Knee joint prosthesis reconstruction method based on double-domain contrast learning network

Abstract

The invention relates to the technical field of medical image processing and personalized medical treatment, and provides a knee joint prosthesis reconstruction method based on a double-domain contrast learning network. According to the method, the problems that the segmentation precision of the complex anatomical structure of the knee joint is low and the adaptability of a prosthesis model is insufficient are solved by combining double-domain contrast learning with image domain and frequency domain characteristics. According to the technical scheme, the method comprises the steps of obtaining and preprocessing knee joint image data, extracting knee joint anatomical features through a double-domain contrast learning network, conducting three-dimensional reconstruction and extracting key geometric features such as cartilage thickness, contact area and skeleton curvature, and finally generating a personalized prosthesis model highly matched with the anatomical structure of a patient in combination with ISTA-Net. The precision of prosthesis design can be remarkably improved, the postoperative recovery effect is optimized, and the method is widely applied to the field of personalized medical treatment and knee joint replacement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and personalized medicine, and particularly relates to a knee MRI image prosthesis reconstruction method based on a dual-domain contrast learning network, especially combining dual-domain contrast learning, image segmentation technology and 3D modeling technology to realize accurate modeling of the knee joint, and designing a personalized prosthesis based on the individual anatomical features of the patient. BACKGROUND

[0002] Knee arthroplasty (TKA) is widely used in the treatment of knee osteoarthritis, rheumatoid arthritis and other diseases, but the traditional knee prosthesis design is usually based on standardized templates, ignoring the individual differences of patients, which may lead to postoperative prosthesis discomfort, unsatisfactory functional recovery and other problems. Personalized prosthesis reconstruction technology is based on medical image data such as CT or MRI data of patients, combined with anatomical and biomechanical requirements, to design a prosthesis that highly matches the patient's anatomical structure, thereby improving postoperative results and prosthesis service life.

[0003] However, existing personalized prosthesis design methods still have many challenges, mainly focusing on how to accurately extract the anatomical structure of the knee joint, how to efficiently convert image data into a 3D model that can be used for prosthesis design, and how to ensure perfect matching of the prosthesis model with key areas such as patient bones and cartilage.

[0004] In order to solve the above problems, in recent years, medical image segmentation technology based on deep learning has been widely applied. Deep learning models, especially convolutional neural networks (CNN), can automatically learn the structural features of the knee joint and extract key areas such as bones and cartilage through accurate image segmentation. In addition, the introduction of 3D modeling technology also enables the prosthesis design to more accurately match the patient's anatomical features.

[0005] Although there are some personalized prosthesis design methods, the existing methods still have some deficiencies, such as low segmentation accuracy for complex knee joint anatomical structures, and incomplete optimization algorithms for prosthesis design. How to improve the segmentation accuracy of knee joint images and effectively convert the segmentation results into personalized prosthesis models is still an important problem faced by current technology. SUMMARY

[0006] The knee joint prosthesis reconstruction method based on the dual-domain contrast learning network aims to solve the problem of low precision in the image segmentation of the complex anatomical structure of the knee joint (such as the femur, patella, tibia and cartilage) in the current medical technology, especially in the case of blurred boundary or poor image quality. The knee joint prosthesis reconstruction method based on the dual-domain contrast learning network combines the knee joint anatomical structure segmentation technology based on the dual-domain contrast learning, the 3D modeling technology and the prosthesis design optimization algorithm to design a new personalized knee joint prosthesis reconstruction method. The method can accurately segment the anatomical structure of the knee joint and generate a prosthesis model that matches the patient's anatomical structure, facilitating efficient and accurate preoperative planning for doctors.

[0007] The technical scheme of the present application is a knee joint prosthesis reconstruction method based on a dual-domain contrast learning network, and the flow chart is as shown in Figure 1 The knee joint prosthesis reconstruction method based on the dual-domain contrast learning network includes the following steps:

[0008] S1, pre-processing the knee joint medical images in the data set

[0009] S2, data enhancement on the pre-processed medical images;

[0010] S3, constructing an anatomical structure segmentation process based on a dual-domain contrast learning network;

[0011] S4, extracting geometric features from the segmentation results;

[0012] S5, generating a prosthesis model that fits the individual knee joint based on the patient's geometric data.

[0013] Further, the pre-processing operation in step S1 includes data cropping, resampling, standardization and noise reduction processing.

[0014] Data cropping aims to remove the redundant background area in the medical image and only keep the key anatomical parts of the knee joint, such as the femur, patella, tibia and cartilage area. This operation can effectively focus on the region of interest, reduce data redundancy, reduce the computational complexity of subsequent processing, and improve the learning effect of the model on the anatomical features of the knee joint.

[0015] Resampling operation unifies the spatial resolution (256x256 pixels) and slice thickness of the image, eliminates the resolution difference caused by different imaging devices or scanning protocols, and ensures the format consistency of the input data. Through resampling, the data complexity can be reduced while preserving the anatomical structure details, providing a standardized input format for the deep learning model.

[0016] The standardization process reduces the influence of differences between imaging devices and scanning conditions on subsequent models by normalizing image intensity values to the range [0, 1]. Standardization also improves the model's generalization ability across different datasets, ensuring consistent feature expression for images from different sources.

[0017] The denoising process removes random noise and device noise from the image while preserving the boundary details of the knee joint anatomy using anisotropic diffusion filtering, Gaussian filtering, and other methods. Denoising not only improves image quality, making subsequent feature extraction more accurate, but also improves the model's robustness in noisy environments. In addition, by suppressing noise, important detail information such as cartilage boundaries and joint spaces can be better preserved, providing high-quality input data for subsequent segmentation and modeling.

[0018] Further, the data augmentation operation in step S2 includes geometric augmentation and pixel augmentation. First, geometric augmentation simulates position and angle changes in image acquisition by random rotation, translation, scaling, and mirror flipping, enhancing the model's adaptability to spatial disturbances. Second, pixel augmentation simulates changes under different imaging conditions by adding random noise, adjusting contrast and brightness, and improving the model's processing ability for multiple data sources.

[0019] Through data augmentation techniques, the diversity of training data can be significantly increased, improving the model's adaptability to different data distributions and thus enhancing its robustness and generalization ability in knee medical image segmentation, feature extraction, and prosthesis modeling tasks. By simulating various disturbances under real imaging conditions, the model's performance in processing actual clinical images will be more stable and reliable.

[0020] Further, the anatomical structure segmentation process based on the dual-domain contrast learning network in step S3 first extracts local detail features of the knee joint image in the image domain using a convolutional neural network (CNN) based deep segmentation network. These features have significant expression ability in the anatomical details of the knee joint, such as cartilage boundaries and bone contact surfaces.

[0021] In the frequency domain, the image is domain-transformed using Fourier transform to extract high-frequency and low-frequency features to reflect the global characteristics of the knee joint anatomy. High-frequency components are mainly used to capture image boundaries and texture information, such as the junction area between cartilage and bone; low-frequency components maintain the overall lighting consistency and global morphology of the anatomical region. In addition, to reduce the interference of frequency domain noise, spectral filtering techniques such as high-pass filtering and low-pass filtering are used to further optimize the quality of frequency domain features. Through this process, the performance deficiency of traditional image domain features in complex anatomical structure boundary blur can be effectively addressed. Feature fusion mechanism

[0022] To effectively combine the features of the image domain and the frequency domain, we designed a feature fusion mechanism based on contrastive learning. Specifically, during the feature fusion process, the features of the same anatomical region are kept consistent between the two domains through the constraint of the contrastive loss function, while the feature distributions of different anatomical regions are separated.

[0023] The feature fusion module takes the feature pyramid network in the deep learning network as the core structure, and realizes the collaborative expression of high-frequency and low-frequency information through multi-scale integration, thereby improving the accuracy and robustness of segmentation. In addition, through the adversarial training strategy, the features of the image domain and the frequency domain are further aligned to ensure that the fused features have stronger robustness when segmenting complex anatomical structures.

[0024] The final output results include the accurate segmentation of the main anatomical regions of the knee joint, such as the femur, patella, tibia, and cartilage. These segmentation results are optimized by the Softmax activation function and the cross-entropy loss optimized by the Dice coefficient, ensuring that the multi-class anatomical regions have clear boundaries and high-resolution detail expression. At the same time, a boundary consistency verification mechanism is introduced in the output stage to evaluate and strengthen the global consistency of the segmentation results, avoiding the omission or boundary ambiguity of key anatomical regions.

[0025] Further, the step S4 includes performing geometric feature extraction on the segmentation results. Based on the knee joint segmentation results, anatomical key points (such as femoral condyles and tibial condyles) are extracted, and cartilage contact surfaces and bone support regions are marked to clearly define the key structural features of the knee joint. The Marching Cubes algorithm is used for three-dimensional surface reconstruction to generate a high-precision three-dimensional model of the knee joint, and through geometric parameter calculation, important features such as cartilage thickness distribution, contact area, and bone curvature are analyzed. These geometric features provide precise data support for personalized prosthesis design, optimizing the adaptability and functional recovery effect of the prosthesis and the knee joint.

[0026] Further, the step S5 includes generating a prosthesis model that adapts to the individual knee joint based on the patient's geometric data. The ISTA-Net (Iterative Sparse Transform Network) is used to generate an initial prosthesis shape, and biomechanical simulation is combined to optimize the adaptability of the prosthesis to the patient's bone and cartilage contact surface, ensuring that the prosthesis design meets the anatomical and functional requirements. Subsequently, finite element analysis is performed to verify the stress distribution and stress concentration of the prosthesis model, ensuring the mechanical stability and long-term reliability of the prosthesis during postoperative use. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 INVENTION FLOWCHART

[0028] Figure 2 Dual-domain contrastive learning network structure diagram

[0029] Figure 3 Knee dataset pre-processing results

[0030] Figure 4 ISTA-Net structure diagram

[0031] Figure 5 Three-dimensional reconstruction schematic diagram DETAILED DESCRIPTION

[0032] The following is a clear and complete description of the technical solutions of the present application in combination with the drawings in the embodiments of the present application, which is part of the embodiments of the present application rather than all the embodiments, and all other embodiments obtained by those skilled in the art without creative labor based on the embodiments in the present application belong to the protection scope of the present application.

[0033] The whole process of the present application is divided into three stages of dataset preparation, segmentation and reconstruction, as shown in Figure 1 .

[0034] I. Dataset preparation stage

[0035] In the data preprocessing stage, first, the knee image data is obtained from the CT or MRI image of the patient. The obtained image may have noise, different resolution and other problems, so the following processing is needed:

[0036] (1) Data preprocessing

[0037] Resampling: Because the resolutions of different image sources are different, the image needs to be resampled to unify the images of different resolutions to the same spatial resolution.

[0038] Assuming that the original resolution of the image is WxH, and the target resolution is W'xH', then the scaling factor of each pixel is calculated by the following formula:

[0039] S x and S y are the scaling factors in the x and y directions, respectively.

[0040] Normalization processing: the intensity value of the image is normalized to unify the intensity values of images from different sources to the same range, usually the intensity value is normalized to [0, 1], the formula is as follows:

[0041] I original is the original pixel value;

[0042] I min and I max are the minimum and maximum values in the image, respectively.

[0043] De-noising: Using classic filtering algorithms to remove noise in the image, ensuring image quality and reducing unnecessary interference.

[0044] The Gaussian filter formula is as follows:

[0045] σ is the standard deviation, which controls the smoothing degree of the filter.

[0046] (2) Data augmentation

[0047] Geometric augmentation enhances the adaptability of the model to spatial disturbances by simulating the spatial position and angle changes that may occur in the actual imaging process. Specific operations include: random rotation, translation transformation, scaling transformation and mirror flip. These geometric transformation operations effectively expand the data distribution and enhance the robustness of the model when processing spatial transformation data.

[0048] Pixel enhancement simulates changes under different imaging devices and lighting conditions by perturbing image pixel-level features such as noise, contrast and brightness, thereby improving the model's ability to adapt to cross-domain data. This includes adding Gaussian noise (formula I' = I + N(0, σ 2 ), where N(0, σ 2 ) is Gaussian noise with mean 0 and variance σ 2 ), adjusting contrast (formula I' = αI + β, where α is the contrast gain control and β is the brightness offset control), and brightness enhancement (by increasing or decreasing the global brightness value to simulate lighting changes in low-light or overexposed scenes). These pixel-level enhancement techniques effectively improve the model's generalization performance in complex imaging environments.

[0049] II. Segmentation phase

[0050] Based on the dual-domain contrast network, the network structure diagram is shown in Figure 2 , learning features in the image domain and frequency domain, and realizing effective fusion of two-domain features through contrast learning strategy, as shown in Figure 3 , the image domain uses a deep convolutional neural network to extract local detail features, focusing on the boundaries and subtle areas of anatomical structures; the frequency domain extracts global features through Fourier transform, high-frequency components capture boundary clarity, and low-frequency components maintain structural consistency. Using contrast learning mechanism, the two-domain features are aligned and optimized through contrast loss function, enhancing their complementarity and reducing domain differences. This method not only improves the model's adaptability to complex data, but also significantly improves the accuracy and robustness of the task.

[0051] (1) Dual-domain contrast learning model

[0052] Traditional segmentation methods usually rely only on image domain for feature extraction, however, the complexity of knee joint images and the details of anatomical structures require more advanced techniques. The invention introduces a dual-domain contrast learning model, combining image domain and frequency domain feature extraction, significantly improving the segmentation accuracy of knee joint anatomical regions.

[0053] Image domain features: Local anatomical features of the knee joint are extracted through a convolutional neural network (CNN), especially for accurate segmentation of the edges of structures such as cartilage and bone. An encoder-decoder structure is used, with the encoder part extracting multi-scale spatial features and the decoder part fusing high-resolution and low-resolution features through skip connections, effectively capturing local edge details and anatomical structure features of the knee joint.

[0054] Frequency domain features: Fourier transform is used to enhance the global consistency modeling capability of the image, especially for the identification of joint space and bone boundaries. The introduction of the frequency domain optimizes the understanding of the overall structure, helping to eliminate noise interference in the image and improve segmentation accuracy.

[0055] Through a feature fusion mechanism, local fine expression in the image domain is combined with global consistent modeling in the frequency domain. The feature pyramid module in the deep network and the contrast learning loss function are used to align and optimize the features of the two domains, ensuring the complementarity of global and local information. This fusion strategy significantly improves the adaptability and robustness of the segmentation model in complex medical image data, especially in low-contrast and blurred boundary image conditions, showing excellent anatomical structure segmentation ability.

[0056] (2) Training and optimization of the segmentation network

[0057] Training phase: The collected knee joint image data is used for model training. During the training process, the model automatically learns how to extract femur, patella, tibia, cartilage and other regions from complex images and generate high-precision segmentation results.

[0058] Hyperparameter optimization: To improve the segmentation accuracy of the model, the invention uses the ISTA-Net model, as shown in Figure 4 , which automatically adjusts the network weights through an iterative optimization strategy, effectively extracting key features in the knee joint image, thereby improving segmentation accuracy.

[0059] In addition, a large number of hyperparameter optimizations are performed during the training process, including learning rate, batch size, number of convolution layers, etc., to ensure that the model can adapt to the differences of various knee joint images and improve segmentation accuracy.

[0060] Model Evaluation: The segmentation performance of the model is verified on the test set using evaluation metrics such as the Dice Similarity Coefficient (Dice). The overlap between the model segmentation results and the true labeled data is compared. The Dice coefficient formula is as follows:

[0061] A is the predicted segmentation area;

[0062] B is the true labeled area.

[0063] (3) Result evaluation and optimization

[0064] The quality of the segmentation results directly affects the accuracy of subsequent prosthesis modeling. Therefore, in the segmentation stage, in addition to calculating the Dice coefficient, the Cross-Validation method is used to further evaluate the robustness of the model. By dividing the dataset into multiple subsets, k-1 subsets are used for training and the remaining 1 subset is used for testing, ensuring that the model can effectively generalize and reduce the risk of overfitting.

[0065] Three, reconstruction stage

[0066] (1) Geometric feature extraction Based on the segmentation results, key geometric features of the knee joint are extracted, including cartilage contact surface, bone support area, and anatomical key points (such as femoral condyle, tibial condyle). By automatically identifying the anatomical key points of the knee joint, the matching area of the prosthesis is precisely positioned, and key point labeling is performed. Calculate key parameters such as cartilage thickness distribution, contact area, and bone curvature to achieve geometric parameter calculation, the formula is as follows:

[0067] The cartilage thickness of point p is based on the minimum distance from point p to the bone surface point set S b .

[0068] Calculate the average unit area of the contact area S to evaluate the contact area of the cartilage and bone.

[0069] max(λ1+λ2) is the maximum principal curvature value of the bone surface, where λ1, λ2 are the eigenvalues of the curvature tensor , reflecting the curvature characteristics of the bone.

[0070] (2) Three-dimensional reconstruction

[0071] As Figure 5As shown, the Marching Cubes algorithm is used to perform three-dimensional surface reconstruction on the segmentation results, generating high-resolution knee joint three-dimensional models. Based on the extracted geometric features and segmented regions, a complete knee joint structure is generated, including the femur, patella, tibia, and cartilage regions. The surface of the reconstructed three-dimensional model is smoothed to reduce model noise caused by segmentation errors.

[0072] (3) Personalized prosthesis modeling Based on the patient's geometric features and three-dimensional reconstruction model, an adapted personalized prosthesis is generated. ISTA-Net (Iterative Sparse Transform Network) is used to generate the initial prosthesis shape, as follows:

[0073] where W represents the network weight, λ is the regularization parameter, X and Y are the input features and target output respectively.

[0074] (4) Model verification and optimization

[0075] The accuracy of the reconstruction results is verified by the Dice coefficient, and the generated prosthesis is evaluated for adaptation to the patient's knee joint anatomy, ensuring that the prosthesis precisely matches the contact surface of the cartilage and bone. The initial prosthesis model is optimized in multiple rounds, adjusting the prosthesis size, contact surface shape, and support area to ensure that its mechanical properties are fully consistent with the patient's anatomical requirements.

[0076] Although the above describes the specific embodiments of the present application in order to facilitate the understanding of the present application by those skilled in the art, it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application as defined and limited by the appended claims, and all inventions utilizing the concept of the present application are within the scope of protection.

Claims

1. A knee prosthesis reconstruction method based on a dual-domain contrast learning network, characterized in that, The method comprises the following steps: S11: acquiring magnetic resonance imaging (MRI) data or CT data in DICOM format of a knee joint of a patient; S12: cropping, resampling and standardizing the data to eliminate imaging bias and enhance data consistency, so that the data can be applied to a deep learning segmentation model; S13: using a dual-domain contrast learning model combining image domain and frequency domain features to segment the knee joint anatomical region of the preprocessed data and extract cartilage, bone and other related structures; S14: generating knee joint geometric features based on the segmentation results, especially cartilage contact surfaces and bone support areas, to provide accurate data for subsequent prosthesis modeling; S15: generating a personalized prosthesis model matching the patient's knee joint anatomy through a prosthesis modeling process.

2. The dual-domain contrastive learning network-based knee prosthesis reconstruction method according to claim 1, characterized in that, The S11 comprises: S21: acquiring MRI or CT image data of the patient's knee joint to ensure clear image quality, covering joint parts such as femur, patella and tibia, extracting data according to different image sources, ensuring that the image storage format meets the needs of subsequent processing, and pre-processing the image data to ensure that there is no missing or damaged data, and recording image metadata such as resolution and scanning parameters.

3. The dual-domain contrastive learning network-based knee prosthesis reconstruction method according to claim 1, characterized in that, The S12 comprises: S31: cropping the knee joint image to remove the background area and focus on the joint area, improving the effective data range of the image, resampling the image to adjust the resolution to a preset standard of 256x256, and ensuring the consistency of the input data; S32: through standardization processing, the intensity values of all images are adjusted to a unified range of [0, 1] to eliminate the influence of imaging differences between devices.

4. The dual-domain contrastive learning network-based knee prosthesis reconstruction method according to claim 1, characterized in that, The S13 comprises: S41: using a deep learning model based on dual-domain contrast learning to segment the knee joint anatomical structure by combining image domain and frequency domain features; S42: in the image domain, the convolutional neural network is used to extract the detailed features of the knee joint anatomical region, focusing on identifying the boundaries of cartilage, bone and other related areas; in the frequency domain, the global features are extracted by Fourier transform and other methods to optimize the global consistency modeling of the image, especially the clarity of the knee joint boundary and joint space; S43: by fusing the features of the image domain and the frequency domain through the contrast loss function, the segmentation accuracy and robustness are improved, especially in the segmentation task of complex knee joint anatomical structure.

5. The dual-domain contrastive learning network-based knee prosthesis reconstruction method according to claim 1, characterized in that, The S14 comprises: S51: extracting the geometric features of the femur, patella, tibia, cartilage and other regions through knee joint image segmentation to ensure that the prosthesis design is based on accurate anatomical data; S52: designing the geometric shape of the knee joint cartilage contact surface and bone support area to provide structured data for prosthesis generation; S53: converting the extracted geometric features into a three-dimensional model to optimize the shape of the prosthesis and ensure that the prosthesis is highly compatible with the patient's anatomical structure.

6. The dual-domain contrastive learning network-based knee prosthesis reconstruction method according to claim 1, characterized in that, The S15 comprises: S61: introducing ISTA-Net for end-to-end optimization of prosthesis modeling to automatically generate personalized prostheses using anatomical data and biomechanical features of the knee joint; S62: ISTA-Net not only optimizes the geometry of the prosthesis, but also improves the computational efficiency of the modeling process, significantly reducing the time complexity of prosthesis modeling; S63: By optimizing the network structure, it automatically adapts to the individualized anatomical features of the patient, generating a personalized prosthesis model that highly matches the patient's knee joint anatomical structure.

Images