Knee osteoarthritis x-ray explainable grading method and system based on knowledge graph fusion

Abstract

This invention discloses a knowledge graph fusion-based method and system for interpretable X-ray grading of knee osteoarthritis. The method includes multi-directional contour extraction and predefined structural contour template matching of target X-ray images of the knee joint to obtain stable contour results; pathological quantitative analysis of the stable contour results to obtain pathological feature vectors; obtaining the similarity between the pathological feature vectors corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples; constructing a symmetric normalized adjacency matrix based on the similarity; and inputting the target X-ray image and the symmetric normalized adjacency matrix into a pre-trained graph convolutional fusion diagnostic network model to output the grading result of knee osteoarthritis. This invention, based on pathological knowledge, converts the stable contour results of knee joint X-ray images into a symmetric normalized adjacency matrix, which is then fused with the extracted data features of the target X-ray image in the graph convolutional fusion diagnostic network model to accurately identify the grading result of knee osteoarthritis.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis and assisted diagnosis of medical images, specifically to a method and system for interpretable grading of knee osteoarthritis X-rays based on knowledge graph fusion. Background Technology

[0002] Clinical grading of knee osteoarthritis typically relies on X-ray images, with the Kellgren-Lawrence grading system being the commonly used standard. Traditional manual interpretation depends on physician experience, resulting in high subjectivity, time-consuming processes, and difficulties in ensuring consistency. To improve efficiency, existing technologies widely employ deep learning methods for automated grading, with a typical approach using convolutional neural networks as the core to directly learn discriminative features from X-ray images and output grading results. However, most existing deep learning-based automated grading methods are black-box models, lacking sufficient interpretability, which poses a significant challenge for medical applications.

[0003] While some studies have attempted to interpret the model's region of interest through visualization heatmaps, attention mechanisms, or class activation mapping, these methods mostly only provide clues in the post-processing stage and struggle to establish clear correspondences with commonly used clinical pathological indicators (such as osteophytes, joint space width, and changes in left-right symmetry). This fails to provide doctors with direct and clear decision-making support. Consequently, the application of existing deep learning models in medical settings is limited, especially in clinical environments where high traceability and reliability are required.

[0004] Furthermore, although some techniques have attempted to incorporate explicit medical priors for grading, such as measuring joint space width, left-right symmetry changes, and detecting osteophytes in the joint region, and then using rule-based or traditional classifiers for grading, these methods rely on the robustness of contour and edge extraction. However, traditional algorithms are highly sensitive to images of varying severity, especially under different shooting postures or image quality conditions, easily leading to incomplete contours, noise interference, and edge distortion, thus affecting the stability and accuracy of the grading results. At the same time, these methods often fail to fully utilize the powerful feature learning capabilities of deep learning models, limiting grading accuracy and generalization ability.

[0005] Therefore, existing technologies have significant shortcomings in terms of interpretability, robustness, and effectiveness in combining deep learning and medical knowledge. There is an urgent need for a new approach to address these issues and further improve the performance and clinical applicability of automated grading systems. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a knowledge graph fusion-based X-ray interpretable grading method and system for knee osteoarthritis. Based on pathological knowledge, the stable contour results of the knee joint X-ray images are converted into a symmetric normalized adjacency matrix. The symmetric normalized adjacency matrix is ​​then fused with the extracted data features of the target X-ray image in a graph convolutional fusion diagnostic network model to accurately identify the grading results of knee osteoarthritis.

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0008] According to a first aspect of this application, a knowledge graph fusion-based method for interpretable X-ray grading of knee osteoarthritis is provided, comprising: Obtain target X-ray images of the knee joint; Multi-directional contour extraction is performed on the target X-ray image to obtain the initial contour result of the knee joint. The initial contour result is then corrected using a predefined structural contour template matching strategy to obtain a stable contour result. The stable contour results were subjected to pathological quantitative analysis to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. Obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples, and construct a symmetric normalized adjacency matrix based on the similarity. The target X-ray image and the symmetric normalized adjacency matrix are input into a pre-trained graph convolutional fusion diagnostic network model, which outputs the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is trained through a sample set, which includes several sets of pre-trained X-ray images and corresponding symmetric normalized adjacency matrices.

[0009] In some embodiments of this application, based on the aforementioned scheme, the multi-directional contour extraction of the target X-ray image is performed to obtain the initial contour result of the knee joint: Obtain a grayscale image of the target X-ray image; Based on the multi-directional Sobel operator, the gradient convolution kernels and corresponding gradient responses of the grayscale image in the horizontal, vertical and two diagonal directions are extracted respectively, and the gradient responses in all directions are fused to obtain the total edge response map. Based on the total edge response map, the set of non-zero edge responses of the joint region to be processed is obtained; The set of non-zero edge responses is filtered based on the threshold search range to generate a set of candidate thresholds. The number of non-zero points for each candidate threshold is obtained, and an objective function is constructed based on the absolute value of the difference between the number of non-zero points of the candidate thresholds and the target number of non-zero points. The candidate threshold corresponding to the minimum objective function value is used as the adaptive dynamic threshold, and the binary edge map extracted from the grayscale image based on the adaptive dynamic threshold is used as the initial contour result.

[0010] In some embodiments of this application, based on the foregoing scheme, the step of correcting the initial contour result using a predefined structural contour template matching strategy to obtain a stable contour result includes: Set a predefined structural contour template set, which includes several predefined structural contour templates; Obtain the set of candidate edge points corresponding to the initial contour result; For each predefined structural contour template, a search displacement set is set, the search displacement set includes several search displacement vectors, and a candidate contour template set is obtained after each predefined structural contour template is translated based on several search displacement vectors, the candidate contour template set includes several candidate contour templates; Based on the candidate edge coverage and the predefined structural contour template overlap rate, a matching function is established to weightedly fuse the candidate edge point set and the candidate contour template, and the candidate contour template corresponding to the search displacement vector that makes the highest matching score is obtained. The initial contour result is corrected based on the candidate contour template to obtain a stable contour result.

[0011] In some embodiments of this application, based on the foregoing scheme, the pathological quantitative analysis of the stable contour results to obtain pathological feature vectors based on osteophyte, joint space changes, and left-right symmetry changes includes: For the stable contour results, two contour curves are generated to describe the joint gap. For each contour curve, the distance sequence between the two contours is calculated point by point along the vertical direction to obtain the distance list corresponding to the two contour curves respectively. Based on the two sets of distance lists, obtain the point pair sets respectively; Based on two sets of distance lists and two sets of point pairs, mean features, variance features, range features, maximum distance features, and overlap features are extracted.

[0012] In some embodiments of this application, based on the foregoing scheme, the step of obtaining the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples, and constructing a symmetric normalized adjacency matrix based on the similarity, includes: The feature matrix is ​​obtained by combining the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples; Based on Gaussian kernel similarity, the feature matrix is ​​converted into a similarity matrix; The similarity matrix is ​​converted into a sparse matrix using the K-nearest neighbor sparsification method. An adjacency matrix is ​​constructed based on a sparse matrix, and a self-loop is added to the adjacency matrix to obtain a degree matrix; Based on the adjacency matrix and the degree matrix, a symmetric normalized adjacency matrix is ​​constructed.

[0013] In some embodiments of this application, based on the aforementioned scheme, the graph convolutional fusion diagnostic network model includes a convolution and pooling unit, a graph convolutional fusion unit, and a fully connected output unit. The convolution and pooling unit is used to extract depth features from the target X-ray image to obtain the initial node features of each sample node. The graph convolutional fusion unit is used to perform feature propagation and aggregation on the initial node features and the normalized adjacency matrix to obtain node features. The fully connected output unit is used to output the classification results, which include the classification results of knee osteoarthritis and the corresponding classification probabilities.

[0014] In some embodiments of this application, based on the foregoing scheme, the convolution and pooling unit extracts depth features from the target X-ray image through multiple sequentially connected CNN feature extraction sub-units; The graph convolutional fusion unit performs feature propagation and aggregation on the initial node features and the normalized adjacency matrix through multiple sequentially connected GCN fusion sub-units, wherein the GCN fusion sub-unit is a graph convolutional network; The fully connected output unit includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used to output the classification result, and the second fully connected layer is used to output the classification probability.

[0015] According to a second aspect of this application, a knowledge graph fusion-based X-ray interpretable grading system for knee osteoarthritis is provided, comprising: The first acquisition module is used to acquire target X-ray images of the knee joint; The second acquisition module is used to extract the contour of the target X-ray image in multiple directions to obtain the initial contour result of the knee joint, and to correct the initial contour result by a predefined structural contour template matching strategy to obtain a stable contour result. The third acquisition module is used to perform pathological quantitative analysis on the stable contour results to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. A construction module is used to obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vector corresponding to several samples, and to construct a symmetric normalized adjacency matrix based on the similarity. The fourth acquisition module inputs the target X-ray image and the symmetric normalized adjacency matrix into a pre-trained graph convolutional fusion diagnostic network model and outputs the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is obtained by training a sample set and the corresponding symmetric normalized adjacency matrix set of the sample set. The sample set includes several sets of pre-trained X-ray images.

[0016] According to a third aspect of this application, a computer-readable storage medium is provided that stores a computer program thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0017] According to a fourth aspect of this application, an electronic device is provided, comprising: One or more processors; A memory for storing executable instructions of the processor, which, when executed by the one or more processors, cause the one or more processors to implement the method described above.

[0018] The beneficial effects of this application are as follows: (1) The knowledge graph fusion-based X-ray interpretable grading method and system for knee osteoarthritis provided in this application uses multi-directional contour extraction, including multi-directional gradient edge calculation, adaptive dynamic threshold selection, contour refinement based on predefined structural contour template matching, and reclassification and optimization strategies for noisy contours, to stably obtain stable contour results of knee joints under different severity and different imaging postures.

[0019] (2) The knowledge graph fusion-based X-ray interpretable grading method and system for knee osteoarthritis provided in this application transforms stable contour results into quantitative features such as mean, variance, range, maximum distance, and overlap, and establishes the correspondence between them and osteophytes, joint space width, and symmetrical joint space width, thereby realizing the computable expression of medical concepts.

[0020] (3) The knowledge graph fusion-based X-ray interpretable grading method and system for knee osteoarthritis provided in this application establishes a symmetric normalized adjacency matrix based on graph knowledge structure, calculates similarity and association weights based on quantitative features, forms a graph structure knowledge representation and uses it as medical knowledge input, and uses the symmetric normalized adjacency matrix to guide the update of image features in graph convolution propagation, thereby achieving endogenous interpretability of grading results. Compared with the end-to-end grading scheme based solely on convolutional neural networks, this embodiment injects medical knowledge into the graph convolutional fusion diagnostic network model in the form of a symmetric normalized adjacency matrix, so that the graph convolutional fusion diagnostic network model is naturally constrained and guided by clinical pathological evidence in the process of feature propagation and aggregation, and shifts from posterior interpretation to endogenous interpretation in terms of interpretation mechanism, which is more in line with the requirements of interpretability and traceability in medical scenarios.

[0021] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are intended to explain the invention, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph fusion-based X-ray interpretable grading method for knee osteoarthritis according to the present invention; Figure 2 This is a schematic diagram of the overall framework and information flow of a knowledge graph fusion-based X-ray interpretable grading method for knee osteoarthritis according to the present invention. Figure 3 (a) is a schematic diagram of the distribution of candidate contour segments obtained by the multi-directional gradient response of the present invention; Figure 3 (b) is the initial contour result obtained on the target X-ray image of the present invention; Figure 3 (c) is a schematic diagram of the present invention segmenting and marking the stable contour results according to anatomical structure; Figure 4 The ExpGCN network architecture of this invention is used to integrate medical knowledge and data knowledge; Figure 5 This is a schematic diagram of an X-ray interpretable grading system for knee osteoarthritis based on knowledge graph fusion, according to the present invention. Figure 6 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0023] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0025] In related technologies, when grading knee osteoarthritis, end-to-end grading methods based solely on convolutional neural networks lack intrinsic interpretability, making it difficult to establish a clear correspondence between predicted results and clinical evidence such as osteophytes, joint space changes, and bilateral symmetry changes, affecting clinical credibility and practical application. Furthermore, traditional schemes based solely on explicit pathological measurements rely on the robustness of contour and edge extraction; traditional operators are prone to contour discontinuities, noise interference, or incomplete extraction in images of different severity levels, leading to unstable indicators and affecting diagnostic stability. Additionally, existing schemes combining medical indicators and deep features often lack a unified structured expression, making it difficult for medical knowledge to participate in deep network training and inference in a computable manner, resulting in insufficient fusion. This embodiment provides a knowledge graph fusion-based X-ray interpretable grading method for knee osteoarthritis to solve the above problems. It extracts complete contours of the knee joint, ensuring diagnostic stability, and provides a unified structured expression for the scheme combining medical indicators and deep features, enabling a clear correspondence between predicted grading results and clinical evidence such as osteophytes, joint space changes, and bilateral symmetry changes.

[0026] According to the first aspect of this application, Figure 1 and Figure 2 As shown, this embodiment provides a knowledge graph fusion-based X-ray interpretable grading method for knee osteoarthritis, including: Step S1: Obtain the target X-ray image of the knee joint.

[0027] Step S2: Extract the contours of the target X-ray image from multiple directions to obtain the initial contour result of the knee joint. Then, correct the initial contour result by using a predefined structural contour template matching strategy to obtain a stable contour result.

[0028] Step S2 in this embodiment is used to obtain stable contour results of key bony structures of the knee joint under different shooting postures, different imaging qualities, and different lesion severity, so as to provide reliable input for subsequent pathological quantitative analysis and graph convolutional network construction.

[0029] In some embodiments of this example, the step of extracting the multi-directional contour of the target X-ray image to obtain the initial contour result of the knee joint includes: Step S201: Obtain a grayscale image of the target X-ray image; Step S202: Based on the multi-directional Sobel operator, extract the gradient convolution kernels and corresponding gradient responses of the grayscale image in the horizontal, vertical and two diagonal directions respectively, and fuse the gradient responses in all directions to obtain the total edge response map. Step S203: Obtain the set of non-zero edge responses for the joint region to be processed based on the total edge response map; Step S204: Filter the set of non-zero edge responses based on the threshold search range to generate a set of candidate thresholds, obtain the number of non-zero points for each candidate threshold, and construct an objective function based on the absolute value of the difference between the number of non-zero points of the candidate thresholds and the number of non-zero points of the target threshold. Use the candidate threshold corresponding to the minimum objective function value as the adaptive dynamic threshold, and use the binary edge map extracted from the grayscale image based on the adaptive dynamic threshold as the initial contour result.

[0030] In this embodiment, the target X-ray image is adjusted to be an orthogonal image, which facilitates the extraction of gradient convolution kernels and corresponding gradient responses of the grayscale image in the horizontal, vertical and two diagonal directions.

[0031] In this embodiment, in step S202, the input grayscale image is set to... ,in, This represents the pixel coordinates in the input grayscale image. This indicates the input grayscale image in pixel coordinates. The grayscale value at that location. Gradient convolution kernels are selected in four directions. These correspond to the horizontal, vertical, and two diagonal directions (0°, 90°, 45°, 135°), respectively. The Sobel convolution kernel is then defined as:

[0032] Multidirectional gradient response is defined as:

[0033] in, Indicates the first Convolutional kernels in each direction Under the influence of the pixel coordinates The gradient response magnitude at the location; Represents a two-dimensional convolution operation; Indicates taking the absolute value; Indicates direction index, This represents a grayscale image. The total edge response map is obtained by fusing gradient responses from all directions.

[0034] in, Represents pixel coordinates The total edge response value after fusing gradient responses from all directions.

[0035] This embodiment employs an adaptive dynamic thresholding strategy to filter and connect edge candidates. Since background noise, soft tissue shadows, and exposure differences in the target X-ray image cause variations in edge response intensity, a fixed threshold is prone to two problems: when the threshold is too high, true contours are easily missing or discontinuous; when the threshold is too low, noisy edges and false contours are easily introduced, increasing the difficulty of subsequent contour selection. Therefore, this embodiment adaptively determines the threshold based on the edge response distribution of the current image to balance contour integrity and noise suppression capabilities.

[0036] In this embodiment, in step S203, the joint region to be processed is set as follows: The total edge response map is And denote that the set of non-zero edge responses within this region includes several non-zero total edge response maps, and the calculation formula is:

[0037] remember For set The quantile value.

[0038] In this embodiment, in step S204, to reduce the impact of extreme noise response on the threshold setting, the lower and upper bounds of the threshold search range are set as follows:

[0039] in, , and The preset quantile coefficient is used to adaptively determine the threshold search interval based on the edge response distribution of the current image; preferably, A value of 0.30~0.50 is acceptable. A value of 0.80 to 0.95 is acceptable.

[0040] In the interval Inside, according to the preset step size The set of non-zero edge responses is filtered based on the threshold search range to generate a candidate threshold set. Candidate threshold set Includes several candidate thresholds For any candidate threshold The corresponding binary edge map is defined as:

[0041] in, This is an indicator function. Indicates at the candidate threshold Next pixel coordinates The binary edge response at the point. The number of non-zero points is defined as:

[0042] in, Indicates at the candidate threshold Lower joint area The total number of edge pixels in the inner binary edge map. To avoid the contour being too sparse, causing breakage, or too dense, introducing noise, let the number of non-zero points of the target be:

[0043] in, This indicates the total number of pixels within the current joint region. The target edge density coefficient, preferably, A value of 0.03 to 0.08 is acceptable. Further, construct the threshold selection objective function:

[0044] Finally, in the candidate threshold set The threshold that minimizes the objective function is selected as the adaptive dynamic threshold.

[0045] in, This represents an adaptive dynamic threshold, which is used to generate the binary edge map required for subsequent contour extraction, and the binary edge map is used as the initial contour result.

[0046] Thus, this embodiment employs a multi-directional edge enhancement strategy to detect the knee joint boundary. Addressing the common issues of tilting, bending, and blurred local boundaries in knee joint contours observed in X-rays, this invention combines a customized convolutional kernel with a multi-directional Sobel operator to simultaneously calculate gradient responses in the horizontal, vertical, and diagonal directions. These gradient responses are then fused to form a more complete set of non-zero edge responses. By introducing multi-directional gradient information, the breakage or loss caused by insufficient response to tilted contours in a single-directional operator can be avoided, ensuring that key contours can be stably extracted under different poses.

[0047] In some implementations of this embodiment, a matching and fine-tuning mechanism based on predefined contours is introduced. For key areas of the knee joint, this embodiment pre-establishes a set of representative structural contour templates. The initial contour results extracted by adaptive dynamic thresholding are matched and filtered with the structural contour templates. The target contour that best matches the anatomical structure is located from the initial contour results, and local gaps, offsets, and burrs are corrected to obtain a stable contour result for pathological analysis. Specifically, the initial contour results are corrected through a predefined structural contour template matching strategy to obtain a stable contour result, including: Step S211: Set a predefined structural contour template set, wherein the predefined structural contour template set includes several predefined structural contour templates; Step S212: Obtain the set of candidate edge points corresponding to the initial contour result; Step S213: Set a search displacement set for each predefined structural contour template, the search displacement set including several search displacement vectors, and obtain a candidate contour template set after each predefined structural contour template is translated based on several search displacement vectors, the candidate contour template set including several candidate contour templates; Step S214: Based on the candidate edge coverage rate and the predefined structural contour template overlap rate, establish a matching function that weights and fuses the candidate edge point set and the candidate contour template to obtain the candidate contour template corresponding to the search displacement vector that makes the highest matching score; Step S215: Correct the initial contour result based on the candidate contour template to obtain a stable contour result.

[0048] In this embodiment, in step S211, the predefined structural contour template set is as follows: , Indicates the first A predefined structural outline template, The total number of predefined structural contour templates is defined as the set of typical bony boundary points obtained by experts after alignment, statistical analysis, and sampling for the corresponding anatomical location. In a specific embodiment, a total of 8 predefined structural contour templates can be set, each corresponding to a different predefined structural contour template.

[0049] In this embodiment, in step S212, the local anatomical sub-region to be matched is set. , The set of candidate edge points corresponding to the initial contour result of the current local anatomical sub-region to be matched is obtained as follows: The set of candidate edge points is Includes several candidate edge points ,in:

[0050] in, This indicates the local anatomical subregion currently to be matched. This represents the set of candidate edge points within the current local anatomical subregion to be matched. Represents the set of candidate edge points any point in, For adaptive dynamic threshold Next pixel coordinates Binary edge response at the location.

[0051] In this embodiment, in step S213, to adapt to individual anatomical differences and imaging position offsets, a search displacement set is set for each predefined structural contour template:

[0052] in, and These represent the local anatomical sub-regions currently to be matched. Width and height, and These represent the displacement search scaling factors in the horizontal and vertical directions, respectively. Indicates the displacement search step size. For the horizontal search displacement, This is the search displacement in the vertical direction. Preferably, A value of 0.10~0.15 is acceptable. A value of 0.08 to 0.12 is acceptable. It can be taken as 1 pixel. Under a set of typical joint dimensions, the above range can be corresponding to... , .

[0053] For any displacement vector , will the Candidate contour templates are obtained by translating a predefined structural contour template. .in Represents a predefined structural outline template In the displacement vector The translational predefined structural contour template obtained under the action.

[0054] The distance from a point to a set is defined as:

[0055] in, Represents a predefined structural contour template point set any point in, Represents the candidate edge point set any point in, Point Point set The minimum Euclidean distance; similarly, Point To the predefined structural contour template point set The minimum Euclidean distance. The candidate edge coverage rate is defined as:

[0056] in, Represents the candidate edge point set Falling into the predefined structural outline template Distance tolerance threshold The proportion of points within the template is used to measure the degree to which candidate edges cover the predefined structural contour template.

[0057] Define the overlap rate of the predefined structural contour template as:

[0058] in, Represents a predefined structural contour template point set Candidate edge point set At the distance tolerance threshold The proportion of points that hit the inner edge is used to measure the degree of overlap between the predefined structural contour template and the candidate edge.

[0059] The matching score is defined by combining the candidate edge coverage rate and the hit rate of the predefined structural contour template: in, These are weighting coefficients used to balance the contributions of the predefined structural contour template hit rate and the candidate edge coverage rate to the matching score; Indicates the translated first... Predefined structural outline template With candidate edge point set The overall matching score.

[0060] in, The distance tolerance threshold is used to determine whether the predefined structural contour template point and the candidate edge point are considered a match. Its value can be set according to the image resolution and the contour point sampling interval, preferably 2 to 3 pixels. A value of 0.5 is preferred. Finally, in the search displacement set... The search displacement vector that yields the highest matching score is selected from the internal selection: .

[0061] in, This represents the independent variable that maximizes the objective function. This represents the optimal translational displacement vector corresponding to the current predefined structural contour template.

[0062] The input and output relationship of the above multi-directional contour extraction and the contour segmentation diagram are as follows: Figure 3 As shown. Figure 3 middle, Figure 3 (a) shows a schematic diagram of the distribution of candidate contour segments obtained from multi-directional gradient responses; Figure 3 (b) represents the initial contour result obtained on the target X-ray image, corresponding to the binary edge map overlay effect in step S2 of the embodiment; Figure 3(c) is a schematic diagram of segmenting and marking the stable contour results according to the anatomical structure. It corresponds to the stable contour results after the predefined structural contour template in step S2 is refined, and is used to support the subsequent zonal quantitative analysis of pathological indicators such as joint space.

[0063] Step S3: Perform pathological quantitative analysis on the stable contour results to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes.

[0064] In some embodiments of this example, the stable contour results extracted from the knee joint X-ray image are further transformed into interpretable and computable pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. Based on these vectors, an adjacency matrix is ​​constructed for graph convolutional fusion reasoning, thereby achieving a structured representation of medical knowledge. Specifically: Step S301: For the stable contour results, two contour curves are formed to describe the joint gap. For each contour curve, the distance sequence between the two contours is calculated point by point along the vertical direction to obtain the distance list corresponding to the two contour curves respectively. Step S302: Obtain point pair sets based on the two sets of distance lists respectively, for subsequent overlap evaluation. This process yields two sets of distance lists and point pair sets, which serve as inputs for the quantization calculation in this section. Step S303: Based on two sets of distance lists and two sets of point pairs, extract mean features, variance features, range features, maximum distance features, and overlap features to characterize interpretable evidence of osteophyte, joint space changes, and left-right symmetry changes. Specifically: The first type is the mean feature. To reduce the interference of outliers and noise on the mean, this embodiment only calculates the mean for the distance values ​​in the smaller half of the distance list, so as to more stably reflect the typical width level of the joint space.

[0065] The second type is variance features, which calculate the degree of dispersion of the distance sequence to reflect the degree of fluctuation of the joint space at different positions, and indirectly characterize phenomena such as irregular joint surface morphology or local narrowing.

[0066] The third category is range features, which describe the span of the distance sequence. This feature reflects the overall variation of the joint space, providing a quantitative basis for assessing uneven stenosis.

[0067] The fourth category is the maximum distance feature, which is used to capture gap changes or contour deviations at local extreme locations, providing supplementary information for identifying local anomalies.

[0068] The fifth category is the overlap feature, which quantifies the degree of matching between left and right or different contours by comparing the proportion of overlap between two sets of point pairs in space, thus providing an interpretable indicator for assessing symmetry changes and structural consistency.

[0069] In this embodiment, the mean feature is mainly used to characterize the typical width level of the joint space; the variance feature, range feature, and maximum distance feature are mainly used to characterize the range of joint space changes, local extreme changes, and abnormal fluctuations caused by osteophytes or contour deviations; the overlap feature is mainly used to characterize changes in left-right symmetry, and the variance feature, range feature, and maximum distance feature can also be used as auxiliary indicators for characterizing left-right symmetry anomalies.

[0070] Specifically, let the vertical distance columns corresponding to the left and right joint space regions in a unified coordinate system be as follows: and ,in, Indicates the left joint space at the 1st Vertical distance values ​​at each sampling location Indicates the right joint space at the 1st Vertical distance values ​​at each sampling location. and These represent the number of sampling point sets in the left and right vertical distance columns, respectively. Let the sampling point sets corresponding to the left and right joint gaps be respectively... and , and Each includes several sampling locations, among which and These represent the sets of sampling location point coordinates used to construct the left and right vertical distance columns, respectively.

[0071] Define the merge distance column as follows:

[0072] in, This indicates a splicing operation. , Indicates the merged distance column The Middle Distance values.

[0073] make ,in, Indicates the merged distance column The 50th percentile, which is the median of the merged distance column. Further define the index set:

[0074] in, Indicates the merged distance column All of the above satisfy The set of indices corresponding to the elements. This indicates the number of elements in the set of indices. Then the mean characteristic... Defined as:

[0075] Variance characteristics Defined as:

[0076]

[0077] Range features Defined as:

[0078] in, Indicates the merged distance column The difference between the maximum and minimum values ​​is used to characterize the discrete range of the joint space distance distribution.

[0079] Maximum distance feature Defined as:

[0080] in, Indicates the merged distance column The maximum distance value is used to characterize the degree of gap variation at local extreme locations.

[0081] Overlap Features Defined as:

[0082] in, Represents the set of sampling points on the left side in a unified coordinate system. With the sampling location point set on the right The number of overlapping sampling points and These represent the total number of sampling locations in the left and right sampling location sets, respectively. It is used to measure the degree of overlap between the left and right joint spaces in space.

[0083] After calculating the above five types of features, this embodiment represents each image sample as a pathological feature vector. The feature vector consists of the mean, variance, range, maximum distance, and overlap. This feature vector serves as a numerical representation of medical knowledge, enabling the model to receive and process pathological concepts in a unified manner.

[0084] Step S4: Obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples, and construct a symmetric normalized adjacency matrix based on the similarity.

[0085] This embodiment constructs a symmetric normalized adjacency matrix based on the similarity of pathological feature vectors between samples to express the medical knowledge relationships between samples, specifically: Step S401: Combine the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples to obtain a feature matrix; Step S402: Based on Gaussian kernel similarity, convert the feature matrix into a similarity matrix; Step S403: Convert the similarity matrix into a sparse matrix using the K-nearest neighbor sparsification method; Step S404: Construct an adjacency matrix based on the sparse matrix, and add self-loops to the adjacency matrix to obtain the degree matrix; Step S405: Construct a symmetric normalized adjacency matrix based on the adjacency matrix and the degree matrix.

[0086] Specifically, based on the feature matrix, the similarity between any two pathological feature vectors is first calculated to obtain a similarity matrix. Let the first... The pathological feature vectors are:

[0087] in, For the first The mean characteristics of each pathological feature vector For the first Variance features of each pathological feature vector For the first Range characteristics of each pathological feature vector For the first The maximum distance feature of each pathological feature vector For the first Overlap feature of pathological feature vectors.

[0088] Similarity matrix The elements can be defined as Gaussian kernel similarity:

[0089] in, For the scale parameter, take The median of For the first Each pathological feature vector.

[0090] Furthermore, to achieve a sparse graph structure, only the Top-K similar neighbors of each sample node can be retained, while non-Top-K neighbors can be ignored. Set the similarity matrix to zero to convert it into a sparse matrix. .here The default value is 10, used to construct each sample node. Nearest neighbor connections.

[0091] Then, an adjacency matrix is ​​constructed based on the sparse matrix, and self-loops are added to this adjacency matrix to obtain a degree matrix that describes the total connection strength between each sample node and other sample nodes. Finally, the similarity matrix is ​​normalized to obtain a normalized adjacency matrix. Let the adjacency matrix... Adding a self-loop yields Degree matrix Defined as:

[0092] in, The adjacency matrix after adding self-loops In the Line 1 The element values ​​of the column, The adjacency matrix after adding self-loops In the graph, the total number of sample nodes participating in the graph construction. Then the symmetric normalized adjacency matrix is:

[0093] in, Used for feature propagation and aggregation in subsequent graph convolutions.

[0094] The normalized adjacency matrix, as input to the graph convolutional fusion diagnostic network model, can be directly used for feature propagation and aggregation in subsequent graph convolutional fusion diagnostic network models, enabling medical knowledge to play a structural guiding role in the model reasoning process.

[0095] Step S5: Input the target X-ray image and the symmetric normalized adjacency matrix into the pre-trained graph convolutional fusion diagnostic network model, and output the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is obtained by training a sample set and the corresponding symmetric normalized adjacency matrix set of the sample set. The sample set includes several sets of pre-trained X-ray images.

[0096] This embodiment proposes a graph convolutional fusion diagnostic network model driven by medical knowledge and data knowledge to classify osteoarthritis from target X-ray images of the knee joint and provide traceable medical evidence during the inference process. The network employs a dual-input structure, receiving the target X-ray image and a normalized adjacency matrix respectively, and completes feature propagation, fusion, and classification output within a unified framework.

[0097] In this embodiment, the graph convolutional fusion diagnostic network model includes a convolution and pooling unit, a graph convolutional fusion unit, and a fully connected output unit. The convolution and pooling unit is used to extract depth features from the target X-ray image to obtain the initial node features of each sample node. The graph convolutional fusion unit is used to perform feature propagation and aggregation on the initial node features and the normalized adjacency matrix to obtain node features. The fully connected output unit is used to output the classification results, which include the classification results of knee osteoarthritis and the corresponding classification probabilities.

[0098] Specifically, such as Figure 4 As shown, the first input to the convolutional fusion diagnostic network model is the data knowledge representing the target X-ray image, which is used to provide data knowledge representations such as image texture, bone morphology and overall structure.

[0099] The second input to the graph convolutional fusion diagnostic network model is medical knowledge represented by a normalized adjacency matrix, which represents the medical similarity relationships and constraints between the target X-ray image data and the sample or between structural parts. By introducing the normalized adjacency matrix, the network no longer relies solely on pixel-level texture during feature update, but explicitly incorporates structural relationships formed by pathological evidence such as osteophytes, joint space changes, and left-right symmetry changes into the inference path.

[0100] Secondly, the network uses a graph convolutional fusion unit as the core entry point for fusion. This unit propagates information between sample nodes based on a symmetric normalized adjacency matrix, aggregating and updating pathological feature vectors. This allows sample nodes with similar pathological feature vectors to form more stable structural relationships in the representation space. This propagation mechanism injects medical knowledge into the feature learning process in the form of structural constraints, thereby reducing the instability caused by relying solely on local textures and making the network output more consistent with clinical pathological patterns. The graph convolutional module can be configured as a multi-level structure, achieving more comprehensive relationship modeling through layer-by-layer propagation.

[0101] In one implementation, convolution and pooling units first extract depth features from the input target X-ray image to obtain the initial node features for each sample node. ; then With normalized adjacency matrix The graph is input into a convolutional fusion unit for feature propagation and aggregation to obtain node features after incorporating medical structural constraints. The hierarchical results are then output via a fully connected output unit.

[0102] Because the medical knowledge graph structure participates in the front-end propagation and aggregation process, the network can link the corresponding pathological quantitative evidence and its structural relationships while outputting results, making the diagnostic process interpretable and traceable.

[0103] The structure of the above-mentioned graph convolutional fusion diagnostic network model combining medical knowledge and data knowledge is as follows: Figure 4 As shown. Figure 4 In the network, input 1 represents the data knowledge corresponding to the target X-ray image, and input 2 represents the medical knowledge corresponding to the symmetric normalized adjacency matrix constructed by quantizing pathological feature vectors. The network includes convolution and pooling units, graph convolution fusion units, and fully connected output units in sequence, and finally outputs hierarchical results.

[0104] Specifically, the convolution and pooling unit extracts depth features from the target X-ray image through multiple sequentially connected CNN feature extraction sub-units. The multiple sequentially connected CNN feature extraction sub-units include a first convolutional layer (ConvBlock1), a second convolutional layer (ConvBlock2), a third convolutional layer (ConvBlock3), a fourth convolutional layer (ConvBlock4), and a global average pooling / flattening layer (GAP / Flatten). The first, second, and third convolutional layers are used to extract texture / edge detail features of the target X-ray image, the fourth convolutional layer is used for channel upscaling, and the global average pooling / flattening layer is used to obtain the initial node features.

[0105] The graph convolutional fusion unit performs feature propagation and aggregation on the initial node features and the normalized adjacency matrix through multiple sequentially connected GCN fusion sub-units. The GCN fusion sub-unit is a GCN (Graph Convolutional Network). In one specific embodiment... =2-layer GCN (Graph Convolutional Network): The first layer reduces the node feature dimension from Mapped to The second layer maps the dimension from 256 to... Each layer is followed by a ReLU activation function, and an application is made. To suppress overfitting. The forward propagation of GCN is defined as:

[0106] in, The node features output by the graph convolutional fusion unit. For the normalized adjacency matrix, For trainable weights, This is the activation function.

[0107] The fully connected output unit includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used to integrate global information and output the classification result through the ReLU activation function. The second fully connected layer is used to integrate global information and output the classification probability through the Softmax activation function.

[0108] Table 1 Examples of ExpGCN Network Parameters

[0109] The following is a detailed description of a knowledge graph fusion-based X-ray interpretable grading method for knee osteoarthritis provided in this embodiment, using a specific example. The statistics and distribution of the dataset used for training the graph convolutional fusion diagnostic network model are shown in Table 2, including the number of samples, the number of images from KL0 to KL4, and the division of the training / validation / test sets.

[0110] Table 2 Statistical Analysis of OAI Final Dataset

[0111] This patient dataset, through its age range, BMI distribution, and gender ratio, reflects to some extent the typical demographic characteristics of knee osteoarthritis, which helps reduce the risk of gender bias in the model.

[0112] Input: A target X-ray image of the knee joint The resolution is 2688×2208 pixels.

[0113] Preprocessed output: for After normalization and scaling to 512×512, we get .

[0114] Multidirectional gradient and threshold output: calculation In the corresponding image of this embodiment, the search range is obtained according to the aforementioned rules. Target non-zero number ,get ,at this time Output a binary edge map. .

[0115] Contour Output: Two key contours are obtained after template matching and contour refinement. , Their point counts are respectively , .

[0116] Quantized feature output: Calculate the vertical distance list The pathological quantitative characteristics were obtained: , , , , ,form .

[0117] Similarity / Adjacency Output: Similarity to two other samples in the sample set , Calculations yielded ,structure .

[0118] The graph convolutional fusion diagnostic network model outputs the KL classification probability. The prediction level is KL2, and the quantitative evidence corresponding to "decreased joint space width, abnormal maximum distance, and decreased symmetry" is also output (i.e. (equal value) Training implementation method: (1) Loss function: Weighted multi-class cross-entropy loss is adopted.

[0119] in, The number of hierarchical categories, One-hot encoding for real labels. Output the probability for Softmax. For class weights. A graph structure smoothing regularization term is added to enhance robustness:

[0120] in, Indicates the first The embedding representation of each sample node in the last layer of GCN Indicates the first The embedding representation of each sample node in the last layer of the GCN. The total loss is defined as:

[0121] in, The default value is 0.1, which is used as a weighting factor.

[0122] (2) Training strategy The initial learning rate is set to The weight decays to The batch size was set to 16, and the number of iterations was 200. The network used the Adam optimizer during training, with a specified learning rate and weight decay. All experiments were performed on an NVIDIA GeForce RTX 4090 GPU workstation equipped with 24GB of VRAM.

[0123] (3) Data augmentation Randomly rotate (e.g., ±10°), randomly translate (e.g., ±5%), randomly scale (e.g., 0.9–1.1), randomly perturb contrast / brightness, add Gaussian noise or slight blurring to the target X-ray image; keep the center alignment of the knee joint area during enhancement to avoid damaging the joint space structure.

[0124] Experimental results show that the ExpGCN network proposed in this embodiment, by integrating medical knowledge and data knowledge, significantly improves classification accuracy compared to classification schemes that only use convolutional neural networks (DatCNN) or rely solely on pathological quantitative features. By introducing the combination of graph convolution and medical knowledge, the overall accuracy of the model can reach approximately 69%.

[0125] As shown in Table 3, ExpGCN demonstrated excellent performance in precision, recall, F1 score, and accuracy across multiple experimental scenarios, particularly exhibiting strong predictive ability at higher KL levels. Compared to traditional models, ExpGCN achieved significant improvements across different KL levels, proving the effectiveness of combining graph convolutional networks with medical knowledge.

[0126] Table 3 Comparison of KOA Classification Experiment Results

[0127] Note: AttGCN is developed based on ExpGCN by integrating an attention mechanism.

[0128] According to the second aspect of this application, such as Figure 5 As shown in the figure, this embodiment provides a knowledge graph fusion-based X-ray interpretable grading method system for knee osteoarthritis, including: The first acquisition module is used to acquire target X-ray images of the knee joint; The second acquisition module is used to extract the contour of the target X-ray image in multiple directions to obtain the initial contour result of the knee joint, and to correct the initial contour result by a predefined structural contour template matching strategy to obtain a stable contour result. The third acquisition module is used to perform pathological quantitative analysis on the stable contour results to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. A construction module is used to obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vector corresponding to several samples, and to construct a symmetric normalized adjacency matrix based on the similarity. The fourth acquisition module inputs the target X-ray image and the symmetric normalized adjacency matrix into a pre-trained graph convolutional fusion diagnostic network model and outputs the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is obtained by training a sample set and the corresponding symmetric normalized adjacency matrix set of the sample set. The sample set includes several sets of pre-trained X-ray images.

[0129] Specifically, this embodiment corresponds one-to-one with the above method embodiments. The functions of each module have been described in detail in the corresponding method embodiments, so they will not be repeated here.

[0130] According to a third aspect of this application, this embodiment provides a computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0131] The present invention can implement all or part of the processes in the above methods, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0132] According to the fourth aspect of this application, such as Figure 6 As shown, an electronic device is provided, comprising: One or more processors; Memory is used to store executable instructions for the processor, which, when executed by one or more processors, cause one or more processors to implement the methods described above.

[0133] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different system components (including memory and processor).

[0134] The processor can be a Central Processing Unit (CPU), or 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. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of a computer system, connecting all parts of the computer system through various interfaces and lines.

[0135] Memory can be used to store computer programs and / or modules. The processor implements various functions of the computer system by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system and at least one application program required for a function (e.g., sound playback, image playback, etc.); the data storage area can store data created based on the use of the mobile phone (e.g., audio data, video data, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMedia Cards (SMC), Secure Digital (SD) cards, Flash Cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, servers, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and memory) containing computer-usable program code.

[0137] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), servers, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0141] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0142] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph fusion-based method for interpretable X-ray grading of knee osteoarthritis, characterized in that, include: Obtain target X-ray images of the knee joint; Multi-directional contour extraction is performed on the target X-ray image to obtain the initial contour result of the knee joint. The initial contour result is then corrected using a predefined structural contour template matching strategy to obtain a stable contour result. The stable contour results were subjected to pathological quantitative analysis to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. Obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples, and construct a symmetric normalized adjacency matrix based on the similarity. The target X-ray image and the symmetric normalized adjacency matrix are input into a pre-trained graph convolutional fusion diagnostic network model, which outputs the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is trained through a sample set, which includes several sets of pre-trained X-ray images and corresponding symmetric normalized adjacency matrices.

2. The method according to claim 1, characterized in that, The initial contour result of the knee joint is obtained by performing multi-directional contour extraction on the target X-ray image: Obtain a grayscale image of the target X-ray image; Based on the multi-directional Sobel operator, the gradient convolution kernels and corresponding gradient responses of the grayscale image in the horizontal, vertical and two diagonal directions are extracted respectively, and the gradient responses in all directions are fused to obtain the total edge response map. Based on the total edge response map, the set of non-zero edge responses of the joint region to be processed is obtained; The set of non-zero edge responses is filtered based on the threshold search range to generate a set of candidate thresholds. The number of non-zero points for each candidate threshold is obtained. An objective function is constructed based on the absolute value of the difference between the number of non-zero points of the candidate threshold and the number of non-zero points of the target threshold. The candidate threshold corresponding to the minimum objective function value is used as the adaptive dynamic threshold. The binary edge map extracted from the grayscale image based on the adaptive dynamic threshold is used as the initial contour result.

3. The method according to claim 1, characterized in that, The step of correcting the initial contour result using a predefined structural contour template matching strategy to obtain a stable contour result includes: Set a predefined structural contour template set, which includes several predefined structural contour templates; Obtain the set of candidate edge points corresponding to the initial contour result; For each predefined structural contour template, a search displacement set is set, the search displacement set includes several search displacement vectors, and a candidate contour template set is obtained after each predefined structural contour template is translated based on several search displacement vectors, the candidate contour template set includes several candidate contour templates; Based on the candidate edge coverage and the predefined structural contour template overlap rate, a matching function is established to weightedly fuse the candidate edge point set and the candidate contour template, and the candidate contour template corresponding to the search displacement vector that makes the highest matching score is obtained. The initial contour result is corrected based on the candidate contour template to obtain a stable contour result.

4. The method according to claim 1, characterized in that, The pathological quantitative analysis of the stable contour results yields pathological feature vectors based on osteophyte, joint space changes, and left-right symmetry changes, including: For the stable contour results, two contour curves are generated to describe the joint gap. For each contour curve, the distance sequence between the two contours is calculated point by point along the vertical direction to obtain the distance list corresponding to the two contour curves respectively. Based on the two sets of distance lists, obtain the point pair sets respectively; Based on two sets of distance lists and two sets of point pairs, mean features, variance features, range features, maximum distance features, and overlap features are extracted.

5. The method according to claim 1, characterized in that, The step of obtaining the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples, and constructing a symmetric normalized adjacency matrix based on the similarity, includes: The feature matrix is ​​obtained by combining the pathological feature vector corresponding to the target X-ray image and the pathological feature vectors corresponding to several samples; Based on Gaussian kernel similarity, the feature matrix is ​​converted into a similarity matrix; The similarity matrix is ​​converted into a sparse matrix using the K-nearest neighbor sparsification method. An adjacency matrix is ​​constructed based on a sparse matrix, and a self-loop is added to the adjacency matrix to obtain a degree matrix; Based on the adjacency matrix and the degree matrix, a symmetric normalized adjacency matrix is ​​constructed.

6. The method according to claim 1, characterized in that: The graph convolutional fusion diagnostic network model includes convolution and pooling units, a graph convolutional fusion unit, and a fully connected output unit. The convolution and pooling units are used to extract depth features from the target X-ray image to obtain the initial node features of each sample node. The graph convolutional fusion unit is used to perform feature propagation and aggregation on the initial node features and the normalized adjacency matrix to obtain node features. The fully connected output unit is used to output the classification results, which include the classification results of knee osteoarthritis and the corresponding classification probabilities.

7. The method according to claim 6, characterized in that: The convolution and pooling unit extracts depth features from the target X-ray image through multiple sequentially connected CNN feature extraction sub-units. The graph convolutional fusion unit performs feature propagation and aggregation on the initial node features and the normalized adjacency matrix through multiple sequentially connected GCN fusion sub-units, wherein the GCN fusion sub-unit is a graph convolutional network; The fully connected output unit includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used to output the classification result, and the second fully connected layer is used to output the classification probability.

8. A knowledge graph fusion-based X-ray interpretable grading system for knee osteoarthritis, characterized in that, include: The first acquisition module is used to acquire target X-ray images of the knee joint; The second acquisition module is used to extract the contour of the target X-ray image in multiple directions to obtain the initial contour result of the knee joint, and to correct the initial contour result by a predefined structural contour template matching strategy to obtain a stable contour result. The third acquisition module is used to perform pathological quantitative analysis on the stable contour results to obtain pathological feature vectors based on osteophytes, joint space changes, and left-right symmetry changes. A construction module is used to obtain the similarity between the pathological feature vector corresponding to the target X-ray image and the pathological feature vector corresponding to several samples, and to construct a symmetric normalized adjacency matrix based on the similarity. The fourth acquisition module inputs the target X-ray image and the symmetric normalized adjacency matrix into a pre-trained graph convolutional fusion diagnostic network model and outputs the grading result of knee osteoarthritis. The graph convolutional fusion diagnostic network model is obtained by training a sample set and the corresponding symmetric normalized adjacency matrix set of the sample set. The sample set includes several sets of pre-trained X-ray images.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program includes executable instructions that, when executed by a processor, implement the method of any one of claims 1-7.

10. An electronic device, characterized in that, include: One or more processors; A memory for storing executable instructions of the processor, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1-7.

Images