Deep learning based hip image classification system

Abstract

The present application relates to the technical field of medical image analysis and artificial intelligence aided diagnosis, in particular to a hip joint image classification system based on deep learning, comprising: acquiring two-dimensional image data and basic physical parameters of a target object; extracting two-dimensional key anatomical point features and pathological texture features, and generating a basic imaging classification probability; combining a generative adversarial network with a three-dimensional statistical shape model prior constraint to reconstruct a three-dimensional implicit geometric surface; discretizing into a three-dimensional finite element grid and combining with the basic physical parameters to deduce three-dimensional contact stress distribution characteristics; and fusing the basic imaging classification probability, pathological texture features and three-dimensional contact stress distribution characteristics to output a functional risk comprehensive evaluation result. The present application realizes comprehensive evaluation from static image classification to load-bearing function risk prediction.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and artificial intelligence-assisted diagnosis technology, specifically a hip joint image classification system based on deep learning. Background Technology

[0002] With the in-depth application of computer image processing technology, the complexity of automatic medical image classification tasks has increased significantly, especially in the comprehensive evaluation of multi-dimensional anatomical structures.

[0003] Currently, the basic static image classification probability is output by extracting features from two-dimensional medical images and using deep neural networks to identify anatomical points and pathological textures. However, traditional two-dimensional image classification methods are limited by the loss of anteroposterior depth information and local curvature changes inherent in two-dimensional projection, and rely solely on the feature representation of planar images for grade determination. Although these methods can provide basic lesion morphology classification results, they cannot extrapolate and generate three-dimensional structural entities and mechanical transmission characteristics under load, which can easily lead to a serious mismatch between the classification degree of two-dimensional image representation and the actual risk.

[0004] Therefore, how to overcome the dimensional limitations of two-dimensional static images to accurately achieve comprehensive classification and evaluation has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a hip joint image classification system based on deep learning, and to solve the following technical problems:

[0006] To avoid the discrepancy between the degree of degeneration and the patient's actual symptoms caused by relying solely on two-dimensional static imaging, the traditional static imaging classification is upgraded to a comprehensive assessment process oriented towards the weight-bearing functional status. This effectively predicts the risk of decompensation and improves the consistency between clinical decisions and the patient's true functional status.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A deep learning-based hip joint image classification system includes:

[0009] The data acquisition module is used to acquire two-dimensional image data and basic physical parameters of the target object;

[0010] The morphological feature extraction module is used to input two-dimensional image data into a pre-trained morphological feature extraction network, extract two-dimensional key anatomical point features and pathological texture features, and generate basic imaging classification probabilities.

[0011] The implicit geometry reconstruction module is used to input two-dimensional key anatomical point features and two-dimensional image data into a pre-trained generative adversarial network, and combine them with the prior constraints of a pre-constructed three-dimensional statistical shape model to inversely map and generate the three-dimensional implicit geometric surface of the target object.

[0012] The virtual biomechanical deduction module is used to discretize the three-dimensional implicit geometric surface into a three-dimensional finite element mesh. Combined with the basic physical parameters, it uses a pre-trained graph neural network to calculate the three-dimensional contact stress distribution characteristics of the target object under the load-bearing state corresponding to the basic physical parameters.

[0013] The functional risk joint classification module is used to fuse basic imaging classification probabilities, pathological texture features and three-dimensional contact stress distribution features into a multimodal feature. It takes a preset fully connected layer as input and outputs a comprehensive functional risk assessment result.

[0014] The model adaptive update module is used to obtain real clinical diagnostic labels, compare the functional risk comprehensive assessment results with real clinical diagnostic labels, obtain risk prediction bias, and adjust the weight parameters of morphological feature extraction network, generative adversarial network, graph neural network and fully connected layer based on risk prediction bias.

[0015] Optionally, the data acquisition module includes:

[0016] The image acquisition unit is used to receive two-dimensional image data sent by a preset acquisition device;

[0017] The two-dimensional image data includes the grayscale matrix of the region corresponding to the target object;

[0018] The parameter acquisition unit is used to acquire the basic physical parameters corresponding to the target object;

[0019] The basic physical parameters include preset gravity load parameters and geometric boundary condition parameters.

[0020] Optional, the morphological feature extraction module includes:

[0021] The feature encoding unit is used to perform feature dimensionality reduction on two-dimensional image data through a preset convolutional layer in the morphological feature extraction network to generate multi-scale feature maps.

[0022] The anatomical point localization unit is used to focus spatial features on the multi-scale feature map through a preset attention mechanism module and output two-dimensional key anatomical point features.

[0023] The texture extraction unit is used to perform high-frequency feature filtering on multi-scale feature maps and output pathological texture features.

[0024] The probability generation unit is used to input multi-scale feature maps into a preset classification head and output basic image classification probabilities.

[0025] Optional, implicit geometry reconstruction modules include:

[0026] The feature stitching unit is used to stitch together the features of two-dimensional key anatomical points with two-dimensional image data along the channel dimension to generate a fused feature vector.

[0027] The surface generation unit is used to input the fused feature vector into the preset generator in the generative adversarial network and output the initial 3D point cloud data.

[0028] Implicit mapping units are used to apply surface smoothing constraints to initial 3D point cloud data through prior constraints of a pre-built 3D statistical shape model, generating 3D implicit geometric surfaces.

[0029] Optionally, the virtual biomechanical simulation module includes:

[0030] Mesh generation units are used to discretize three-dimensional implicit geometric surfaces to generate three-dimensional finite element meshes;

[0031] Boundary condition application elements are used to map basic physical parameters onto multiple mesh nodes of a three-dimensional finite element mesh to construct a mechanical simulation model;

[0032] The stress calculation unit is used to transform the mechanical simulation model into a graph topology, where mesh nodes are used as graph nodes and the connection relationships of mesh elements are used as graph edges. The graph topology is input into a graph neural network for iterative solution and outputs three-dimensional contact stress distribution characteristics.

[0033] The three-dimensional contact stress distribution characteristics include the stress peak value and stress gradient vector of each node.

[0034] Optional, the functional risk joint classification module includes:

[0035] Modal alignment unit is used to map basic imaging classification probabilities, pathological texture features and three-dimensional contact stress distribution features to the same preset dimension space to generate an alignment feature matrix.

[0036] The feature fusion unit is used to calculate the association weights of each feature in the aligned feature matrix through a pre-defined cross-attention network to generate global fused features;

[0037] The classification output unit is used to input the global fusion features into the fully connected layer and output the comprehensive assessment result of functional risk.

[0038] Optionally, the classification output unit includes a preset risk assessment subunit, and the execution logic of the risk assessment subunit includes:

[0039] Extract the maximum stress peak value from the three-dimensional contact stress distribution characteristics;

[0040] If the maximum stress peak exceeds the preset stress failure threshold, a high-risk classification label is generated, and a comprehensive functional risk assessment result containing preset surgical intervention instructions is output.

[0041] If the maximum stress peak value is less than or equal to the stress failure threshold and greater than the preset conservative observation threshold, a medium-risk classification label is generated, and a functional risk comprehensive assessment result containing preset conservative treatment instructions is output.

[0042] If the maximum stress peak is less than or equal to the conservative observation threshold, a low-risk classification label is generated, and a functional risk comprehensive assessment result containing preset routine follow-up instructions is output.

[0043] Among them, the stress failure threshold and the conservative observation threshold are pre-set based on the bone mineral density parameters of the target object or the statistical distribution of historical clinical samples obtained in advance, and the stress failure threshold is greater than the conservative observation threshold.

[0044] Optional, the model adaptive update module includes:

[0045] The deviation calculation unit is used to compare the comprehensive assessment results of functional risks with the actual clinical diagnostic labels and calculate the risk prediction deviation.

[0046] The gradient backpropagation unit is used to calculate the gradient of the loss function based on the risk prediction bias.

[0047] The weight update unit is used to backpropagate the gradient of the loss function to the morphological feature extraction network, generative adversarial network, graph neural network and fully connected layer according to the preset learning rate, and update the weight parameters.

[0048] The beneficial effects of this invention are:

[0049] 1. This invention reverse maps two-dimensional images into three-dimensional implicit geometric surfaces and uses graph neural networks to deduce the three-dimensional contact stress distribution under load. This mechanism breaks through the limitation that traditional static images cannot reflect the real stress state, effectively identifies local stress concentrations, and realizes early prediction of functional failure risk.

[0050] 2. This invention integrates basic imaging probability, pathological texture features, and three-dimensional contact stress in a multimodal manner. This mechanism avoids the misleading effects of single morphological features, effectively combines the degree of structural degradation with the risk of mechanical functional failure, and forms a comprehensive risk profile that takes into account both the imaging appearance and the actual stress state, thereby improving the accuracy of auxiliary diagnosis.

[0051] 3. This invention introduces individualized basic physical parameters into the simulation model, extracts the maximum stress peak value and compares it with a dynamically set threshold, and directly outputs a comprehensive result including surgical, conservative or follow-up instructions; this avoids the assumption bias of the average model, transforms the algorithm deduction into a clinical treatment closed loop that can be directly executed, and enhances clinical applicability.

[0052] 4. This invention calculates the prediction deviation by comparing the risk assessment results with the real clinical diagnostic labels and jointly updates the weights of each network. This mechanism enables the system to absorb real clinical feedback and continuously evolve, effectively adapting to individual differences in bone morphology of different hospital equipment and populations, and ensuring the stability and generalization ability of the assessment results in the long term. Attached Figure Description

[0053] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0054] Figure 1 This is a schematic diagram of a deep learning-based hip joint image classification system provided in an embodiment of this application. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 A deep learning-based hip joint image classification system includes: a data acquisition module for acquiring two-dimensional image data and basic physical parameters of the target object;

[0057] The morphological feature extraction module is used to input two-dimensional image data into a pre-trained morphological feature extraction network, extract two-dimensional key anatomical point features and pathological texture features, and generate basic imaging classification probabilities.

[0058] The implicit geometry reconstruction module is used to input two-dimensional key anatomical point features and two-dimensional image data into a pre-trained generative adversarial network, and combine them with the prior constraints of a pre-constructed three-dimensional statistical shape model to inversely map and generate the three-dimensional implicit geometric surface of the target object.

[0059] The virtual biomechanical deduction module is used to discretize the three-dimensional implicit geometric surface into a three-dimensional finite element mesh. Combined with the basic physical parameters, it uses a pre-trained graph neural network to calculate the three-dimensional contact stress distribution characteristics of the target object under the load-bearing state corresponding to the basic physical parameters.

[0060] The functional risk joint classification module is used to fuse basic imaging classification probabilities, pathological texture features and three-dimensional contact stress distribution features into a multimodal feature. It takes a preset fully connected layer as input and outputs a comprehensive functional risk assessment result.

[0061] The model adaptive update module is used to obtain real clinical diagnostic labels, compare the functional risk comprehensive assessment results with real clinical diagnostic labels, obtain risk prediction bias, and adjust the weight parameters of morphological feature extraction network, generative adversarial network, graph neural network and fully connected layer based on risk prediction bias.

[0062] This embodiment provides an implementation mechanism for a hip joint image classification system based on deep learning. Specifically, the system is deployed in the hospital's orthopedic auxiliary diagnostic platform to serve the continuous clinical diagnosis and treatment process from initial outpatient diagnosis to preoperative assessment, treatment pathway suggestion, and follow-up result feedback.

[0063] The main scenario is as follows: A patient with long-term hip pain and decreased walking distance is received in the orthopedic clinic of a tertiary hospital. The doctor obtains anteroposterior images of the pelvis and the patient's basic weight-bearing information. The system not only provides traditional imaging classification, but also further infers the trend of joint stress concentration under actual weight-bearing conditions, and outputs functional risk conclusions for clinical decision-making based on this.

[0064] Specifically, the data acquisition module receives two-dimensional image data and the basic physical parameters corresponding to the patient; the two-dimensional image can be an anteroposterior X-ray of the pelvis or other single-plane medical images that meet the hospital imaging standards.

[0065] The basic physical parameters include at least gravity load information to describe the value of the load-bearing load and geometric boundary conditions to define the force posture; the basic physical parameters here are not abstract labels, but are used to characterize the actual force constraint relationship between the pelvis, proximal femur and joint contact surface when the patient is standing, bearing weight on one leg or initiating gait.

[0066] After acquiring the input, the morphological feature extraction module performs structured analysis on the two-dimensional image; the two-dimensional key anatomical point features extracted by this module are used to describe the femoral head contour, acetabular coverage edge, Shenton line continuity, and main boundaries of the joint space.

[0067] The extracted pathological texture features are used to reflect changes in osteophytes, cystic changes, sclerotic zones, and local bone density texture. At the same time, the system also generates basic imaging classification probabilities, such as the probability distribution of mild, moderate, and severe degeneration.

[0068] The probabilities mentioned above are used to characterize the two-dimensional static morphological state, providing preliminary information for subsequent three-dimensional functional deduction;

[0069] Because two-dimensional projection lacks anteroposterior depth, joint coverage spatial relationship, and local curvature changes in terms of physical mechanism, relying solely on two-dimensional classification results is prone to discrepancies between the degree of degradation represented by two-dimensional static images and the patient's actual complaints; therefore, the system further introduces an implicit geometric reconstruction module.

[0070] This module inputs two-dimensional key anatomical points and original images into a generative adversarial network. The network first generates an initial three-dimensional morphology of the patient's hip joint, and then uses a three-dimensional statistical shape model for prior constraints, so that the generated results retain individual patient differences and do not deviate from the acceptable range of real anatomy.

[0071] The implicit geometric surface here can be understood as the reconstruction of the continuous boundary of the femoral head, acetabulum and main contact area. Its significance lies in transforming the two-dimensional image representation into a three-dimensional structural entity that can bear mechanical analysis.

[0072] After obtaining the three-dimensional geometric surface, the virtual biomechanics deduction module discretizes it into a three-dimensional finite element mesh and maps the load-bearing state by combining the patient's weight, posture constraints and pelvic stability relationship.

[0073] Furthermore, graph neural networks do not simply replace conventional finite element solvers, but rather fit the mechanical transmission law of local morphological changes—neighborhood force transmission—global stress redistribution;

[0074] For example, when the anterolateral curvature of the femoral head becomes abnormally steep and the acetabulum coverage is insufficient, the load is more likely to form a local peak in this area, and the stress gradient has a higher rate of spatial variation between neighboring nodes.

[0075] Graph neural networks utilize the connections between grid nodes to quickly deduce this stress chain; the resulting three-dimensional contact stress distribution features include at least the stress peak value and stress gradient vector of each node, used to describe the contact stress peak region and the magnitude of the contact stress gradient variation.

[0076] The functional risk joint classification module integrates the aforementioned two-dimensional imaging classification probabilities, pathological texture features, and three-dimensional contact stress distribution features into a multimodal fusion.

[0077] Its core significance lies in avoiding the misleading effect of a single source of information; for example, both patients may present with narrowing of the joint space, but one patient has a more uniform distribution of stress peaks, while the other patient shows a significant concentration of stress extremes in the anterior and superior contact area.

[0078] The former may be more suitable for conservative observation, while the latter suggests a higher risk of load-bearing decompensation; after fusion, the comprehensive results output by the fully connected layer no longer remain at the lesion grade label, but form a clinical auxiliary conclusion that takes into account both the degree of structural degeneration and the risk of functional failure.

[0079] The model adaptive update module compares the system output with real clinical diagnostic labels, surgical records, follow-up pain scores or subsequent treatment outcomes, and corrects the system parameters accordingly.

[0080] Its purpose is not to make the system overfit to a particular case, but to gradually correct the mapping relationship between two-dimensional images, three-dimensional structures, stress risks, and clinical outcomes, so as to adapt to the real feedback under different hospital equipment, different population bone morphology differences, and different treatment pathways.

[0081] As an anomaly handling mechanism, if the image has obvious artifacts, excessive pelvic rotation, or abnormal exposure, causing the key anatomical boundaries to be unstable and unrecognizable, the system will pause 3D reconstruction and mechanical deduction, only output a prompt that the image quality is insufficient, and suggest re-acquiring the image.

[0082] If the basic physical parameters are incomplete, such as lack of load-bearing status or height and weight information, the system can use the standardized adult reference boundary within the hospital for transient extrapolation, but it will be marked as low confidence in the output.

[0083] If the reconstructed three-dimensional surface deviates too much from the statistical shape prior, it indicates that the input image may be abnormal or there may be postoperative implant interference. In this case, the system will prioritize maintaining the two-dimensional classification result and will not directly give a high-confidence mechanical conclusion to avoid misleading the clinic.

[0084] In the anteroposterior pelvic X-ray of the outpatient, the system identified mild flattening of the femoral head edge on the affected side, insufficient coverage of the upper outer edge of the acetabulum, and detected bone sclerosis texture in the area above the joint space.

[0085] Two-dimensional classification results indicate that the degeneration probability is mainly moderate to severe; three-dimensional reconstruction results show a sudden change in local curvature of the anterolateral femoral head, and contact stress concentration occurs under the single-leg standing boundary during load-bearing simulation.

[0086] After multimodal fusion, the system output showed a high functional risk and an increased probability of conservative treatment failure. It is recommended to further evaluate the surgical indications in conjunction with clinical symptoms and to include this case in the follow-up database.

[0087] The purpose of this step is to upgrade the traditional static image classification into a comprehensive assessment process oriented towards load-bearing functional status, thereby realizing the technological transformation from seeing lesions to predicting the risk of decompensation, and improving the consistency between clinical decisions and the patient's actual functional status.

[0088] In a preferred embodiment of the present invention, the data acquisition module includes: an image acquisition unit for receiving two-dimensional image data sent by a preset acquisition device; wherein the two-dimensional image data includes a grayscale matrix of the region corresponding to the target object; and a parameter acquisition unit for acquiring basic physical parameters corresponding to the target object; wherein the basic physical parameters include preset gravity load parameters and geometric boundary condition parameters.

[0089] This embodiment provides a refined implementation mechanism for a data acquisition module. Specifically, in the aforementioned outpatient scenario, if only images are acquired without simultaneously acquiring basic physical parameters related to load-bearing capacity, the subsequent three-dimensional stress deduction is easily reduced to an average assumption under a standard body type, which cannot accurately reflect the patient's true mechanical state. Therefore, this embodiment designs image acquisition and parameter acquisition in a coordinated manner.

[0090] Specifically, the image acquisition unit is connected to the hospital's medical image archiving and communication system or digital imaging equipment to receive two-dimensional image data of the patient's hip joint area; the received grayscale matrix can be understood as the arrangement of pixel brightness and darkness, and the changes in brightness and darkness correspond to the differences in projection of cortical bone, cancellous bone, joint space and soft tissue.

[0091] For this system, the grayscale matrix is ​​not only an image file, but also the basic input for subsequent anatomical boundary localization, texture anomaly recognition, and 2D to 3D inverse mapping. Therefore, during the acquisition stage, it is preferable to ensure that the pelvis is symmetrically positioned and the lower limb rotation posture meets the hospital's standards, so as to reduce anatomical distortion caused by projection angle deviation from the source.

[0092] The parameter acquisition unit is used to synchronously receive the patient's basic physical parameters, including gravity load parameters which can be derived from the patient's weight, the weight-bearing ratio on the affected side, or the standardized standing load model in the hospital; geometric boundary condition parameters can include pelvic posture, relative orientation of the proximal femur, foot support pattern, and standard standing or single-leg standing scene labels.

[0093] The above parameters are important because hip joint stress depends not only on the bone morphology itself, but also on the magnitude of the body load and the force transmission path; with the same degree of joint degeneration, local contact pressure is more likely to increase when the body weight is higher, the pelvic compensatory tilt is obvious, or the gait is unbalanced.

[0094] To clarify the data flow process, a standardized data mapping structure is used for expression. Assuming that the image acquisition unit receives a grayscale matrix G of the affected hip joint area, the parameter acquisition unit simultaneously obtains a set of parameters P, where P includes three fields: load-bearing scenario = bipedal standing load level = medium to high pelvic tilt = mild.

[0095] The subsequent modules do not directly treat P as a text label, but instead transform it into the meaning of force boundary. For example, standing on both feet corresponds to the force distribution of the left and right hips, medium to high load corresponds to the overall compressive stress of the contact surface shifting upward, and slight pelvic tilt corresponds to the force center shifting to one side.

[0096] As a system fault-tolerant control mechanism, if the image acquisition unit detects that the image lacks pelvic reference landmarks, has a truncated field of view, or has a metal artifact obscuring a key area, it will automatically mark the image as low availability and restrict it from entering the reconstruction process; if the parameter acquisition unit lacks weight data, it can call the most recent valid weight record in the electronic medical record.

[0097] If boundary condition parameters cannot be obtained, such as when the patient cannot complete the standard standing position, the system allows switching to the default supine reference model, but at the same time, it will indicate in the report that the result is biased towards structural assessment and the functional extrapolation output is conservative.

[0098] After the patient enters the X-ray room, the digital imaging equipment generates a grayscale image of the pelvis in the frontal position and automatically transmits it back to the platform. At the same time, the system reads the patient's height, weight and the information entered by the doctor that the side of the pain is the affected side when standing, and converts it into force modeling parameters. Since the patient's body position is basically symmetrical, this set of data is determined to be able to directly enter the subsequent analysis.

[0099] The purpose of this step is to ensure that the input data simultaneously covers structural image information and load-bearing condition information, so that subsequent deductions are based on the patient's individualized mechanical background, rather than a static average model detached from the clinical setting.

[0100] In a preferred embodiment of the present invention, the morphological feature extraction module includes: a feature encoding unit, used to perform feature dimensionality reduction on two-dimensional image data through a preset convolutional layer in the morphological feature extraction network to generate a multi-scale feature map;

[0101] The anatomical point localization unit is used to focus spatial features on the multi-scale feature map through a preset attention mechanism module and output two-dimensional key anatomical point features.

[0102] The texture extraction unit is used to perform high-frequency feature filtering on multi-scale feature maps and output pathological texture features; the probability generation unit is used to input multi-scale feature maps into a preset classification head and output basic imaging classification probabilities.

[0103] This embodiment provides a refined implementation mechanism for the morphological feature extraction module. Specifically, based on the aforementioned acquisition process, if only the entire image is roughly classified, the system is easily interfered with by irrelevant structures around the pelvis, soft tissue shadows, or differences in shooting posture, resulting in a reduction in the feature extraction weight of the real lesion area.

[0104] Therefore, this embodiment refines the two-dimensional morphological analysis into a continuous process of multi-scale encoding, anatomical point focusing, pathological texture extraction, and classification probability generation;

[0105] Specifically, the feature coding unit performs multi-layer convolutional coding on the two-dimensional image to form a multi-scale feature map. The so-called multi-scale means that coarse structural information and fine local information are preserved at the same time: the former helps to identify the overall spatial relationship between the femoral head and the acetabulum, while the latter helps to identify serrated changes at the joint space edge, osteophyte protrusions, and local bone sclerosis texture.

[0106] In hip joint scenarios, looking at the coarse scale alone can easily overlook small osteophytes, while looking at the fine scale alone may misjudge local noise. Therefore, multi-scale combination is necessary.

[0107] After obtaining the multi-scale feature map, the anatomical point localization unit focuses on the spatial region through an attention mechanism; its technical meaning is that the system assigns higher attention weight to anatomically significant regions, such as the boundary arc of the femoral head spherical contour, the coverage point of the upper edge of the acetabulum, the direction of the femoral neck axis, and pelvic reference landmarks near the teardrop area.

[0108] For ease of explanation, a certain layer of feature map can be simplified into a preset grid partition, where the upper and middle grids correspond to the upper outer edge of the acetabulum, the middle grid corresponds to the center of the femoral head, and the lower inner grid corresponds to the edge of the obturator foramen. The attention mechanism will enhance the grid response directly related to the joint configuration and suppress soft tissue projection interference that is irrelevant to diagnosis. The output can be represented as the coordinates of several key points and their neighborhood description information.

[0109] The texture extraction unit further performs high-frequency feature filtering on the multi-scale feature map to extract pathological texture features. Here, high frequency is not a simple mathematical concept, but refers to areas where the gray-level gradient exceeds a preset threshold, the edge contrast is greater than a specific parameter, or the surface roughness parameter exceeds the standard range. In hip joint images, these often correspond to the edges of osteophytes, the cortical sclerosis zone, the periphery of cystic changes, or irregular contact interfaces after wear.

[0110] The function of this unit is to transform local texture changes with significant gray-level anomalies into structured features that can be used for classification and subsequent fusion; the probability generation unit inputs multi-scale feature maps into the classification head to form basic image classification probabilities;

[0111] As an example: Suppose the classification head outputs three channels, representing the possibility of mild, moderate and severe degeneration respectively. When a patient has a high response in both the moderate and severe channels, it indicates that their two-dimensional morphology is near the boundary of the traditional classification. Further judgment should rely on three-dimensional mechanical results, rather than classification based on a single fixed standard.

[0112] As an anomaly handling mechanism, if the anatomical point localization results show that the spatial relationship between key points obviously violates basic anatomical common sense, such as the femoral head center falling outside the acetabular contour, it indicates that the image may have serious rotation, truncation or algorithm interference, and the system will trigger relocalization or fall back to coarse classification mode.

[0113] If the texture extraction unit detects that the high-frequency information mainly comes from metal artifacts or edge enhancement noise, the features in this part are downweighted to avoid mistaking device artifacts for pathological changes; if the probability of each category output by the classification head is close, the system marks the case as two-dimensional discrimination uncertain, reserving more decision space for subsequent multimodal fusion.

[0114] In the images of this outpatient, the coding unit identified abnormalities in the coverage relationship between the femoral head and acetabulum on the affected side at a coarse scale, and identified sclerosis and small osteophytes at the superior joint margin at a fine scale.

[0115] After anatomical point localization, the system stably determines the center of the femoral head, the upper outer edge of the acetabulum, and the upper edge of the joint space; texture extraction results show that the texture abnormalities are more obvious in the upper weight-bearing area; the classification head gives the basic probability output of moderate to severe degeneration tendency based on this;

[0116] The purpose of this step is to transform the original two-dimensional grayscale image into structured features that have both anatomical location and pathological texture meanings, so as to provide a stable and interpretable front-end input for subsequent three-dimensional reconstruction and functional risk assessment.

[0117] In a preferred embodiment of the present invention, the implicit geometric reconstruction module includes: a feature stitching unit, used to stitch two-dimensional key anatomical point features with two-dimensional image data along the channel dimension to generate a fused feature vector; a surface generation unit, used to input the fused feature vector into a preset generator in a generative adversarial network to output initial three-dimensional point cloud data; and an implicit mapping unit, used to apply surface smoothing constraints to the initial three-dimensional point cloud data through a pre-constructed three-dimensional statistical shape model prior constraint to generate a three-dimensional implicit geometric surface.

[0118] This embodiment provides a refined implementation mechanism for an implicit geometric reconstruction module. Specifically, while two-dimensional key points and classification probabilities can describe where anomalies occur, they cannot directly obtain the three-dimensional boundary of how the stressed surfaces contact each other. If the three-dimensional shape is directly extrapolated from the two-dimensional projection, distortion in the depth direction is likely to occur. Therefore, this embodiment constructs a three-dimensional implicit geometric surface that can be used for mechanical analysis through feature stitching, surface generation, and statistical prior constraints.

[0119] Specifically, the feature stitching unit fuses the features of two-dimensional key anatomical points with the original two-dimensional image in the channel dimension; instead of a simple linear stitching of non-feature data, it inputs explicit positioning information and original grayscale morphological information into the subsequent generator; for example, key points can clearly define the approximate location of the femoral head center, acetabular edge and joint contact zone, while the original image retains edge continuity, cortical thickness changes and local contour details.

[0120] When the two are combined, the system no longer makes structural inferences based solely on a single grayscale texture, nor does it rigidly fit surfaces based on a small number of points.

[0121] The surface generation unit uses a generative adversarial network to output initial 3D point cloud data; its significance is illustrated through a specific deduction example: assuming that the 2D input indicates that the femoral head is approximately circular in the orthogonal projection, but the outer upper edge is slightly flattened, the point cloud output by the generator will maintain the overall spherical trend, while showing local collapse or curvature changes in the corresponding spatial region.

[0122] The initial point cloud represents a discrete set of spatial coordinates representing the individualized bony structure of the patient, serving as a basic three-dimensional topological skeleton to support subsequent surface continuity processing.

[0123] The implicit mapping unit further constrains the surface smoothing through a three-dimensional statistical shape model. The scientific basis for this step is that although there are individual differences in the actual anatomy of the hip joint, it is still constrained by the biological morphological boundary in general, and there will be no abnormal curvature abrupt change, breakage or discontinuous overturning surface.

[0124] The statistical shape model gathers a large sample of common morphological changes in the pelvis and proximal femur, thus allowing for reasonable constraints on the initial point cloud. This ensures that the model preserves the patient's local degenerative features while preventing the generator from producing non-physical artifacts in the direction of missing information. The resulting implicit geometric surface can continuously describe the spatial relationship between the femoral head surface, the acetabular bearing surface, and the area where the two are close together.

[0125] If the two-dimensional keypoint localization is unstable, leading to spatial relationship conflicts between different keypoints, the feature stitching unit can prioritize the reconstruction path dominated by the original image and reduce the influence of keypoint branches.

[0126] If the initial point cloud output by the generator has obvious discrete isolated points or local overlaps, then point cloud cleaning and topology repair are performed first, and then statistical prior constraints are fed in.

[0127] If the statistical shape model determines that the initial result exceeds the acceptable morphological distribution range, such as in cases of suspected postoperative prosthesis or severe bone defects, the system can mark the reconstruction as an abnormal anatomical structure and switch to conservative mode or prompt manual review during subsequent simulations.

[0128] In the outpatient's case, the two-dimensional key points showed insufficient coverage of the upper lateral margin of the acetabulum on the affected side, while the original image also showed a slightly flattened outline of the weight-bearing area of ​​the femoral head; after stitching, the image was input into the generator to obtain a set of initial point clouds showing local morphological abnormalities in the anterolateral aspect of the femoral head.

[0129] Constrained by a statistical shape model, the system generates a continuous three-dimensional implicit surface that conforms to the anatomical rules of the hip joint, providing a basis for the next stage of load-bearing simulation.

[0130] The purpose of this step is to restore the two-dimensional static projection into a three-dimensional structural expression with anatomical continuity and individual specificity, so that subsequent mechanical analysis can be carried out on the actual contact interface rather than the two-dimensional image representation.

[0131] In a preferred embodiment of the present invention, the virtual biomechanics derivation module includes: a mesh generation unit, used to discretize a three-dimensional implicit geometric surface to generate a three-dimensional finite element mesh;

[0132] Boundary condition application elements are used to map basic physical parameters onto multiple mesh nodes of a three-dimensional finite element mesh to construct a mechanical simulation model;

[0133] The stress calculation unit is used to transform the mechanical simulation model into a graph topology, where grid nodes are used as graph nodes and the connection relationships of grid cells are used as graph edges. The graph topology is input into a graph neural network for iterative solution, and the output is three-dimensional contact stress distribution characteristics. The three-dimensional contact stress distribution characteristics include the stress peak value and stress gradient vector of each node.

[0134] This embodiment provides a refined implementation mechanism for a virtual biomechanical deduction module; specifically, although the three-dimensional implicit geometric surface has been obtained in the previous stage, if it remains at the level of three-dimensional shape display, it can only provide a three-dimensional geometric shape description and cannot output the mechanical instability position under load.

[0135] Therefore, this embodiment transforms the three-dimensional structure into an interpretable functional stress result through mesh generation, boundary condition application, and stress calculation;

[0136] Specifically, the mesh generation unit discretizes the three-dimensional implicit geometric surface into a three-dimensional finite element mesh; the essence of this process is to divide the continuous bone surface into multiple interconnected small units, enabling the system to track the transmission path of external loads between local regions.

[0137] For the hip joint, the mesh density of the weight-bearing zone, the edge zone, and the relatively non-weight-bearing zone can be different. Usually, the mesh density near the contact zone needs to be more refined because this area is more prone to stress peaks and gradient abrupt changes. This is not to increase computational complexity, but to more accurately describe local stress concentration.

[0138] Boundary condition application elements map basic physical parameters onto mesh nodes to form a mechanical simulation model. The mapping here means: distributing the axial load corresponding to body weight to the force direction of the pelvis and femur, transforming the standing or single-leg support scenario into the configuration of constrained nodes and free nodes, and introducing postural effects such as pelvic tilt or femoral adduction and abduction. For example, with the same structural morphology, the stress center of the contact surface is often significantly different in the two cases of balanced bipedal standing and single-leg weight-bearing on the affected side.

[0139] The stress calculation unit further transforms the mechanical simulation model into a graph topology and inputs it into a graph neural network for iterative solution. Here, the mesh nodes are used as graph nodes and the element connection relationships are used as graph edges, which can express how the deformation and stress changes in a certain region spread along the adjacent bone surfaces.

[0140] Before inputting the topology into the graph neural network, the system constructs an initial feature vector containing the semantics of the mechanical boundaries for each graph node;

[0141] Specifically, the feature vector not only includes the node's three-dimensional spatial coordinates and local patch normal vectors, but also splices together the external load components mapped to the node and the boundary constraint state variables;

[0142] The graph neural network receives a feature matrix composed of the features of the nodes mentioned above and an adjacency matrix composed of grid edges, and updates the features through a multi-layer message passing mechanism.

[0143] Specifically, in each iteration, the node feature update follows the following structured business rules: In the graph topology, the set of nodes directly connected to the current node through a single grid edge is predefined as the first-order neighbor node set; the features of the current node's first-order neighbor node set are extracted.

[0144] The propagation weights are assigned based on the reciprocal of the Euclidean distance between the neighboring nodes and the current node in three-dimensional space to reflect the physical property that the mechanical effect gradually decreases with spatial distance; the features of all neighboring nodes are weighted and summed according to the assigned propagation weights to obtain the neighborhood aggregation features.

[0145] The neighborhood aggregated features are concatenated with the current node's own features and input into a pre-defined multilayer perceptron network to generate a new feature vector for the next layer.

[0146] For ease of explanation, it can be simplified to a contact chain consisting of nodes N1, N2, and N3: when the local contact surface corresponding to N2 is narrow and the force path is restricted, the network calculates that its neighborhood features shrink sharply through feature aggregation during iteration, and infers that its stress level will be higher than that of the nodes on both sides.

[0147] The graph neural network fits this mechanical transmission law of geometric constraint-neighborhood load compression-local peak increase; the final output three-dimensional contact stress distribution features include at least node-level stress peaks and stress gradient vectors, where the former reflects the high pressure area and the latter reflects whether the pressure change is steep.

[0148] Both are equally important for assessing the risk of cartilage wear and bone surface micro-damage, because it is not only high absolute pressure that is dangerous; a sudden change in pressure can also indicate insufficient local tolerance.

[0149] As a fault-tolerant control mechanism, if malformed cells or local topological breaks occur after mesh generation, the system first performs mesh smoothing and reconstruction to avoid treating geometric noise as anomalies in stress. If the boundary condition parameters are insufficient, the system prioritizes the standard bipedal standing condition as the baseline scenario and reduces the extrapolation intensity of dynamic risks.

[0150] If the stress distribution output by the graph neural network clearly conflicts with basic mechanics, such as the highest stress falling in a clearly non-contact area, the result verification process is triggered to check whether the reconstructed geometry and boundary mapping are abnormal, and if necessary, revert to the simplified finite element reference result.

[0151] On the three-dimensional hip joint surface of the outpatient, the system divides the weight-bearing area on the affected side into a relatively dense mesh and applies load constraints corresponding to the patient's weight and standing posture.

[0152] The graph neural network solution shows that a high stress concentration zone is formed between the anterolateral aspect of the femoral head and the superior lateral edge of the acetabulum, and the stress gradient increases from the center to the periphery, indicating that this area is more likely to experience continuous mechanical stimulation during daily weight-bearing.

[0153] The purpose of this step is to make the force imbalance state that cannot be directly expressed by two-dimensional images explicit, thereby enabling early identification of the risk of clinical misjudgment and providing functional evidence support for the selection of treatment pathways.

[0154] In a preferred embodiment of the present invention, the functional risk joint classification module includes: a modal alignment unit, used to map the basic imaging classification probability, pathological texture features and three-dimensional contact stress distribution features to the same preset dimensional space to generate an alignment feature matrix;

[0155] The feature fusion unit is used to calculate the association weights of each feature in the aligned feature matrix through a pre-set cross-attention network to generate global fused features; the classification output unit is used to input the global fused features into the fully connected layer and output the comprehensive functional risk assessment result.

[0156] This embodiment provides a refined implementation mechanism for a functional risk joint classification module. Specifically, in the aforementioned process, the system has obtained two-dimensional classification probability, pathological texture features, and three-dimensional stress distribution features respectively. However, if the three are output separately, doctors still need to make a comprehensive judgment on their own, which can easily reduce usability due to the dispersed information sources.

[0157] Therefore, this embodiment introduces modal alignment, feature fusion and classification output mechanisms to unify structural degradation evidence and functional stress evidence into a single risk conclusion;

[0158] Specifically, the modal alignment unit first maps features from different sources to the same preset dimensional space. The necessity of this is that two-dimensional classification probability usually reflects category tendency, pathological texture features reflect the local microscopic lesion morphology, and three-dimensional stress distribution features reflect the mechanical response under load. The three have different semantic scales, and direct splicing can easily cause one type of information to dominate the final result.

[0159] After alignment, different modalities can be represented as multiple row vectors or column vectors in the same feature matrix; for example, the stress concentration on the anterolateral side of the weight-bearing area with obvious sclerosis in the weight-bearing area of ​​moderate to severe degeneration can be represented as aligned feature fragments F1, F2, and F3, respectively, so that the subsequent network can learn the clinical coupling relationship between them.

[0160] The feature fusion unit uses a cross-attention network to calculate the correlation weights between the above features; the core here is not the abstract weight calculation process, but the identification of which structural anomalies truly correspond to functional decompensation.

[0161] For example, if pathological texture suggests hardening above, but the stress peak is mainly located on the anterolateral side, it indicates that the two pieces of evidence may reflect the same long-term load shift process, and their correlation should be increased; conversely, if the two-dimensional classification emphasizes severe degeneration, but the stress distribution is relatively dispersed, it indicates that morphological deterioration has not necessarily transformed into high-risk functional instability, and the structural evidence should not be overemphasized.

[0162] The classification output unit inputs the fused global features into the fully connected layer to form a comprehensive evaluation result; this result may include risk level, main risk source region, trend judgment of response to conservative treatment, etc.

[0163] In actual deployment, the output results are preferably accompanied by a prompt indicating whether the risk is structure-dominant or mechanically dominant, so that doctors can understand whether the conclusion is more inclined to severe image degradation or more inclined to obvious weight-bearing imbalance.

[0164] As a system fault-tolerant control mechanism, if a certain mode is missing, for example, if stress characteristics are unavailable due to reconstruction failure, the mode alignment unit can align only the existing modes and reduce the certainty level of the risk conclusion in the output.

[0165] If there is a significant conflict between different modalities, such as a two-dimensional indication of severe but a mechanical result close to low risk, the feature fusion unit will not force an extreme judgment, but will instead add a suggestion to combine clinical symptoms for review; if the output of the fully connected layer is close to the boundary of multiple categories, the system can retain the risk suggestions of two adjacent categories for clinical manual confirmation.

[0166] In this outpatient, the two-dimensional classification probability indicated that the degeneration was close to severe, the pathological texture features showed sclerosis in the upper weight-bearing area and marginal osteophytes, and the three-dimensional stress features showed local peak concentration on the anterolateral contact surface.

[0167] After alignment, the cross-attention network calculates a higher association weight between inadequate coverage, increased anterolateral stress, and tendency to fail conservative treatment, ultimately outputting a comprehensive assessment result with a higher risk.

[0168] The purpose of this step is to transform previously fragmented, multi-source evidence into unified, interpretable clinical support judgments, so that the system output is no longer a single indicator, but a comprehensive risk profile oriented towards actual treatment decisions.

[0169] In a preferred embodiment of the present invention, the classification output unit includes a preset risk determination subunit. The execution logic of the risk determination subunit includes: extracting the maximum stress peak value in the three-dimensional contact stress distribution features; if the maximum stress peak value is greater than a preset stress failure threshold, generating a high-risk classification label and outputting a functional risk comprehensive assessment result containing preset surgical intervention instructions.

[0170] If the maximum stress peak value is less than or equal to the stress failure threshold and greater than the preset conservative observation threshold, a medium-risk classification label is generated, and a functional risk comprehensive assessment result containing preset conservative treatment instructions is output.

[0171] If the maximum stress peak is less than or equal to the conservative observation threshold, a low-risk classification label is generated, and a functional risk comprehensive assessment result containing preset routine follow-up instructions is output.

[0172] Among them, the stress failure threshold and the conservative observation threshold are pre-set based on the bone mineral density parameters of the target object or the statistical distribution of historical clinical samples obtained in advance, and the stress failure threshold is greater than the conservative observation threshold.

[0173] This embodiment provides a refined implementation mechanism for a risk assessment subunit; specifically, relying solely on comprehensive features to output an abstract risk score, while beneficial for the algorithm to express complex relationships, may still result in assessment results lacking direct and actionable clinical guidance in actual clinical applications.

[0174] Therefore, this embodiment further introduces risk stratification logic based on the maximum stress peak value, so that the system output can be directly mapped into executable diagnosis and treatment suggestions;

[0175] Specifically, the risk assessment subunit extracts the maximum stress peak value from the three-dimensional contact stress distribution characteristics. The reason for choosing this indicator as the key assessment criterion is that during the long-term weight-bearing process of the hip joint, an excessively high local peak value often means a reduction in the contact area, a concentration of load, or that the tolerance boundary of the osteocartilage interface is being approached. Such situations are more likely to lead to persistent pain, accelerated wear, or poor results from conservative treatment.

[0176] At the same time, the threshold is not fixed and applicable to all populations, but should be set in combination with bone mineral density parameters or historical sample statistical results; those with lower bone mineral density usually have poorer tissue tolerance, so their failure threshold can be relatively more conservative.

[0177] When a certain population group is more likely to progress to surgical cases under specific stress levels in historical samples, the threshold range can also be calibrated accordingly; in the specific numerical conversion rules, the system pre-sets the baseline stress failure threshold of standard healthy adults in the database. Compared with the baseline conservative observation threshold ;

[0178] When the actual bone mineral density T-value of the target object is obtained, the system calculates an individualized threshold based on structured quantification rules, which incorporates an attenuation coefficient. ;

[0179] The specific mapping rule is: when When bone mineral density is determined to be within the normal range, a decay factor is set. ;when When osteopenia is detected, a linear attenuation strategy is used to calculate the attenuation coefficient. ;when At that time, osteoporosis was diagnosed, and a more stringent attenuation strategy was used to calculate the attenuation coefficient. The system calculates the individualized stress failure threshold for the target object:

[0180]

[0181] And individualized conservative observation thresholds:

[0182]

[0183] This threshold dynamic calibration method based on quantitative rules avoids the mistaken identification of people with fragile bones as low-risk by a single fixed assessment standard.

[0184] When the peak maximum stress exceeds the stress failure threshold, the system generates a high-risk classification label and outputs a comprehensive result including surgical intervention recommendations.

[0185] The surgical intervention instructions mentioned here do not replace the doctor's decision to perform surgery, but rather refer to suggestions to enter the preoperative assessment pathway and to further refine high-level clinical procedures such as computed tomography or lower limb alignment assessment.

[0186] The underlying logic is that excessive local stress often indicates that the joint is on the verge of mechanical decompensation, and it may be difficult to reverse the situation by simply relying on pain relief and simple rehabilitation training.

[0187] When the peak stress is between two thresholds, the system generates a medium-risk label and outputs conservative treatment recommendations. Such cases usually present as structural abnormalities that have affected function, but no obvious mechanical failure has yet occurred. Therefore, weight loss, physical therapy, muscle strength training, gait adjustment or drug treatment can be given priority. At the same time, more frequent follow-up examinations should be arranged to observe whether stress-related symptoms persist.

[0188] When the peak maximum stress is not higher than the conservative observation threshold, the system generates a low-risk label and outputs routine follow-up recommendations. This does not mean that the imaging is completely normal, but rather that no obvious high-risk stress concentration has been observed under the current load-bearing conditions. It is appropriate to focus on monitoring and avoid overtreatment.

[0189] As an anomaly handling mechanism, if bone mineral density parameters are missing, the threshold can be temporarily called according to the historical statistical range of people of the same age and gender, but the result is marked as the threshold set by group compensation.

[0190] If the maximum stress peak is close to the threshold boundary of the two categories, the system can simultaneously combine stress gradient, symptom score or the length of previous disease course for auxiliary correction to avoid the boundary cases being mechanically classified; if the confidence of the three-dimensional stress result itself is insufficient, such as low reconstruction quality, the risk judgment subunit will not directly output clear intervention instructions, but only give a prompt that it is necessary to combine image re-acquisition or manual review.

[0191] In this outpatient setting, the system extracted that the maximum peak stress in the anterolateral contact area of ​​the affected side was significantly higher than the upper limit of the failure reference for patients with similar bone mineral density. Therefore, a high-risk label was generated, and the comprehensive results suggested that the patient should be included in the surgical evaluation pathway. At the same time, the choice of surgical procedure was improved by combining the clinical pain level and the degree of activity limitation.

[0192] If the patient's peak stress level is only in the middle range, the system will switch to recommending conservative treatment and short-term follow-up examinations instead of directly pushing for high-level intervention.

[0193] The purpose of this step is to further solidify the results of multimodal analysis into actionable risk stratification and treatment recommendations, thereby achieving a closed-loop connection from algorithm scoring to clinical pathway triggering.

[0194] In a preferred embodiment of the present invention, the model adaptive update module includes: a bias calculation unit, used to compare the functional risk comprehensive assessment result with the real clinical diagnostic label and calculate the risk prediction bias; and a gradient backpropagation unit, used to calculate the gradient of the loss function based on the risk prediction bias.

[0195] The weight update unit is used to backpropagate the gradient of the loss function to the morphological feature extraction network, generative adversarial network, graph neural network and fully connected layer according to the preset learning rate, and update the weight parameters.

[0196] This embodiment provides a refined implementation mechanism for a model adaptive update module; specifically, in a real clinical environment, the imaging equipment of different hospitals, the structure of the patient population, and the treatment habits will all affect the model performance;

[0197] If the system uses the initial training parameters for a long time without absorbing the posterior results, it may be insufficient to adapt to the cases in our hospital. Therefore, this embodiment forms an update mechanism that continuously evolves with clinical feedback through bias calculation, gradient backpropagation and weight update.

[0198] Specifically, the deviation calculation unit compares the comprehensive evaluation results with the true clinical diagnostic labels; these true labels can come from the final discharge diagnosis, actual findings during surgery, follow-up pain improvement, whether the disease progresses after conservative treatment, and postoperative pathology or imaging review results; its significance is that the system can not only fit the current clinical diagnostic labels, but also establish a mapping relationship with the patient's true clinical outcome.

[0199] For example, some cases do not appear to be very serious on two-dimensional images at the initial diagnosis, but progress rapidly in a short period of time and require surgical intervention. Such results indicate that the system should improve its sensitivity to specific stress patterns.

[0200] The gradient backpropagation unit calculates the gradient of the loss function based on the prediction bias and transmits the error information back to each submodule. To better adapt to medical safety requirements, the loss function is constructed using asymmetric weighted cross-entropy loss.

[0201] The system represents the comprehensive evaluation results as a predicted probability distribution vector, converts the real clinical diagnostic labels into one-hot codes of the real state, and calculates the cross-entropy between the two.

[0202] During the calculation process, the system introduces an asymmetric penalty sensitivity matrix: for false negative samples that are actually high-risk but are misclassified as medium- or low-risk (i.e., missed diagnosis), a penalty weight that is significantly greater than that for false positive samples that are actually medium- or low-risk but are misclassified as high-risk (i.e., misdiagnosis).

[0203] This is because the physical harm caused by the high risk of missed diagnosis in clinical practice leading to patients missing the opportunity for surgery is greater than the slight escalation of conservative recommendations; by multiplying the above asymmetric penalty weights with the basic cross-entropy, the target loss value is finally output and its derivative is used to obtain the gradient of the loss function;

[0204] For ease of explanation, a simplified scenario can be used: If the system previously classified a case as medium risk and recommended conservative treatment, but the case was confirmed as high risk during a short-term follow-up and underwent surgery, the error signal will drive the network to increase the feature extraction weights for the combination of local stress concentration and specific texture degradation.

[0205] Conversely, if the system has over-classified a risk level but the patient is actually stable with conservative treatment, it will suppress the over-response to certain simple imaging abnormalities without obvious functional decompensation patterns.

[0206] The weight update unit performs joint updates on the morphological feature extraction network, generative adversarial network, graph neural network, and final classification layer according to a preset learning rate. The significance of using joint updates here is that this system is not a single classifier, but is composed of two-dimensional recognition, three-dimensional reconstruction, mechanical deduction, and risk output coupled together.

[0207] If the conclusions at a certain stage are inaccurate, the reasons may be due to the offset of the front-end dissection points, the oversmoothing of the three-dimensional surface, insufficient learning of force propagation, or the bias of fusion judgment. Therefore, error information should be allowed to flow back along the link, rather than being repaired only locally at one layer.

[0208] As a system fault tolerance control mechanism, if there is a delay in the actual clinical label, such as the patient has not yet completed the follow-up, the case will not be entered into the formal update queue for the time being, but will only be used as a cache of samples to be labeled.

[0209] If labels from different sources conflict with each other, such as inconsistencies between outpatient diagnoses and intraoperative findings, the final result with higher confidence will be used first, or it will be manually reviewed before being included in the training set. If too few new samples are added in a short period of time, which is insufficient for stable updates, the system can adopt a batch accumulation strategy and update uniformly after reaching the preset sample size to avoid the model being disturbed by individual extreme cases.

[0210] After systematic evaluation, the outpatient entered the surgical evaluation pathway. Subsequent clinical records showed that there was indeed significant wear of the anterolateral cartilage during the operation, which was consistent with the high stress concentration area predicted by the system.

[0211] This case was therefore recorded as a predicted matching sample to consolidate existing parameters; if another similar imaging case is judged as high risk by the system, but the six-month follow-up shows that the symptoms have been significantly relieved and have not progressed after weight loss and rehabilitation treatment, then this case will be used as an over-predicted sample to participate in the update, prompting the system to more carefully distinguish people with severe morphological conditions but no functional loss of compensation in the future.

[0212] The purpose of this step is to establish a continuous optimization link that is linked to real clinical outcomes, so as to enable the system to gradually adapt to the characteristics of cases in our hospital during long-term deployment, thereby improving the stability and applicability of the comprehensive assessment results.

[0213] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A hip joint image classification system based on deep learning, characterized in that, include: The data acquisition module is used to acquire two-dimensional image data and basic physical parameters of the target object; The morphological feature extraction module is used to input the two-dimensional image data into a pre-trained morphological feature extraction network, extract two-dimensional key anatomical point features and pathological texture features, and generate basic imaging classification probabilities. The implicit geometry reconstruction module is used to input the two-dimensional key anatomical point features and the two-dimensional image data into a pre-trained generative adversarial network, and combine them with the prior constraints of a pre-constructed three-dimensional statistical shape model to inversely map and generate the three-dimensional implicit geometric surface of the target object. The virtual biomechanical deduction module is used to discretize the three-dimensional implicit geometric surface into a three-dimensional finite element mesh, and, in combination with the basic physical parameters, calculate the three-dimensional contact stress distribution characteristics of the target object under the load-bearing state corresponding to the basic physical parameters through a pre-trained graph neural network. The functional risk joint classification module is used to fuse the basic imaging classification probability, the pathological texture features and the three-dimensional contact stress distribution features into a multimodal feature, input a preset fully connected layer, and output a comprehensive functional risk assessment result. The model adaptive update module is used to obtain real clinical diagnostic labels, compare the functional risk comprehensive assessment results with the real clinical diagnostic labels, obtain risk prediction bias, and adjust the weight parameters of the morphological feature extraction network, the generative adversarial network, the graph neural network, and the fully connected layer based on the risk prediction bias.

2. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The data acquisition module includes: An image acquisition unit is used to receive the two-dimensional image data sent by a preset acquisition device; The two-dimensional image data includes a grayscale matrix of the region corresponding to the target object; A parameter acquisition unit is used to acquire the basic physical parameters corresponding to the target object; The basic physical parameters include preset gravity load parameters and geometric boundary condition parameters.

3. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The morphological feature extraction module includes: The feature encoding unit is used to perform feature dimensionality reduction on the two-dimensional image data through a preset convolutional layer in the morphological feature extraction network to generate a multi-scale feature map. The anatomical point localization unit is used to focus spatial features on the multi-scale feature map through a preset attention mechanism module and output the two-dimensional key anatomical point features. The texture extraction unit is used to perform high-frequency feature filtering on the multi-scale feature map and output the pathological texture features. The probability generation unit is used to input the multi-scale feature map into a preset classification head and output the basic image classification probability.

4. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The implicit geometry reconstruction module includes: The feature stitching unit is used to stitch the two-dimensional key anatomical point features with the two-dimensional image data along the channel dimension to generate a fused feature vector; The surface generation unit is used to input the fused feature vector into the preset generator in the generative adversarial network and output the initial three-dimensional point cloud data. An implicit mapping unit is used to apply surface smoothing constraints to the initial three-dimensional point cloud data through the prior constraints of the pre-constructed three-dimensional statistical shape model, thereby generating the three-dimensional implicit geometric surface.

5. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The virtual biomechanical deduction module includes: Mesh generation units are used to discretize the three-dimensional implicit geometric surface to generate the three-dimensional finite element mesh; Boundary condition application elements are used to map the basic physical parameters onto multiple mesh nodes of the three-dimensional finite element mesh to construct a mechanical simulation model; The stress calculation unit is used to transform the mechanical simulation model into a graph topology, wherein the mesh nodes are used as graph nodes and the connection relationships of the mesh cells are used as edges of the graph. The graph topology is input into the graph neural network for iterative solution and outputs the three-dimensional contact stress distribution characteristics. The three-dimensional contact stress distribution characteristics include the stress peak value and stress gradient vector of each node.

6. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The functional risk joint classification module includes: A modal alignment unit is used to map the basic imaging classification probability, the pathological texture features and the three-dimensional contact stress distribution features to the same preset dimension space to generate an alignment feature matrix. The feature fusion unit is used to calculate the association weights of each feature in the aligned feature matrix through a preset cross-attention network to generate global fused features; The classification output unit is used to input the global fusion features into the fully connected layer and output the comprehensive functional risk assessment result.

7. The deep learning-based hip joint image classification system according to claim 6, characterized in that, The classification output unit includes a preset risk assessment subunit, and the execution logic of the risk assessment subunit includes: Extract the maximum stress peak value from the three-dimensional contact stress distribution characteristics; If the maximum stress peak value is greater than the preset stress failure threshold, a high-risk classification label is generated, and the functional risk comprehensive assessment result containing the preset surgical intervention instructions is output. If the maximum stress peak value is less than or equal to the stress failure threshold and greater than the preset conservative observation threshold, a medium-risk classification label is generated, and the functional risk comprehensive assessment result containing preset conservative treatment instructions is output. If the maximum stress peak is less than or equal to the conservative observation threshold, a low-risk classification label is generated, and the functional risk comprehensive assessment result containing preset routine follow-up instructions is output. The stress failure threshold and the conservative observation threshold are pre-set based on the bone density parameters of the target object or the statistical distribution of historical clinical samples, and the stress failure threshold is greater than the conservative observation threshold.

8. The deep learning-based hip joint image classification system according to claim 1, characterized in that, The model adaptive update module includes: The deviation calculation unit is used to compare the functional risk comprehensive assessment result with the actual clinical diagnostic label and calculate the risk prediction deviation; The gradient backpropagation unit is used to calculate the gradient of the loss function based on the risk prediction bias. The weight update unit is used to backpropagate the gradient of the loss function to the morphological feature extraction network, the generative adversarial network, the graph neural network, and the fully connected layer according to a preset learning rate, and update the weight parameters.

Images