiotemporal coordinate system to obtain the fused multimodal dataset. S202, The multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. S203, the dynamic mechanical feature matrix is ​​input into the performance evaluation model based on the multi-scale attention mechanism to perform quantitative analysis of the prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

[0102] As can be seen, multimodal data is collected based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients. A deep learning-based multimodal alignment network is then used for data fusion to obtain a fused multimodal dataset. This multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical characteristics of the knee joint prosthesis under dynamic load, resulting in a dynamic mechanical feature matrix. This dynamic mechanical feature matrix is ​​then input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance, generating a performance evaluation report. This enables a comprehensive quantitative analysis of prosthesis performance, improving the scientific rigor and accuracy of prosthesis design.

[0103] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for evaluating knee joint prostheses, characterized in that, The method includes: Based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, multimodal data were collected, and a deep learning-based multimodal alignment network was used for data fusion to map data from different sources into a unified spatiotemporal coordinate system, resulting in a fused multimodal dataset. The multimodal dataset is input into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. The dynamic mechanical feature matrix is ​​input into a performance evaluation model based on a multi-scale attention mechanism to perform quantitative analysis of prosthesis performance. The performance evaluation model evaluates the stability, wear resistance, and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

2. The method according to claim 1, characterized in that, The process involves collecting multimodal data based on biomechanical performance test data of knee prostheses, medical imaging data, and postoperative motion capture data from patients. A deep learning-based multimodal alignment network is then used for data fusion, mapping data from different sources to a unified spatiotemporal coordinate system to obtain a fused multimodal dataset, including: Based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, an edge computing-based data acquisition framework is adopted to acquire multimodal data in real time. A lightweight data caching mechanism is used to ensure the real-time and continuous nature of data acquisition. For multimodal data, a deep learning-based format recognition model is used to automatically identify the format types of different data sources. An adaptive data cleaning algorithm is used to filter noise and impute missing values ​​in the data to generate a preliminary standardized dataset. For the initial standardized dataset, a deep learning-based multimodal alignment network is used to map data from different sources to a unified spatiotemporal coordinate system. Through timestamp alignment algorithms and spatial registration techniques, the temporal and spatial differences between data are eliminated, generating an initial aligned multimodal dataset. For the initially aligned multimodal dataset, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse the features of biomechanical performance data, medical imaging data and motion capture data. Through cross-modal attention mechanism, the correlation between different data sources is captured to generate the fused multimodal dataset.

3. The method according to claim 2, characterized in that, The process involves inputting the multimodal dataset into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network uses temporal convolution to capture the stress distribution and strain changes of the prosthesis at different motion stages, obtaining a dynamic mechanical feature matrix, including: For multimodal datasets, a feature extraction method based on adaptive convolutional neural networks is used to extract stress-strain features from biomechanical performance data, structural features from medical imaging data, and motion trajectory features from motion capture data. Through multi-scale convolutional kernels, mechanical features at different scales are captured to generate preliminary feature representations. For the initial feature representation, a dynamic modeling method based on temporal convolution modules is adopted. By combining the load changes of the prosthesis at different motion stages, the temporal features of stress distribution and strain change are captured. Through the sliding window mechanism, short-term fluctuation and long-term trend features are extracted to generate dynamic mechanical feature representation. For the representation of dynamic mechanical features, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse stress distribution features, strain change features and motion trajectory features. Through cross-modal attention mechanism, the correlation between different features is captured to generate a fused dynamic mechanical feature matrix. For the fused dynamic mechanical feature matrix, a Bayesian optimization-based calibration method is adopted. Combining the material properties of the prosthesis and the actual use scenario, the feature weights are dynamically adjusted. Through regularization constraints, overfitting is prevented, and the final dynamic mechanical feature matrix is ​​generated.

4. The method according to claim 3, characterized in that, The dynamic mechanical feature matrix is ​​input into a performance evaluation model based on a multi-scale attention mechanism for quantitative analysis of prosthesis performance. The performance evaluation model, by combining local and global features, assesses the stability, wear resistance, and biocompatibility of the prosthesis and generates a performance evaluation report, including: For the dynamic mechanical feature matrix, a feature extraction method based on multi-scale convolutional neural networks is adopted to extract the local features of the prosthesis at the microscale. Through adaptive convolution kernels, local mechanical properties at different scales are captured to generate local feature representations. For the dynamic mechanical feature matrix, a global feature extraction method based on graph convolutional networks is adopted to abstract the overall mechanical behavior of the prosthesis into a graph structure, extract global features, capture global mechanical properties through multi-layer graph convolution operations, and generate global feature representations. For local and global feature representations, a feature fusion method based on multi-scale attention mechanism is adopted to perform weighted fusion of local and global features. Through cross-scale attention mechanism, the correlation between local and global features is captured to generate fused multi-scale feature representations. The fused multi-scale feature representation is used to perform quantitative analysis by adopting a performance evaluation model based on deep neural networks, combined with the stability, wear resistance and biocompatibility indicators of the prosthesis. Through the interpretability module in the performance evaluation model, a performance evaluation report containing quantitative results and evaluation basis is generated.

5. A knee joint prosthesis evaluation system, characterized in that, The system includes: The acquisition module is used to collect multimodal data based on biomechanical performance test data of knee joint prostheses, medical imaging data and postoperative motion capture data of patients, and to perform data fusion using a deep learning-based multimodal alignment network to map data from different sources into a unified spatiotemporal coordinate system to obtain a fused multimodal dataset. The extraction module is used to input the multimodal dataset into an adaptive convolutional neural network to extract the mechanical features of the knee joint prosthesis under dynamic load. The adaptive convolutional neural network captures the stress distribution and strain changes of the prosthesis at different motion stages through temporal convolution to obtain a dynamic mechanical feature matrix. The analysis module is used to input the dynamic mechanical feature matrix into a performance evaluation model based on a multi-scale attention mechanism to perform quantitative analysis of prosthesis performance. The performance evaluation model evaluates the stability, wear resistance and biocompatibility of the prosthesis by combining local and global features, and generates a performance evaluation report.

6. The system according to claim 5, characterized in that, The acquisition module is specifically used for: Based on biomechanical performance test data of knee joint prostheses, medical imaging data, and postoperative motion capture data of patients, an edge computing-based data acquisition framework is adopted to acquire multimodal data in real time. A lightweight data caching mechanism is used to ensure the real-time and continuous nature of data acquisition. For multimodal data, a deep learning-based format recognition model is used to automatically identify the format types of different data sources. An adaptive data cleaning algorithm is used to filter noise and impute missing values ​​in the data to generate a preliminary standardized dataset. For the initial standardized dataset, a deep learning-based multimodal alignment network is used to map data from different sources to a unified spatiotemporal coordinate system. Through timestamp alignment algorithms and spatial registration techniques, the temporal and spatial differences between data are eliminated, generating an initial aligned multimodal dataset. For the initially aligned multimodal dataset, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse the features of biomechanical performance data, medical imaging data and motion capture data. Through cross-modal attention mechanism, the correlation between different data sources is captured to generate the fused multimodal dataset.

7. The system according to claim 6, characterized in that, The extraction module is specifically used for: For multimodal datasets, a feature extraction method based on adaptive convolutional neural networks is used to extract stress-strain features from biomechanical performance data, structural features from medical imaging data, and motion trajectory features from motion capture data. Through multi-scale convolutional kernels, mechanical features at different scales are captured to generate preliminary feature representations. For the initial feature representation, a dynamic modeling method based on temporal convolution modules is adopted. By combining the load changes of the prosthesis at different motion stages, the temporal features of stress distribution and strain change are captured. Through the sliding window mechanism, short-term fluctuation and long-term trend features are extracted to generate dynamic mechanical feature representation. For the representation of dynamic mechanical features, a feature fusion method based on multi-head attention mechanism is adopted to weight and fuse stress distribution features, strain change features and motion trajectory features. Through cross-modal attention mechanism, the correlation between different features is captured to generate a fused dynamic mechanical feature matrix. For the fused dynamic mechanical feature matrix, a Bayesian optimization-based calibration method is adopted. Combining the material properties of the prosthesis and the actual use scenario, the feature weights are dynamically adjusted. Through regularization constraints, overfitting is prevented, and the final dynamic mechanical feature matrix is ​​generated.

8. The system according to claim 7, characterized in that, The analysis module is specifically used for: For the dynamic mechanical feature matrix, a feature extraction method based on multi-scale convolutional neural networks is adopted to extract the local features of the prosthesis at the microscale. Through adaptive convolution kernels, local mechanical properties at different scales are captured to generate local feature representations. For the dynamic mechanical feature matrix, a global feature extraction method based on graph convolutional networks is adopted to abstract the overall mechanical behavior of the prosthesis into a graph structure, extract global features, capture global mechanical properties through multi-layer graph convolution operations, and generate global feature representations. For local and global feature representations, a feature fusion method based on multi-scale attention mechanism is adopted to perform weighted fusion of local and global features. Through cross-scale attention mechanism, the correlation between local and global features is captured to generate fused multi-scale feature representations. The fused multi-scale feature representation is used to perform quantitative analysis by adopting a performance evaluation model based on deep neural networks, combined with the stability, wear resistance and biocompatibility indicators of the prosthesis. Through the interpretability module in the performance evaluation model, a performance evaluation report containing quantitative results and evaluation basis is generated.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-4 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-4.

Images