Total knee arthroplasty postoperative image automatic analysis method and system based on cascaded prior guidance

Abstract

The application relates to the technical field of computer-aided medical diagnosis, and discloses a total knee arthroplasty postoperative image automatic analysis method and system based on cascade prior guidance. The method comprises the following steps: inputting an X-ray image into a first semantic segmentation network to generate a coarse segmentation mask; automatically generating visual prompt information based on the coarse segmentation mask; inputting the prompt information and the image into a second visual basic model to output a sub-pixel level fine segmentation mask; extracting a prosthesis edge pixel point set, performing linear fitting by singular value decomposition, and calculating a lower limb force line angle and / or a prosthesis tilt angle. The application also realizes risk stratification and follow-up tracking through radiolucent line detection, weighted K-Means clustering and time series analysis. The application solves the sub-pixel level edge extraction problem under the interference of metal artifacts. The application can be deployed in a hospital imaging department or a cloud platform and is used for TKA postoperative automatic evaluation and early warning.

Description

Technical Field

[0001] This invention relates to the field of computer-aided medical diagnostic technology, and in particular to an automatic image analysis method and system based on cascaded prior guidance after total knee arthroplasty. Background Technology

[0002] Total knee arthroplasty (TKA) is the standard surgical procedure for treating disabling osteoarthritis such as end-stage knee arthritis and rheumatoid arthritis. With the increasing aging of the global population, the number of TKA surgeries has exploded. However, TKA is not a permanent solution; postoperative complications such as prosthesis loosening, polyethylene liner wear, periprosthetic fractures, and infections are the main causes of surgical failure and revision.

[0003] Long-term, regular postoperative follow-up is the core method for monitoring surgical outcomes and identifying early complications. Currently, the gold standard for clinical follow-up remains morphological assessment based on X-rays. By comparing X-ray images at different postoperative time points, doctors can observe whether there is a radiolucent line (RLL) between the prosthesis and bone, whether the prosthesis has shifted or sunken, and the correction of the lower limb mechanical axis.

[0004] Currently, the evaluation of postoperative imaging after TKA is mainly divided into the following three approaches: Existing Method 1 (Manual Measurement): Clinicians manually mark the center of the femoral head, the center of the knee joint, and the center of the ankle joint on a standing X-ray film with long legs, and use the line connecting the two points to calculate parameters such as the hip-knee-ankle mechanical axis (HKA).

[0005] Existing Method 2 (Traditional Deep Learning): Use convolutional neural networks (CNNs) such as U-Net and DeepLab for semantic segmentation, extract the skeleton or prosthesis contour, and then perform geometric calculations.

[0006] Existing method 3 (visual foundation model): such as Segment Anything Model (SAM) and its upgraded version SAM2, performs general image segmentation through a prompting and guidance mechanism.

[0007] However, the aforementioned existing technology has the following problems: (1) Subjective bias and intra- and inter-observer variability: Manual measurements are highly subjective. Doctors with different experience may have subtle differences in their definition of anatomical landmarks (such as the selection of the center point of the tibial plateau). Studies have shown that even when the same doctor makes two measurements on the same X-ray, the difference in the HKA angle may exceed 2°. This inconsistency makes it easy for slight prosthesis displacement (a sign of early loosening) to be masked by measurement errors.

[0008] (2) "Artifact Masking" and "Contrast Trap": Numerous metallic prostheses are present in TKA postoperative images. These prostheses exhibit strong absorption in X-ray imaging, leading to severe "hardening artifacts" and "stripe artifacts" at the prosthesis edges. At the interfaces between the prosthesis and bone, and between the prosthesis and bone cement, image contrast is extremely low, and edges are blurred. Existing conventional image processing algorithms (such as traditional edge detection operators and thresholding) often fail in these areas, failing to extract true, continuous physical edges, resulting in distorted subsequent quantization calculations.

[0009] (3) Sensitivity to imaging position (rotation): TKA assessment is extremely sensitive to lower limb rotation during imaging. If the patient's lower limbs are internally or externally rotated (deviation greater than 15°) during imaging, the prosthesis morphology presented on the two-dimensional radiograph will undergo significant projection deformation, leading to a systematic error of 4°-10° in the force line measurement value. Currently, existing technologies lack an effective automatic calibration or rotation assessment mechanism, resulting in a large number of false "risk signals" caused by non-standard imaging.

[0010] Furthermore, in recent years, automatic segmentation models based on convolutional neural networks (CNNs) (such as U-Net and the DeepLab series) have been attempted for medical image analysis. However, they have the following limitations in TKA scenarios: the scarcity of training samples leads to weak model generalization ability; general CNN models often exhibit a "smoothing effect" when dealing with subpixel-level edges, failing to capture extremely subtle osteolysis radiolucency bands (usually less than 1 mm); strong semantic models (such as DeepLabv3) can identify "this is a prosthesis," but lose fine contour details; basic large models (such as SAM) have strong zero-shot segmentation capabilities, but lack medical anatomical priors, often mistaking metal artifacts for anatomical structures. At the same time, most existing auxiliary diagnostic systems only remain at the "measurement" level, failing to establish a deep link between measurement results and clinical prognosis. Doctors still need to spend a lot of effort on secondary evaluation after receiving the measurement data, failing to truly achieve intelligent risk stratification and early warning. Summary of the Invention

[0011] The purpose of this invention is to overcome at least one of the shortcomings of existing TKA postoperative image analysis, such as large manual measurement errors, strong interference from metal artifacts, and inability to automatically perform clinical risk stratification. It provides an automatic image analysis method and system based on cascaded prior guidance for total knee arthroplasty, aiming to achieve subpixel-level prosthesis edge extraction and provide objective risk warnings in combination with clinical expert experience.

[0012] The present invention adopts the following technical solution: On the one hand, the present invention provides an automatic image analysis method for postoperative total knee arthroplasty based on cascaded prior guidance, comprising the following steps: S1: Input the postoperative X-ray images of total knee arthroplasty into the first semantic segmentation network to generate a coarse segmentation mask indicating the prosthesis and bone regions; S2: Based on the coarse segmentation mask, automatically generate at least one set of positive sample points and / or an outer rectangle as visual cue information; S3: Input the visual cue information and the X-ray image together into the second visual base model. This model is a general image segmentation model based on cue guidance, which outputs sub-pixel level fine segmentation masks for prostheses and bones. S4: Extract the set of prosthesis edge pixels from the fine segmentation mask, use singular value decomposition (SVD) to perform linear fitting on the set of pixels, and calculate the lower limb force line angle and / or prosthesis tilt angle.

[0013] In addition to any of the possible implementations described above, another implementation is provided, wherein the first semantic segmentation network is selected from the DeepLab series networks (such as DeepLabv3, DeepLabv3+) and the U-Net series networks (such as U-Net, U-Net++, etc.). The second visual base model is selected from at least one of the following: Attention U-Net, nnU-Net series networks, PSPNet series networks, or semantic segmentation networks based on Transformer architecture (such as Swin-Unet, TransUNet, SegFormer); the second visual base model is selected from at least one of the SAM (Segment Anything Model) series models or general image segmentation base models based on prompt-guided mechanisms.

[0014] In addition to any of the possible implementations described above, another implementation is provided in which the first semantic segmentation network is a DeepLabv3 network, which includes a Spatial Pyramid Pooling (ASPP) module for capturing multi-scale contextual information; and the second visual base model is a Segment Anything Model 2 (SAM2), which includes an image encoder and a cue decoder based on a Transformer architecture.

[0015] In addition to any of the possible implementations described above, another implementation is provided in which the method for automatically generating the positive sample point set in step S2 is as follows: based on the coarse segmentation mask, the maximum inscribed circle algorithm is used to automatically sample positive sample points in the prosthesis and bone regions, with a sampling quantity of 5 to 15, and the bounding rectangle of the coarse segmentation mask is extracted.

[0016] In addition to any of the possible implementations described above, an implementation is further provided in which step S4 includes at least one of the following parallel quantization sub-steps: Angle measurement sub-step: Extract the set of spurious edge pixels P = {(x1, y1),(x2, y2), …, (x...} from the fine segmentation mask. n , y n )}, where n is the number of pixels in the pixel set; calculate the centroid coordinates of the pixel set P, construct a decentralized matrix M, perform singular value decomposition on M, and take the eigenvector corresponding to the smallest singular value as the direction vector of the fitted line, thereby calculating the lower limb force line angle and / or prosthesis tilt angle. Translucent line detection sub-step: Perform grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantify the millimeter value of the osteolysis translucent line based on the width of the grayscale gradient change.

[0017] In addition to any of the possible implementations described above, another implementation is provided in which, after step S4, a risk stratification step is further included, specifically: The lower limb force line angle and / or prosthesis tilt angle output by the angle measurement sub-step, and the millimeter value of the translucent line output by the translucent line detection sub-step are obtained as input features for risk assessment. The weighted K-Means clustering algorithm is used to perform cluster analysis on the input features and classify risk levels. The feature weights used in the weighted K-Means clustering algorithm are determined in advance by an index weight matrix constructed using the Analytic Hierarchy Process (AHP). The weight coefficients of each index are determined by the AHP after a consistency test (CR < 0.1).

[0018] In addition to any of the possible implementations described above, another implementation is provided in which, after step S4, a timing analysis step is further included, specifically: For multiple X-ray images of the same patient at different follow-up time points, calculate the lower limb alignment angle and / or prosthesis tilt angle at each time point according to steps S1 to S4. The calculated angle values ​​at each time point are arranged in chronological order to construct a chronological evolution feature vector, in order to identify the trend changes in the displacement or subsidence of the prosthesis.

[0019] On the other hand, the present invention also provides an automated image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance, the system being used to implement the above-described method, the system comprising: The coarse semantic segmentation module is used to input the postoperative X-ray images of total knee arthroplasty into the first semantic segmentation network and generate coarse segmentation masks that indicate the prosthesis and bone regions. The prompt generation module is used to automatically generate at least one set of positive sample points and / or an outer rectangle based on the coarse segmentation mask, as visual prompt information; The fine segmentation module is used to input the visual cue information and the X-ray image into the second visual base model. This model is a general image segmentation model based on cue guidance, which outputs subpixel-level fine segmentation masks for prostheses and bones. The geometric quantization module is used to extract the set of prosthesis edge pixels in the fine segmentation mask, perform linear fitting on the set of pixels using singular value decomposition (SVD), and calculate the lower limb force line angle and / or prosthesis tilt angle.

[0020] In addition to any of the possible implementations described above, another implementation is provided, wherein the first semantic segmentation network is selected from at least one of DeepLab series networks, U-Net series networks, nnU-Net series networks, PSPNet series networks or semantic segmentation networks based on Transformer architecture, and is used to extract multi-scale contextual semantic features and generate a coarse segmentation mask. The second visual base model is selected from at least one of the SAM series models or a general image segmentation base model based on a prompting and guidance mechanism, and is used to receive visual prompting information and output a fine segmentation mask; The system also includes a translucency line detection module, which performs grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantifies the millimeter value of the osteolysis translucency line based on the width of the grayscale gradient change.

[0021] In addition to any of the possible implementations described above, another implementation is provided in which the first semantic segmentation network is a DeepLabv3 network that includes a Spatial Pyramid Pooling (ASPP) module; The second visual foundation model is Segment Anything Model 2 (SAM2), which includes an image encoder and a cue decoder based on the Transformer architecture.

[0022] In addition to any of the possible implementations described above, another implementation is provided in which the system further includes: The risk stratification module includes: The weight storage unit is used to store the index weight matrix pre-constructed using the analytic hierarchy process (AHP); The input feature acquisition unit is used to acquire the lower limb force line angle and / or prosthesis tilt angle output by the geometric quantization module, and the millimeter value of the translucent line output by the translucent line detection module. The clustering analysis unit is connected to the weight storage unit and the input feature acquisition unit, respectively. It is used to read the index weight matrix from the weight storage unit and receive the lower limb force line angle and / or prosthesis tilt angle and the millimeter value of the translucent line from the input feature acquisition unit as input features. It uses the weighted K-Means clustering algorithm to perform clustering analysis and classify risk levels.

[0023] In addition to any of the possible implementations described above, another implementation is provided in which the system further includes: The time-series analysis module is used to acquire multiple X-ray images of the same patient at different follow-up time points. It calls the coarse semantic segmentation module, prompt generation module, fine segmentation module, and geometric quantization module to process the images at each time point, extract the geometric quantization features at each time point, construct the time-series evolution feature vector, and identify the trend changes of prosthesis displacement or subsidence.

[0024] The beneficial effects of this invention are as follows: 1. Improved Accuracy: This invention combines the global semantic capabilities of DeepLabv3 with the local refinement capabilities of SAM2 through the Cascaded Prior Guided Segmentation (CPGR) framework. Without compromising the generalization ability of large-scale models, it utilizes the prior knowledge of specialized models to constrain the search space of large-scale models. Experimental results show that this invention achieves a Dice coefficient of 0.968 on the dataset, improving the edge capture accuracy of metallic prostheses by more than 12% compared to the traditional U-Net, effectively solving the problem of sub-pixel-level edge extraction under the interference of metallic artifacts.

[0025] 2. Measurement Consistency: This invention introduces a linear fitting method based on Singular Value Decomposition (SVD). By performing singular value decomposition on the decentralized matrix, the eigenvector corresponding to the smallest singular value is taken as the direction of the fitted line. Compared with the traditional least squares method, SVD has higher robustness to edge burrs or metallic noise, and the angle measurement error is controlled within 0.5°. Experiments show that this invention exhibits high consistency across images with different exposure levels, with an inter-observer coefficient of variation (CV) of only 3.2%, far lower than the 15.8% of manual measurement.

[0026] 3. Clinical Decision Support: This invention uses the Analytic Hierarchy Process (AHP) to decompose complex clinical assessments into multi-level indicators. By constructing pairwise comparison matrices and performing a consistency test (CR < 0.1), the contribution weight of each indicator to postoperative failure risk is determined. Based on this, a weighted K-Means clustering algorithm is used to classify patients into three levels: low-risk, medium-risk, and high-risk. Experimental results show that the system's output risk level has a consistency (Kappa value) of 0.89 with the clinical assessment results of senior orthopedic experts, effectively achieving the combination of "AI measurement + expert logic".

[0027] 4. Efficiency Improvement: This invention automates the entire process of follow-up after TKA surgery. The time for complete analysis of a single image (from reading the image to outputting the risk report) is reduced from about 15-20 minutes to less than 3.5 seconds, which greatly reduces the labor cost of follow-up.

[0028] 5. Time-series tracking capability: Through the time-series analysis module, this invention can longitudinally track images of the same patient at different follow-up time points, construct a time-series evolution feature vector, identify trend changes in prosthesis displacement or subsidence, and provide a powerful tool for early detection of postoperative complications. Attached Figure Description

[0029] Figure 1 : A schematic diagram of the CPGR cascaded segmentation architecture provided by this invention.

[0030] Figure 2 The risk stratification model framework diagram of multi-source information fusion provided by this invention.

[0031] Figure 3 The following is an architecture diagram of the automatic image analysis and risk stratification assessment system provided by the present invention after total knee arthroplasty.

[0032] Figure 4 : A schematic diagram of the system interface provided by this invention.

[0033] Figure 5 Comparison of the effects of the cascaded segmentation model of this invention with single models (DeepLabv3, SAM2).

[0034] Figure 6 Comparison of the performance of the cascaded segmentation model of this invention with other state-of-the-art (SOTA) models. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0036] The accompanying drawings illustrate a layer structure according to an embodiment of the present invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0037] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0038] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0039] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0040] An embodiment of the present invention provides an automatic image analysis method for postoperative total knee arthroplasty based on cascaded prior guidance, comprising the following steps: Step S1: Input the postoperative X-ray image of total knee arthroplasty into the first semantic segmentation network to generate a coarse segmentation mask indicating the prosthesis and bone regions; Step S2: Based on the coarse segmentation mask, automatically generate at least one set of positive sample points and / or a bounding rectangle as visual cue information; Step S3: Input the visual cue information and the X-ray image together into the second visual basic model. This model is a general image segmentation model based on cue guidance, which outputs subpixel level fine segmentation masks for prostheses and bones. Step S4: Extract the set of prosthesis edge pixels from the fine segmentation mask, use singular value decomposition (SVD) to perform linear fitting on the set of pixels, and calculate the lower limb force line angle and / or prosthesis tilt angle.

[0041] In one specific embodiment, step S4 includes at least one of the following parallel quantization sub-steps: Angle measurement sub-step: Extract the set of spurious edge pixels P = {(x1, y1),(x2, y2), …, (x...} from the fine segmentation mask. n , y n )}, where n is the number of pixels in the pixel set; calculate the centroid coordinates of the pixel set P, construct a decentralized matrix M, perform singular value decomposition on M, and take the eigenvector corresponding to the smallest singular value as the direction vector of the fitted line, thereby calculating the lower limb force line angle and / or prosthesis tilt angle. Translucent line detection sub-step: Perform grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantify the millimeter value of the osteolysis translucent line based on the width of the grayscale gradient change.

[0042] In one specific embodiment, the method further includes the following risk stratification steps: obtaining the lower limb force line angle and / or prosthesis tilt angle output by the angle measurement sub-step, and the millimeter value of the translucent line output by the translucent line detection sub-step, as input features for risk assessment; using a weighted K-Means clustering algorithm to perform cluster analysis on the input features and classify risk levels; wherein, the feature weights used in the weighted K-Means clustering algorithm are pre-determined through an indicator weight matrix constructed by the analytic hierarchy process (AHP), and the weight coefficients of each indicator are determined by the AHP after a consistency test (CR<0.1).

[0043] In one specific embodiment, the method further includes the following time-series analysis steps: for multiple X-ray images of the same patient at different follow-up time points, calculate the lower limb force line angle and / or prosthesis tilt angle at each time point according to steps S1 to S4; arrange the calculated angle values ​​at each time point in time sequence to construct a time-series evolution feature vector to identify the trend changes of prosthesis displacement or subsidence.

[0044] An embodiment of the present invention provides an automatic image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance, comprising: The coarse semantic segmentation module is used to input the postoperative X-ray images of total knee arthroplasty into the first semantic segmentation network and generate coarse segmentation masks that indicate the prosthesis and bone regions. The prompt generation module is used to automatically generate at least one set of positive sample points and / or an outer rectangle based on the coarse segmentation mask, as visual prompt information; The fine segmentation module is used to input the visual cue information and the X-ray image into the second visual base model. This model is a general image segmentation model based on cue guidance, which outputs subpixel-level fine segmentation masks for prostheses and bones. The geometric quantization module is used to extract the set of prosthesis edge pixels in the fine segmentation mask, perform linear fitting on the set of pixels using singular value decomposition (SVD), and calculate the lower limb force line angle and / or prosthesis tilt angle.

[0045] In one specific embodiment, the first semantic segmentation network is selected from at least one of the DeepLab series networks (such as DeepLabv3, DeepLabv3+), U-Net series networks (such as U-Net, U-Net++, Attention U-Net), nnU-Net series networks, PSPNet series networks, or semantic segmentation networks based on the Transformer architecture (such as Swin-Unet, TransUNet, SegFormer), used to extract multi-scale contextual semantic features and generate a coarse segmentation mask; the second visual base model is selected from at least one of the SAM (Segment Anything Model) series models or a general image segmentation base model based on a prompt-guided mechanism, used to receive visual prompt information and output a fine segmentation mask; the system also includes a translucency line detection module, used to perform grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantize the millimeter value of the osteolysis translucency line based on the width of the grayscale gradient change.

[0046] In one specific embodiment, the first semantic segmentation network is a DeepLabv3 network, which includes a Spatial Pyramid Pooling (ASPP) module; the second visual base model is a Segment Anything Model 2 (SAM2), which includes an image encoder and a cue decoder based on the Transformer architecture.

[0047] In one specific embodiment, the system further includes a risk stratification module, comprising: a weight storage unit for storing an indicator weight matrix pre-constructed using the Analytic Hierarchy Process (AHP); an input feature acquisition unit for acquiring the lower limb force line angle and / or prosthesis tilt angle output by the geometric quantification module, and the millimeter value of the translucent line output by the translucent line detection module; and a clustering analysis unit connected to the weight storage unit and the input feature acquisition unit, for reading the indicator weight matrix from the weight storage unit and receiving the lower limb force line angle and / or prosthesis tilt angle and the millimeter value of the translucent line from the input feature acquisition unit as input features, performing clustering analysis using a weighted K-Means clustering algorithm to classify risk levels.

[0048] In one specific embodiment, the system further includes a time series analysis module, which is used to acquire multiple X-ray images of the same patient at different follow-up time points, and call the coarse semantic segmentation module, prompt generation module, fine segmentation module and geometric quantization module to process the images at each time point respectively, extract the geometric quantization features at each time point, construct a time series evolution feature vector, and identify the trend changes of prosthesis displacement or subsidence.

[0049] Example 1: CPGR Cascaded Partition Architecture like Figure 1 As shown, this embodiment provides a CPGR (Cascade Prior-Guided Refinement) cascaded partitioning architecture process.

[0050] The main process nodes in the diagram include: 101 - First Stage: Using DeepLabv3, coarse semantic segmentation is performed on the standardized X-ray images to generate preliminary class probability maps and coarse binary masks (M_coarse), quickly locating the approximate regions of the prosthesis and bone. The DeepLabv3 network includes a Hollow Spatial Pyramid Pooling (ASPP) module, which can capture multi-scale contextual information and remove interference objects (such as surgical staples and skin edges) in complex backgrounds.

[0051] 102-Intermediate Processing: A prompt message (M_prompt) is generated based on the coarse mask, and the search band is expanded using edge buffering to enhance alignment robustness. Specifically, the maximum inscribed circle algorithm is used to automatically sample positive sample points within the prosthesis and bone regions, with a sampling quantity of 5 to 15 (optimal value 10), and the bounding rectangle of the coarse segmentation mask is extracted. These points and boxes are then encoded as prompts and input into the SAM2 model.

[0052] 103-Post-processing and structural decoupling: SAM2 utilizes its Transformer architecture image encoder to extract features and combines them with prompts for fine-grained mask decoding, outputting a high-precision final segmentation result. The fine-grained mask (M_final) output by SAM2 is further decoupled from the topology of the prosthesis and skeleton, separating them into their respective independent structural representations.

[0053] 104-Skeleton Extraction and Branch Analysis: A minimum cut algorithm based on distance transformation is used to generate a distance weight map D(x,y), and a single-pixel-width skeletal skeleton S is extracted from it. At the same time, key bifurcation points P_branch are identified for subsequent morphological analysis or clinical measurement.

[0054] This model not only achieves high-precision pixel-level segmentation results, but also extracts anatomically significant structural information (such as skeleton and branch points), providing support for downstream tasks such as prosthesis localization and alignment evaluation.

[0055] Example 2: Risk Stratification Model Based on Multi-Source Information Fusion like Figure 2 As shown in the figure, this embodiment provides a risk stratification model framework for multi-source information fusion.

[0056] The main process nodes in the diagram include: 201 - Data Input Layer: Acquires feature inputs in three dimensions: geometric quantification features, temporal evolution features, and deep semantic features. Among them, geometric quantification features include lower limb force line angle and prosthesis tilt angle; temporal evolution features include the angle change trend at multiple time points; and deep semantic features include the width of the translucent line.

[0057] 202-Feature Fusion Layer: The Analytic Hierarchy Process (AHP) is used to transform clinical expert experience into subjective weights for each feature. Specifically, pairwise comparison judgment matrices for each assessment indicator are established through expert surveys. After consistency testing (CR < 0.1), the weight coefficients of each indicator are determined. For example, the HKA deviation weight W1 = 0.45, the RLL width weight W2 = 0.35, etc. The three types of heterogeneous features after fusion are weighted to form a unified high-dimensional feature vector for subsequent analysis.

[0058] 203-Computational Layer: The K-Means clustering algorithm is improved by replacing Euclidean distance with Mahalanobis distance that incorporates feature weights, enhancing the clustering's ability to distinguish risk patterns. After normalizing the continuous measured values, patients are divided into three clusters: Cluster 1 (low risk): indicators are within the clinical safety threshold; Cluster 2 (medium risk): indicators show deviations (e.g., HKA deviation between 3° and 5°), observation is recommended; Cluster 3 (high risk): indicators exceed the warning line (e.g., RLL > 2 mm), and the system automatically identifies "revision recommended".

[0059] 204-Output Layer: Maps clustering results to clinically interpretable risk levels (low, medium, and high risk) to support decision-making and intervention.

[0060] These modules together constitute an intelligent risk stratification system that integrates multi-source information and combines expert knowledge with data-driven methods.

[0061] Example 3: System Overall Architecture As attached Figure 3 As shown in the figure, this embodiment provides the overall architecture of an automatic image analysis and risk stratification assessment system after total knee arthroplasty.

[0062] The main process nodes in the diagram include: DICOM Image and User Data Upload: Responsible for receiving image data requests from the front end and sending back the processing results in JSON format.

[0063] Cascaded Segmentation and Quantitative Assessment Model: This model encapsulates the Cascaded Segmentation Model (CPGR) and the Quantitative Assessment Model, and is responsible for converting the model inference results into clinically usable geometric parameters and risk labels.

[0064] Data storage: Store user information, patient files, historical follow-up records, etc., to ensure data traceability.

[0065] Visualization and Report Generation: Provides functions such as visualization, overlay rendering of segmentation results, and generation of risk assessment reports.

[0066] This figure illustrates the complete data flow from the input of the original DICOM image to the final generation of the risk stratification report.

[0067] Example 4: System Interface like Figure 4 As shown in the figure, this embodiment provides an interface diagram of an automatic image analysis and risk stratification assessment system after total knee arthroplasty.

[0068] The main modules in the diagram include: 401 - Upper Interactive Area: Clinicians can upload various patient data here, such as ID, follow-up time, height and weight, etc. Click to upload and analyze to get the results.

[0069] 402-Intermediate Report Area: Displays detailed measurement data of the current image (such as joint space ratio, force line angle), as well as risk scores and risk groups obtained from clustering.

[0070] 403-Lower Image Area: Displays the original X-ray image and the segmentation mask generated by the algorithm, as well as the automatically drawn force line reference plane, etc.

[0071] The image shows the user interaction page of the evaluation system.

[0072] Example 5: Comparison of Segmentation Results like Figure 5 As shown in the figure, this embodiment provides a comparison chart of the effects of the cascaded segmentation model of the present invention and the single model (DeepLabv3, SAM2).

[0073] After cascading the first two models, the visual effect changed significantly: Compared to DeepLabv3, its edges are smoother and more continuous, allowing it to better follow the contours of the implant; Compared to SAM2, it effectively reduces metallic artifacts and avoids over-segmentation.

[0074] This confirms that the cascading paradigm proposed in this invention is not a simple stacking of modules, but rather the optimal solution to the contradiction between "global semantics" and "local precision".

[0075] Example 6: Comparison with other state-of-the-art (SOTA) models like Figure 6 As shown in the figure, this embodiment provides a comparison chart of the performance of the cascaded segmentation model of the present invention with other state-of-the-art (SOTA) models.

[0076] As shown in the figure, other models often exhibit "fracture" or "spillover" phenomena at the prosthesis-bone interface, frequently failing to accurately locate the boundary, resulting in a visually blurred boundary that cannot meet the requirements for sub-millimeter-level measurements. In contrast, the model proposed in this invention maintains a leading position in regional overlap and surpasses existing mainstream models in the edge distance metric, which is of utmost clinical concern.

[0077] This verifies that introducing the visual foundation model (SAM2) and supplementing it with strong semantic prior guidance is an effective way to resolve the contradiction between "metal artifact interference" and "sub-pixel level edge localization" in medical images.

[0078] Specific implementation parameters range: In the specific implementation of this invention, the following parameter ranges are preferably adopted: Image resolution: The longer side of the input image must be no less than 2048 pixels to ensure sub-pixel extraction effect.

[0079] Model training hyperparameters: The suggested range for Learning Rate is [1e-5, 5e-4], and the Batch Size is 8-16.

[0080] AHP consistency threshold: CR must be less than 0.1, otherwise the weight matrix needs to be readjusted.

[0081] Number of positive sample points: 5-15, with an optimal value of 10.

[0082] Industrial applicability: This invention can be deployed in hospital radiology departments and orthopedic follow-up centers, and can also be integrated into PACS systems or cloud AI platforms. Through fully automated TKA postoperative image analysis, this invention can significantly improve measurement consistency and accuracy (reaching sub-millimeter level), and can identify high-risk patients at an early stage, which has important clinical application value for alleviating the pressure on medical resources and improving the quality of follow-up.

[0083] The above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​this application; at the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0084] Certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The terms "comprising" and "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising / including but not limited to". "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error. The following descriptions in the specification are preferred embodiments for carrying out this application; however, these descriptions are for the purpose of illustrating the general principles of this application and are not intended to limit the scope of this application. The scope of protection of this application shall be determined by the appended claims.

[0085] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0086] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0087] The foregoing description illustrates and describes several preferred embodiments of this application. However, as previously stated, it should be understood that this application is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the application concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of this application should be within the protection scope of the appended claims.

Claims

1. A method for automatic image analysis after total knee arthroplasty based on cascaded prior guidance, characterized in that, Includes the following steps: S1: Input the postoperative X-ray images of total knee arthroplasty into the first semantic segmentation network to generate a coarse segmentation mask indicating the prosthesis and bone regions; S2: Based on the coarse segmentation mask, automatically generate at least one set of positive sample points and / or an outer rectangle as visual cue information; S3: Input the visual cue information and the X-ray image together into the second visual base model. This model is a general image segmentation model based on cue guidance, which outputs sub-pixel level fine segmentation masks for prostheses and bones. S4: Extract the set of prosthesis edge pixels from the fine segmentation mask, use singular value decomposition (SVD) to perform linear fitting on the set of pixels, and calculate the lower limb force line angle and / or prosthesis tilt angle.

2. The automatic image analysis method for total knee arthroplasty based on cascaded prior guidance as described in claim 1, characterized in that, The first semantic segmentation network is selected from at least one of the DeepLab series networks, U-Net series networks, nnU-Net series networks, PSPNet series networks, or semantic segmentation networks based on the Transformer architecture; the second visual base model is selected from at least one of the SAM series models or general image segmentation base models based on prompting and guidance mechanisms.

3. The automatic image analysis method for postoperative total knee arthroplasty based on cascaded prior guidance as described in claim 1, characterized in that, The method for automatically generating the positive sample point set in step S2 is as follows: based on the coarse segmentation mask, the maximum inscribed circle algorithm is used to automatically sample positive sample points in the prosthesis and bone regions, with a sampling quantity of 5 to 15, and the outer rectangle of the coarse segmentation mask is extracted.

4. The automatic image analysis method for total knee arthroplasty based on cascaded prior guidance as described in claim 1, characterized in that, Step S4 includes at least one of the following parallel quantization sub-steps: Angle measurement sub-step: Extract the set of spurious edge pixels P = {(x1, y1), (x2, y2), …, (x...} from the fine segmentation mask. n , y n )}, where n is the number of pixels in the pixel set; Calculate the centroid coordinates of the pixel set P, construct a decentered matrix M, perform singular value decomposition on M, and take the feature vector corresponding to the smallest singular value as the direction vector of the fitted line, thereby calculating the lower limb force line angle and / or prosthesis tilt angle. Translucent line detection sub-step: Perform grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantify the millimeter value of the osteolysis translucent line based on the width of the grayscale gradient change.

5. The automatic image analysis method for total knee arthroplasty based on cascaded prior guidance as described in claim 4, characterized in that, Following step S4, a risk stratification step is also included, specifically: The lower limb force line angle and / or prosthesis tilt angle output by the angle measurement sub-step, and the millimeter value of the translucent line output by the translucent line detection sub-step are obtained as input features for risk assessment. The weighted K-Means clustering algorithm is used to perform cluster analysis on the input features and classify risk levels. The feature weights used in the weighted K-Means clustering algorithm are determined in advance by constructing an index weight matrix using the Analytic Hierarchy Process (AHP). The AHP determines the weight coefficients of each index after a consistency check.

6. The automatic image analysis method for total knee arthroplasty based on cascaded prior guidance as described in claim 1, characterized in that, Following step S4, a timing analysis step is also included, specifically: For multiple X-ray images of the same patient at different follow-up time points, calculate the lower limb alignment angle and / or prosthesis tilt angle at each time point according to steps S1 to S4. The calculated angle values ​​at each time point are arranged in chronological order to construct a chronological evolution feature vector, in order to identify the trend changes in the displacement or subsidence of the prosthesis.

7. An automated image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance, characterized in that, The system is used to implement the method as described in any one of claims 1-6, the system comprising: The coarse semantic segmentation module is used to input the postoperative X-ray images of total knee arthroplasty into the first semantic segmentation network and generate coarse segmentation masks that indicate the prosthesis and bone regions. The prompt generation module is used to automatically generate at least one set of positive sample points and / or an outer rectangle based on the coarse segmentation mask, as visual prompt information; The fine segmentation module is used to input the visual cue information and the X-ray image into the second visual base model. This model is a general image segmentation model based on cue guidance, which outputs subpixel-level fine segmentation masks for prostheses and bones. The geometric quantization module is used to extract the set of prosthesis edge pixels in the fine segmentation mask, perform linear fitting on the set of pixels using singular value decomposition (SVD), and calculate the lower limb force line angle and / or prosthesis tilt angle.

8. The automatic image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance as described in claim 7, characterized in that, The first semantic segmentation network is selected from at least one of the DeepLab series network, U-Net series network, nnU-Net series network, PSPNet series network or semantic segmentation network based on Transformer architecture, and is used to extract multi-scale contextual semantic features and generate a coarse segmentation mask. The second visual base model is selected from at least one of the SAM series models or a general image segmentation base model based on a prompting and guidance mechanism, and is used to receive visual prompting information and output a fine segmentation mask; The system also includes a translucency line detection module, which performs grayscale scanning along the normal direction of the prosthesis edge in the fine segmentation mask, and quantifies the millimeter value of the osteolysis translucency line based on the width of the grayscale gradient change.

9. The automatic image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance as described in claim 8, characterized in that, The system also includes: The risk stratification module includes: The weight storage unit is used to store the index weight matrix pre-constructed using the analytic hierarchy process (AHP). The input feature acquisition unit is used to acquire the lower limb force line angle and / or prosthesis tilt angle output by the geometric quantization module, and the millimeter value of the translucent line output by the translucent line detection module. The clustering analysis unit is connected to the weight storage unit and the input feature acquisition unit, respectively. It is used to read the index weight matrix from the weight storage unit and receive the lower limb force line angle and / or prosthesis tilt angle and the millimeter value of the translucent line from the input feature acquisition unit as input features. It uses the weighted K-Means clustering algorithm to perform clustering analysis and classify risk levels.

10. The automatic image analysis system for postoperative total knee arthroplasty based on cascaded prior guidance as described in claim 7, characterized in that, The system also includes: The time-series analysis module is used to acquire multiple X-ray images of the same patient at different follow-up time points. It calls the coarse semantic segmentation module, prompt generation module, fine segmentation module, and geometric quantization module to process the images at each time point, extract the geometric quantization features at each time point, construct the time-series evolution feature vector, and identify the trend changes of prosthesis displacement or subsidence.

Images