Joint disc modeling method based on structure mask and contrast learning mechanism

Abstract

The invention discloses a joint disc modeling method based on a structure mask and a comparative learning mechanism, and relates to the technical field of artificial intelligence and medical image processing, and the method comprises the following steps: S100, generating a mask and initializing mask parameters; s200, optimizing mask parameters; s300, determining an association loss function; s400, boundary reinforcement contrast learning is carried out; s500, training a joint disc model; and S600, outputting the joint disc model. According to the articular disc model constructed by the method, the technical problems of fuzzy boundary, incomplete form reconstruction, poor boundary point connectivity and lack of structural symmetry in the articular disc segmentation and reconstruction process are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and medical image processing, and particularly relates to a method for modeling the articular disc based on a structure mask and a contrast learning mechanism. Background Art

[0002] As a key soft tissue structure in the temporomandibular joint (TMJ) system, the articular disc is located between the mandibular condyle and the temporal bone, and its main functions are to buffer mechanical loads, guide mandibular movement, and maintain joint stability. The integrity, morphology, and movement trajectory of the articular disc have an important impact on the normal operation of oral functions such as chewing, swallowing, and speech pronunciation.

[0003] Clinically, pathological changes of the articular disc are closely related to various temporomandibular joint disorders (TMDs), including but not limited to abnormal states such as anterior displacement, posterior displacement, perforation, atrophy of the articular disc, and reduction of articular disc dislocation. If accurate diagnosis and intervention cannot be carried out in a timely manner, serious consequences such as joint noise, pain, limited mouth opening, and even osteoarthritic degeneration may be caused. Therefore, the morphological evaluation and dynamic monitoring of the articular disc play an irreplaceable role in the imaging diagnosis and treatment plan formulation of temporomandibular joint diseases.

[0004] Magnetic resonance imaging (MRI) is considered the "gold standard" for evaluating the position and morphology of the articular disc, especially for evaluating its dynamic changes in different mouth opening states. However, current artificial intelligence image analysis technologies mainly focus on fields such as lung CT images, breast cancer screening, cardiac MRI, and brain MRI images. The research on MRI image analysis of the temporomandibular joint, especially anterior displacement of the articular disc, is still in its infancy. Due to the small volume and complex structure of the articular disc, and problems such as significant deformation and signal blurring that may occur during lesions, the performance of traditional image segmentation models significantly decreases for such targets. In addition, the articular discs of different patients vary greatly in morphology, which makes the traditional diagnosis process relying on empirical judgment prone to misjudgment when facing complex or atypical cases, especially more challenging for young doctors with insufficient clinical experience. Traditional methods also include: doctors initially judge by asking patients about symptoms such as pain in front of the ear, joint noise, chewing difficulties, limited mouth opening, etc., combined with facial palpation, mouth opening observation, etc., which is the most basic and common means; including mouth opening measurement (maximum mouth opening distance), deviation degree assessment, muscle tender point examination, combined with occlusion conditions for comprehensive analysis and functional evaluation; used to observe the structure of the hard tissues of the temporomandibular joint, such as the position of the condyle and the joint space, but with poor display ability for soft tissues (especially the articular disc), etc.

[0005] The main technical challenges faced by the artificial intelligence image analysis of the articular disc in the prior art are: 1. Strong boundary ambiguity and low contrast. Since the gray-scale contrast between the articular disc and surrounding tissues (such as the joint capsule, synovial cavity, etc.) in MRI images is extremely close, and affected by factors such as imaging slice thickness and noise interference, the edge contour of the articular disc is blurred, making it difficult to manually label and even more difficult to accurately and automatically segment through traditional image processing methods.

[0006] 2. Large structural variability. The anatomical differences between different patients are significant, and the morphological changes of the articular disc during mouth opening and closing are complex, making it difficult for the model to capture its stable geometric features; especially in the pathological state, the articular disc morphology shows irregular changes such as atrophy, fracture, and displacement, increasing the complexity of segmentation and modeling.

[0007] 3. Lack of spatial structure modeling and prior guidance mechanism. Although existing deep learning-based automatic segmentation methods have made some progress, most models ignore the anatomical trend and movement trajectory of the articular disc in three-dimensional space and fail to fuse structural priors for constraint, resulting in phenomena such as morphological loss, false positive enhancement, or fracture at the lesion or boundary regions.

[0008] 4. Weak point-to-point connection and poor boundary continuity. Especially in the boundary point detection task, the model is trained only based on pixel or point-level labels, unable to effectively capture the geometric relationship, semantic association, and structural coherence between points, and difficult to achieve boundary integrity modeling.

[0009] 5. Lack of symmetry and geometric equilibrium modeling. The temporomandibular joint has bilaterally symmetric anatomical features, while current models generally do not establish symmetry constraints, resulting in asymmetric bias or geometric distortion in the model output results in some severe lesion regions, seriously affecting subsequent three-dimensional reconstruction and clinical judgment.

[0010] In summary, the existing technologies cannot meet the requirements of accurate identification, continuous boundary restoration, and structural integrity reconstruction of the articular disc in complex clinical images, lacking a complete mechanism that integrates structural guidance, semantic enhancement, and boundary optimization.

[0011] Therefore, the technicians in this field are committed to developing a method for articular disc modeling based on a structural mask and contrast learning mechanism. Summary of the Invention

[0012] In view of the above-mentioned defects of the existing technologies, the technical problem to be solved by the present invention is the boundary blurring, incomplete morphological reconstruction, poor boundary point connectivity, and lack of structural symmetry in the process of articular disc segmentation and reconstruction.

[0013] This application introduces a structured masking mechanism, combines a point-to-semantic association loss function for deep supervision, constructs a boundary contrast loss using the symmetric replacement region of the articular disc centerline, enhances the global consistency of the articular disc morphology in the edge and central axis regions, solves the problems of blurred boundaries, structural occlusion, and discontinuous boundaries, and significantly improves the reconstruction accuracy of boundary points while maintaining structural integrity.

[0014] Specifically, based on the structured masking mechanism, straight pipe masks, bent pipe masks, and Gaussian block mask structures are proposed in combination with articular disc features to establish a spatial prior in the articular disc region, and reinforcement learning is used to dynamically optimize the mask, angle, and curvature parameters; an association loss function combining boundary points and inner points is constructed, local pigment contrast is performed through a perceptual scale map structure and a multi-layer perceptron (MLP) to enhance edge geometric continuity, a global dependence relationship between point pairs is established, and the discrimination ability of the articular disc tissue edge and internal structure is improved; based on the boundary-enhanced contrast learning strategy of articular disc centerline symmetry replacement, a geometric enhancement loss term is constructed using the definition of the articular disc centerline of boundary points and symmetric area replacement, and adjustable parameters are introduced to maximize the maximum coincidence of boundary points and the structural difference, further enhancing the characteristic distinction of edge points and the reconstruction of structural integrity.

[0015] In one embodiment of the present invention, a method for modeling an articular disc based on a structure mask and a contrast learning mechanism is provided, including the following steps: S100. Mask generation and mask parameter initialization. Input the articular disc image into a conventional segmentation network, extract the articular disc centerline, generate straight pipe masks, bent pipe masks, and Gaussian block masks, set the initial values of the mask parameters, and initialize the number of training rounds to 0; S200. Mask parameter optimization. Optimize the mask parameters based on reinforcement learning, perform mask prediction, obtain the optimal mask parameters, and save the set of articular disc boundary points; S300. Determination of the association loss function. Construct association point pairs, including boundary association point pairs and inner point association point pairs, learn high-dimensional embedding features, determine the joint association loss function, screen and optimize the set of articular disc boundary points, and obtain the association learning of the association point pairs to strengthen the boundary; S400. Boundary-enhanced contrast learning. Define the center of the articular disc centerline and divide the region of interest ROI into the left half region and the right half region , construct an area difference loss function, construct a contrast learning loss function, calculate the predicted boundary of the articular disc, construct a predicted boundary loss function of the articular disc, and calculate the set of symmetric boundary points of the articular disc centerline and the symmetric boundary of the articular disc centerline to complete the initial construction of the articular disc model; S500. Joint disc model training: Construct the overall loss function of the joint disc model, train the joint disc model using the Adam optimizer, increment the number of training epochs by one. When it is determined that the number of training epochs is less than the maximum number of epochs, return to step S200. When the number of training epochs is greater than or equal to the maximum number of epochs, end the training of the joint disc model; S600. Joint disc model output: Evaluate the joint disc model using model evaluation metrics. When the joint disc model meets the model evaluation metrics, save and output the joint disc model. Otherwise, update the initial values of the mask parameters according to the initial mask update strategy, set the number of training epochs to 0, and return to step S200.

[0016] Optionally, in the joint disc modeling method based on structural masks and contrast learning mechanism in the above embodiments, the mask includes: Straight pipe mask: Taking the center line of the joint disc as the main axis, generating a cylindrical region with a radius of r in the natural direction of the joint disc, and obtaining the straight pipe mask; Straight pipe mask: Taking the center line of the joint disc as the main axis, generating a cylindrical region with a radius of Curved pipe mask: Rotating the center line of the joint disc by as the new main axis, with a rotation angle of , and obtaining the region surrounded by the deflection curve around the new main axis, that is, the curved pipe mask; Gaussian block mask: Fitting an elliptical region at the boundary points of the joint disc to obtain an ellipsoidal mask based on boundary fitting and Gaussian distribution, that is, the Gaussian block mask.

[0017] Optionally, in the joint disc modeling method based on structural masks and contrast learning mechanism in any of the above embodiments, step S100 includes: S110. Extraction of the center line of the joint disc: Input the joint disc image into a conventional segmentation network to obtain the joint disc region mask and the set of joint disc boundary points. Use the skeleton extraction algorithm for the joint disc region mask to obtain the center line of the joint disc C ; S120. Mask generation: Generate masks based on the center line of the joint disc C and the set of joint disc boundary points, including straight pipe masks, curved pipe masks, and Gaussian block masks, and generate the corresponding mask regions; S130. Initialization of mask parameters: Set the initial values of the mask parameters. The mask parameters include a radius of , a rotation angle of , and Gaussian parameters . The initial radius is 10 mm, the initial rotation angle is 0, and the initial Gaussian parameter Is (10, 10), apply the initial value of the mask parameter to the articular disc image, perform local enhancement processing on the articular disc area, and improve feature representation.

[0018] Optionally, in the articular disc modeling method based on the structure mask and contrast learning mechanism in any of the above embodiments, the conventional segmentation network uses U-Net.

[0019] Optionally, in the articular disc modeling method based on the structure mask and contrast learning mechanism in any of the above embodiments, the skeleton extraction algorithm includes the Zhang-Suen thinning algorithm and the Medial Axis Transform algorithm.

[0020] Further, in the articular disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, the articular disc center line C Is represented as a set of a series of discrete points: ; Where Is the discrete point coordinate of the articular disc center line, i is the serial number of the discrete point of the articular disc center line, from 1 to M , M Is a positive integer and is the number of discrete points.

[0021] Further, in the articular disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, step S120 includes: S121. Generate a straight pipe mask. The straight pipe mask is used for the non-deformed area. Taking the articular disc center line C As the main axis, generate a cylindrical area with a radius of Along the natural direction (Y-axis) of the articular disc, and the direction is ; Let the articular disc center line be , and the coordinate of each point is, then the straight pipe mask Is defined as: ; Where Represents whether the point In the two-dimensional space belongs to the straight pipe mask area, and the value is 1 or 0. The mask value of the area centered on the point With a radius of And within is 1, and the others are 0; S122. Generate a bent pipe mask, The bent pipe mask is used for the mild to moderate deformation area. Rotate the articular disc center line C By As the new main axis, let the rotation angle be , and the new main axis function is denoted as , the deflection curve formula is: ; The elbow mask is defined as: ; where represents a point in the two-dimensional space whether it belongs to the Boolean value of the elbow mask area, taking values of 1 or 0, is the point on the new main axis, is the radius, with as the center, and the mask value of the area within and including the radius of is 1, and others are 0; S123, Gaussian block mask generation. The Gaussian block mask is used for the large deformation area. Perform local PCA (Principal Component Analysis) on the set of boundary points of the articular disc to extract the principal curvature direction , for the deformed area of the articular disc, generate an ellipsoidal mask based on boundary fitting and Gaussian distribution, that is, the Gaussian block mask. Fit an ellipse at the point . The mask intensity follows a two-dimensional Gaussian distribution, indicating the response weight of each pixel point in the two-dimensional plane to the final structure mask. Its value is between [0, 1], the center point value is 1, and it approaches 0 closer to the edge. The Gaussian block mask is defined as: ; where, is the standard deviation of the Gaussian distribution in the horizontal direction, used to control the expansion range of the fitted ellipse in the lateral edge deformation direction, with a range of 1.0 - 3.0 mm; is the standard deviation of the Gaussian distribution in the vertical direction, used to control the expansion range of the fitted ellipse in the longitudinal boundary extension direction, with a range of 0.5 - 2.0 mm.

[0022] Furthermore, in the articular disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, the non-deformed area is the area where the center line of the articular disc or the boundary curvature has no obvious change or the proportion of the deformed area is small. Specifically: the boundary curvature is in the range of ; or the offset distance is less than 1 mm; or the deformed area accounts for less than 10% of the total structure area of the articular disc.

[0023] Furthermore, in the articular disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, the mild to moderate deformation area is the area where the center line of the articular disc or the boundary curvature changes gently or the proportion of the deformed area is medium. Specifically: the boundary curvature is in the range of ; or the offset distance is in the range of [1 mm, 5 mm]; or the deformed area accounts for the total structural area of the articular disc in the range of [10%, 40%].

[0024] Further, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the large deformation area is the area where the center line or the boundary curvature of the articular disc changes violently or the proportion of the deformed area is large. Specifically, the boundary curvature ; or the offset distance is greater than 5 mm; or the deformed area accounts for more than 40% of the total structural area of the articular disc.

[0025] Further, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, step S200 includes: S210. Mask parameter adjustment: Adjust the mask parameters according to the adjustment strategy of the mask. The straight pipe mask reduces the radius with a fixed step size to refine the central area until the minimum radius; The bent pipe mask increases the rotation angle with a fixed angle until the maximum angle; the Gaussian block mask is enhanced with a fixed ellipsoidal flattening rate to strengthen the boundary envelope until the maximum change in the ellipsoidal flattening rate; S220. Mask parameter optimization: Optimize the mask parameters based on reinforcement learning to perform mask prediction, obtain the set of articular disc boundary points and the predicted boundary of the articular disc. When the boundary IoU (Intersection over Union) is lower than the set boundary IoU threshold, return to step S210 until the continuous N times the boundary IoU improvement rate is lower than the minimum change amount or the maximum number of iterations is reached, complete the mask prediction, obtain the optimal mask parameters, and save the set of articular disc boundary points.

[0026] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the fixed step size is 0.02 mm.

[0027] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the minimum radius is 5 mm.

[0028] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the fixed angle is 1 degree.

[0029] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the maximum angle is 90 degrees.

[0030] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, the oblateness of the ellipsoid is fixed at 0.2.

[0031] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, the maximum change in the oblateness of the ellipsoid is 2.

[0032] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, the set boundary IoU threshold is 0.7.

[0033] Optionally, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in any of the above embodiments, N the range of is greater than or equal to 3 and less than or equal to 5.

[0034] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, N = 3.

[0035] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, the set minimum change amount is 0.001.

[0036] Preferably, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, the maximum number of iterations = 500.

[0037] Furthermore, in the method for modeling the articular disc based on the structural mask and the contrast learning mechanism in the above embodiment, step S220 includes: S221. State space design, where the mask selection and mask parameter setting constitute the current state, denoted as: , wherein, represents the mask, i.e., the straight pipe mask, the bent pipe mask, and the Gaussian block mask, represents the spindle rotation angle, which is used for the bent pipe mask, , respectively represent the standard deviation of the Gaussian distribution in the horizontal direction and the standard deviation of the Gaussian distribution in the vertical direction, which are collectively called the Gaussian distribution width, is the radius of the tubular mask, k is an integer, representing the optional boundary curvature change rate; S222. Action space design, where an action is selected according to the current state, including switching the mask T and adjusting the mask parameters, including , and the action space is defined as: ; S223. Reward function design, define the reward function as the improvement in the detection of the boundary points of the articular disc improvement amount and the overall detection accuracy of the articular disc improvement amount weighted combination, and the reward function formula is as follows: ; wherein, , is the empirical weight coefficient, , and , preferably ; S224. Mask parameter optimization, use the Actor-Critic framework based on policy gradient to optimize the mask parameters, complete the mask prediction, obtain the optimal mask parameters, and save the set of articular disc boundary points. The policy gradient formula is as follows: ; wherein, represents the state and action distribution sampled from the current policy, represents the probability distribution of generating an action from the action space according to the policy network with parameters a in the current state s; represents the logarithmic probability gradient of the action a under the current policy; is the advantage function, which measures how much better the execution of the action a is than the average policy. The advantage function is defined as: wherein, is the discount factor, preferably taken as 0.95, which controls the influence degree of the future state value; is the value estimation of the next state a after the execution of the action ; is the value estimation of the current state s.

[0038] Optionally, in the articular disc modeling method based on the structure mask and the contrast learning mechanism in any of the above embodiments, step S300 includes: S310. Associated point pair construction, for the straight pipe mask, The different regions of the articular disc formed by the elbow pipe mask and the Gaussian block mask are divided, and each point is assigned a spatial role to form associated point pairs, which are classified into boundary-associated point pairs (Boundary-Paired Points, BPP) and interior-associated point pairs (Interior-Paired Points, IPP); S320. High-dimensional embedding feature learning: Use a multi-layer perceptron MLP embedding network to map the associated point pairs for high-dimensional embedding feature learning; S330. Construction of the associated point pair loss function: Construct a boundary-associated point pair (Boundary-Paired Points, BPP) loss function and an interior-associated point pair (Interior-Paired Points, IPP) loss function, and synthesize the boundary-associated point pair (BPP) loss function and the interior-associated point pair (IPP) loss function to generate a joint associated point pair loss function; S340. Screening of the boundary point set: Use the high-dimensional embedding feature and the associated point pair loss function to screen the articular disc boundary point set to obtain an enhanced constrained articular disc boundary point set and an associated point pair association learning to strengthen the boundary.

[0039] Furthermore, in the above-described articular disc modeling method based on a structure mask and a contrast learning mechanism, the spatial roles include boundary points, interior points, and boundary adjacent points.

[0040] Furthermore, in the above-described articular disc modeling method based on a structure mask and a contrast learning mechanism, the associated point pairs are classified as: Boundary-associated point pairs (Boundary-Paired Points, BPP), including the boundary points generated by the Gaussian block mask and the boundary point neighborhoods; Interior-associated point pairs (Interior-Paired Points, IPP), including the straight pipe mask, the interior point regions covered by the elbow pipe mask.

[0041] Furthermore, in the above-described articular disc modeling method based on a structure mask and a contrast learning mechanism, the input dimension of the multi-layer perceptron MLP embedding network is the original feature of the associated point pair, the hidden layer is [128, 64], and the output dimension is the final embedding dimension, and the activation function uses ReLU (Rectified Linear Unit, rectified linear unit) and BatchNorm (Batch Normalization, batch normalization).

[0042] Further, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the original features of the associated point pairs include the positions, intensity values, normal vectors, and masks of the associated point pairs.

[0043] Further, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the final embedding dimension is 32 or 64.

[0044] Optionally, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in any of the above embodiments, step S330 includes: S331. Construction of the boundary associated point pair (BPP) loss function. The boundary associated point pair (BPP) loss is the geometric continuity loss. Let any boundary associated point pair be , and the normal vectors are respectively . The boundary associated point pair (BPP) loss function The formula is: ; Among them, is the embedded feature of the boundary associated point pair after being mapped by the MLP, called the boundary associated point pair embedded feature, is the weight for balancing the consistency between the normal vector and the boundary associated point pair embedded feature represents the total number of boundary associated point pairs, that is, the number of all boundary associated point pairs participating in the calculation of ; S332. Construction of the inlier associated point pair (IPP) loss function. The inlier associated point pair (IPP) loss is the feature consistency loss. Let any inlier associated pair (IPP) be . The inlier associated point pair (IPP) loss function The formula is: ; Among them, represents the total number of inlier associated point pairs, that is, the number of all inlier associated point pairs (IPP) participating in the calculation of , represents the pigment value or voxel value of the inlier associated point pair, used to reflect its local color or density, represents the inlier associated pair (IPP) embedded feature, is a hyperparameter, , adjusting the weight between the color and density difference term and the embedded feature difference term of the inlier associated point pair (IPP); S333. Construction of the associated point pair loss function. The associated point pair loss function is the overall loss function, integrating the boundary associated point pair (BPP) loss function and the inlier associated point pair (IPP) loss function. The associated point pair loss function LThe formula is as follows: ; Among them, R is the reward value of the self-masking strategy, which is dynamically adjusted according to the correct classification or recognition degree of the masked category area. The larger the reward value, the more accurate the correct classification or recognition degree of the masked category area, and it is defined as: ; Among them, is the true mask, is the predicted mask, is the indicator function, which takes 1 when the predicted mask is correct and 0 otherwise; is a hyperparameter, , and the value of the hyperparameter is determined by cross-validation, which is used to balance the geometric continuity loss of the boundary pair of associated points (BPP), the feature consistency loss of the inlier pair of associated points (IPP), and the contribution of the reward function for the accuracy of the mask category.

[0045] Furthermore, in the above-mentioned joint disc modeling method based on the structural mask and the contrast learning mechanism, 5-fold cross-validation is used for cross-validation. In each fold of cross-validation, performance scoring and ranking are performed on each set of hyperparameter combinations , and the one with the highest score is selected.

[0046] Furthermore, in the above-mentioned joint disc modeling method based on the structural mask and the contrast learning mechanism, step S340 includes: S341. Cosine similarity calculation. For each pair of associated points P (i,j) , including the boundary pair of associated points BPP and the inlier pair of associated points IPP, the feature consistency score is calculated using the high-dimensional embedding features, that is, the cosine similarity is used to evaluate whether the pair of points is in a consistent semantic relationship in the feature space. The cosine similarity s ij The formula is as follows: ; Among them, s ij The closer the value is to 1, i , j represent the numbers of the two associated points in the pair of associated points, which means that the higher the similarity degree of the pair of associated points P (i,j) in the semantic relationship. The closer the s ij value is to 0, it means that the pair of points has almost no semantic association. If s ij the value is less than 0, it means that the semantic relationship between the two points is incorrect and may come from different categories; ​​S342. Calculate the comprehensive confidence for each associated point pair P (i,j) Calculate the comprehensive confidence , and the formula is as follows: ; Wherein, represents the number of associated point pairs, ranging from [0, 1]; S343. Screening of the enhanced constraint joint disc boundary point set, based on cosine similarity and comprehensive confidence, screen the boundary point set, select the boundary points with cosine similarity greater than or equal to the cosine similarity threshold and comprehensive confidence greater than or equal to the comprehensive confidence threshold, and obtain the enhanced constraint joint disc boundary point set and the associated point pair association learning enhanced boundary.

[0047] Further, in the joint disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, the cosine similarity threshold is equal to 0.9.

[0048] Further, in the joint disc modeling method based on the structure mask and contrast learning mechanism in the above embodiment, the comprehensive confidence threshold is equal to 0.8.

[0049] Optionally, in the joint disc modeling method based on the structure mask and contrast learning mechanism in any of the above embodiments, step S400 includes: S410. Definition of the center of the joint disc center line, the joint disc boundary point set is expressed as: ; Based on y the selection of the maximum spacing edge point pair of the axis, select among all boundary points y points with approximately the same coordinates, that is, the difference is less than a predetermined threshold , x the point pair with the largest coordinate difference as the left and right limit points: ; Wherein, , are the left and right limit points respectively, and the abscissas are the smallest abscissa and the largest abscissa among the points with approximately the same ordinate (being ) among all boundary points; S420. Region replacement and area difference construction, define the region of interest ROI , take the joint disc center line as the axis of symmetry, and divide all boundary points and the mask region into the left half region and the right half region , perform region replacement operations through a region replacement function to construct an area difference loss function; S430. Construction of the contrastive learning loss function: Extract the high-dimensional embedding features of the boundary points, define positive and negative sample pairs, and construct a contrastive learning loss function based on cosine similarity; S440. Boundary reconstruction and weighted fusion: Calculate the predicted boundary of the articular disc, construct a predicted boundary loss function of the articular disc, calculate the set of symmetric boundary points of the center line of the articular disc and the symmetric boundary of the center line of the articular disc, and complete the initial construction of the articular disc model.

[0050] Furthermore, in the articular disc modeling method based on a structure mask and a contrastive learning mechanism in the above embodiment, the predetermined threshold is set for normalization according to the actual resolution of the image, specifically the product of the vertical resolution and 0.01.

[0051] Optionally, in the articular disc modeling method based on a structure mask and a contrastive learning mechanism in any of the above embodiments, step S420 includes: S421. Define the region of interest, and construct a rectangle covering the main boundary region, that is, the region of interest ROI, The formula is as follows: ; where is an adjustment parameter that controls the expansion range of the region of interest ROI in the y axis direction; S422. Define the region replacement function. The formula of the region replacement function is as follows: ; where , that is, construct a mirror sample by reflecting through the center line of the articular disc, is the abscissa of the point after mirroring along the center line of the articular disc; is the position of the center line of the articular disc, that is, the dividing line between the left half region and the right half region ; S423. Construct the area difference loss function. Use the boundary point mask region as the basis for area calculation. Let the original boundary mask region be , the mask region after region replacement be , and define the area difference loss function as: ; The area difference loss function By comparing the spatial coincidence degree between the original boundary mask region and the mask region after region replacement, the symmetry of the articular disc structure is reflected, and the area difference loss function The larger the value, the greater the difference in the original structure in the left-right mirror state, and the worse the symmetry.

[0052] Optionally, in the articular disc modeling method based on the structure mask and the contrast learning mechanism in any of the above embodiments, step S430 includes: S431. Extract high-dimensional embedding features of boundary points. The formula for extracting high-dimensional embedding features of boundary points is as follows: ; Wherein, p is the boundary point; S432. Define positive sample pairs and negative sample pairs. The original boundary mask region undergoes geometric transformation to obtain positive samples, which are used to train the model to enhance boundary invariance and construct reasonable boundary variants close to the real structure as similar reference samples for the boundary structure; the original boundary mask region undergoes mirror transformation to obtain negative samples, which are used to identify asymmetry and improve the model's discrimination ability for inconsistent left and right structures, especially to identify lesions or irregular morphologies; the original boundary mask region and the positive sample constitute a positive sample pair, and the original boundary mask region and the negative sample constitute a negative sample pair; S433. Define the contrast learning loss. Construct a contrast learning loss function based on cosine similarity , which is defined as follows: ; Wherein, represents the cosine similarity, is the temperature constant, is the feature representation of the original boundary mask region, is the feature representation of the positive sample, is the feature representation of the negative sample.

[0053] Optionally, in the articular disc modeling method based on the structure mask and the contrast learning mechanism in any of the above embodiments, the geometric transformation includes rotation and translation.

[0054] Optionally, in the articular disc modeling method based on the structure mask and the contrast learning mechanism in any of the above embodiments, step S440 includes: S441. Boundary weighted fusion. Perform weighted fusion on the predicted boundary, the associated point pair for associated learning to strengthen the boundary, and the symmetric boundary of the articular disc center line to obtain the predicted boundary of the articular disc as follows: ; ; Among them, is the predicted boundary, emphasizing the boundary position perception ability; is the associated learning to strengthen the boundary, emphasizing the perception of associated point pairs and improving local geometric continuity; is the symmetric boundary of the center line of the articular disc, considering the symmetry of the articular disc edge and the area reconstruction difference, emphasizing the repair of blurred areas or missing parts and reflecting the boundary complementation ability; are respectively weighting coefficients of, as learnable parameters, and normalizing through the gradient descent Softmax function for to ensure the following constraints are met: ; Among them, ; S442. Construction of the loss function for the predicted boundary of the articular disc. The loss function for the predicted boundary of the articular disc is constructed as follows: ; Among them, are respectively the hyperparameters of the area difference loss function and the hyperparameters of the contrast learning loss function; S443. Calculate the set of symmetric boundary points of the center line of the articular disc and the symmetric boundary of the center line of the articular disc. For any point in the set of articular disc boundary points, the point is obtained by mapping through the center line of the articular disc, and a set of symmetric point pairs is constructed, specifically expressed as: ; The set of symmetric boundary points of the center line of the articular disc is expressed as: ; That is, the union of the set of articular disc boundary points and the set of symmetric point pairs ; Use the curve reconstruction method to perform boundary fitting on the set of symmetric boundary points of the center line of the articular disc , specifically expressed as: ; Among them, Rec(.) represents using the curve reconstruction method to perform boundary fitting on the set of symmetric boundary points of the center line of the articular disc to obtain the symmetric boundary of the center line of the articular disc.

[0055] Optionally, in the method for modeling the articular disc based on the structure mask and contrast learning mechanism in any of the above embodiments, the curve reconstruction method includes the α-Shape algorithm, the Delaunay algorithm, and the B-spline algorithm.

[0056] Optionally, in the joint disc modeling method based on the structural mask and contrast learning mechanism in any of the above embodiments, step S500 includes: S510. Maintaining boundary structure symmetry. To ensure continuous reinforcement of the joint disc structure symmetry, the consistency loss function formula is defined as follows: ; where, is the number of boundary points, is the feature extraction function, is the boundary point, is the symmetric reflection point, that is, the joint disc model learns the structural symmetry attribute of the boundary point; S520. Defining the overall loss function. The overall loss function formula of the joint disc model is as follows: ; where, is the weight coefficient, ; S530. Training the joint disc model. Train the joint disc model through the Adam optimizer, increment the number of training rounds by one. When it is determined that the number of training rounds is less than the maximum number of training rounds, return to step S200. When the number of training rounds is greater than or equal to the maximum number of training rounds, end the training of the joint disc model. The joint disc model is represented as follows: .

[0057] Preferably, in the joint disc modeling method based on the structural mask and contrast learning mechanism in the above embodiment, the maximum number of rounds T max = 2000.

[0058] Optionally, in the joint disc modeling method based on the structural mask and contrast learning mechanism in any of the above embodiments, the model evaluation metrics include the boundary accuracy evaluation metric and the structural symmetry evaluation metric.

[0059] Optionally, in the joint disc modeling method based on the structural mask and contrast learning mechanism in any of the above embodiments, step S600 includes: S610. Evaluating the prediction boundary accuracy. Use the boundary accuracy evaluation metric to evaluate the accuracy of the joint disc boundary predicted by the joint disc model, and evaluate the overlap degree between the predicted joint disc boundary region and the true joint disc boundary region; S620. Evaluating the prediction structural symmetry. Use the structural symmetry evaluation metric to evaluate the symmetry of the joint disc structure predicted by the joint disc model; S630. Evaluate and output the articular disc model. When the articular disc model meets the boundary accuracy evaluation index and the structural symmetry evaluation index, save and output the articular disc model. Otherwise, update the initial value of the mask parameter according to the initial mask update strategy, set the number of training rounds to 0, and return to step S200.

[0060] Further, in the articular disc modeling method based on the structural mask and the contrast learning mechanism in the above embodiment, the boundary accuracy evaluation index includes: Dice coefficient, the formula is as follows: ; Where, P is the predicted boundary region of the articular disc, G is the true boundary region of the articular disc. The Dice coefficient reflects the overlap degree between the predicted boundary region of the articular disc and the true boundary region of the articular disc. The value range is [0,1]. The closer the Dice coefficient value is to 1, the more consistent the predicted boundary region of the articular disc is with the true boundary region of the articular disc, the higher the coincidence degree, and the more accurate the articular disc model. The closer the Dice value is to 0, it means that the coincidence degree between the predicted boundary region of the articular disc and the true boundary region of the articular disc is lower, indicating that the construction of the articular disc model fails; IoU (Jaccard), the formula is as follows: ; IoU (Jaccard) measures the overlap degree between the predicted boundary region of the articular disc and the true boundary region of the articular disc by calculating the intersection-over-union ratio of the predicted boundary region of the articular disc and the true boundary region of the articular disc, so as to reflect whether the articular disc model is overfitting or underfitting, IoU (Jaccard) has a value range of [0,1], IoU (Jaccard) The closer the value is to 1, the higher the coincidence degree between the predicted boundary region of the articular disc and the true boundary region of the articular disc, and the more accurate the articular disc model, IoU (Jaccard) The closer the value is to 0, it means that the complete coincidence degree between the predicted boundary region of the articular disc and the true boundary region of the articular disc is lower, indicating that the construction of the articular disc model fails.

[0061] Further, in the articular disc modeling method based on the structural mask and the contrast learning mechanism in the above embodiment, the structural symmetry evaluation index uses the structural symmetry index (Symmetry Consistency Score, SCS) to quantify the error of the left-right symmetric structure. The formula is as follows: ; SCS Measures the left-right structural consistency of the predicted boundary structure of the articular disc model relative to the axis. The predicted boundary point The average spatial deviation between it and the symmetric mapping point about the center line of the articular disc When predicting the high symmetry of the boundary structure, SCS the value approaches 0. When predicting the contour deformation of the boundary structure, SCS the larger the value.

[0062] Furthermore, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the initial mask update strategy is: combining the evaluation index, when the Dice coefficient value is greater than the Dice threshold or IoU (Jaccard) is less than IoU the threshold, the initial radius increases by a specified radius increment, and the initial rotation angle increases by a specified angle increment; when SCS the value is greater than SCS the threshold, the Gaussian parameter increases by a specified parameter increment.

[0063] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the Dice threshold is equal to 0.85.

[0064] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, IoU the threshold is equal to 0.75.

[0065] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, SCS the threshold is equal to 5%.

[0066] Furthermore, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the range of the specified radius increment is [0.01, 0.03].

[0067] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the specified radius increment is 0.02.

[0068] Furthermore, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the range of the specified angle increment is [1, 3].

[0069] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the specified angle increment is 2.

[0070] Furthermore, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the range of the specified parameter increment is [1, 2].

[0071] Preferably, in the method for modeling the articular disc based on the structure mask and the contrast learning mechanism in the above embodiment, the specified parameter increment is 1.

[0072] The present invention realizes the reconstruction of the articular disc structure and the enhancement of the edge details through an articular disc model constructed by combining mask generation, point-to-point association learning, and boundary reinforcement contrast learning, improves the detection sensitivity and positioning accuracy of the articular disc area, enhances the structural consistency between the boundary points and the internal points, improves the restoration of the missing area of the morphology and enhances the reliability of the boundary reconstruction, and improves the response ability in the weak signal area of the boundary. The articular disc model constructed by the present invention solves the technical problems of blurred boundaries, incomplete morphological reconstruction, poor connectivity of boundary points, and lack of structural symmetry in the process of articular disc segmentation and reconstruction.

[0073] The following will further illustrate the concept, specific structure and technical effects of the present invention with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention. Description of the Drawings

[0074] Figure 1 is a flowchart of the method for modeling the articular disc based on the structure mask and the contrast learning mechanism of the exemplary embodiment. Detailed Embodiments

[0075] The following introduces multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification, so that its technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0076] In the drawings, components with the same structure are denoted by the same reference numerals, and components with similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the drawings clearer, the thickness of some parts in the drawings is schematically exaggerated appropriately.

[0077] The present application provides a method for modeling the articular disc based on the structure mask and the contrast learning mechanism, as Figure 1 shown, including the following steps: S100. Mask generation and mask parameter initialization. Input the articular disc image into a conventional segmentation network, extract the center line of the articular disc, and generate a mask, including: Straight pipe mask. Taking the center line of the articular disc as the main axis, generate a cylindrical area with a radius of r and a direction of along the natural direction of the articular disc to obtain the straight pipe mask; Curved pipe mask. Rotate the center line of the articular disc by As the new principal axis, the rotation angle is , the area obtained by the deflection curve around the new principal axis, that is the elbow mask; The Gaussian block mask, fitting the elliptical area at the boundary points of the articular disc, obtains the ellipsoidal mask based on boundary fitting and Gaussian distribution, that is, the Gaussian block mask; Step S100 specifically includes: S110. Extract the center line of the articular disc. Input the articular disc image into the conventional segmentation network U-Net to obtain the articular disc region mask and the set of articular disc boundary points. Use the skeleton extraction algorithm for the articular disc region mask to obtain the center line of the articular disc C， The skeleton extraction algorithm uses the Medial Axis Transform algorithm, and the center line of the articular disc C is represented as a set of a series of discrete points: ; Among them, is the discrete point coordinate of the center line of the articular disc, i is the serial number of the discrete point of the center line of the articular disc, from 1 to M , M is a positive integer, and is the number of discrete points; S120. Mask generation, generating masks based on the center line of the articular disc C and the set of articular disc boundary points, including the straight pipe mask, the elbow mask and the Gaussian block mask, and generating the corresponding mask regions; specifically including: S121. Straight pipe mask generation, the straight pipe mask is used for the non-deformed area, and the non-deformed area is the area where the center line of the articular disc or the boundary curvature has no obvious change or the proportion of the deformed area is small. Specifically: the boundary curvature range is ; or the offset distance is less than 1 mm; or the deformed area accounts for less than 10% of the total structural area of the articular disc. Taking the center line of the articular disc C as the principal axis, generating a cylindrical area with a radius of along the natural direction (Y-axis) of the articular disc, and the direction is ; Let the center line of the articular disc be , and the coordinate of each point is , then the straight pipe mask is defined as: ; Among them represents the Boolean value of whether the point in the two-dimensional space belongs to the straight pipe mask area, taking values of 1 or 0. The mask value of the area centered on the point with a radius of and within is 1, and others are 0; S122. Elbow mask generation The elbow mask is used for the mild to moderate deformation area, which is the area where the center line of the articular disc or the boundary curvature changes gently or the proportion of the deformed area is medium. Specifically: the boundary curvature ranges from ; or the offset distance is in the interval [1mm, 5mm]; or the deformed area accounts for the total structural area of the articular disc in the interval [10%, 40%]. Taking the center line of the articular disc C rotation as the new main axis, setting the rotation angle as and denoting the new main axis function as , the deflection curve formula is: ; The elbow mask is defined as: ; where represents the point in the two-dimensional space whether it belongs to the boolean value of the elbow mask area, taking values of 1 or 0, is the point on the new main axis, is the radius, with as the center and the radius of and within the area, the mask value is 1, and others are 0; S123, Gaussian block mask generation. The Gaussian block mask is used for the large deformation area, which is the area where the center line of the articular disc or the boundary curvature changes violently or the proportion of the deformed area is large. Specifically: the boundary curvature ; or the offset distance is greater than 5mm; or the deformed area accounts for more than 40% of the total structural area of the articular disc. Perform local PCA (Principal Component Analysis) on the set of boundary points of the articular disc to extract the principal curvature direction . For the deformed area of the articular disc, generate an ellipsoidal mask based on boundary fitting and Gaussian distribution, that is, the Gaussian block mask. Fit an ellipse at the point . The mask intensity follows a two-dimensional Gaussian distribution, representing the response weight of each pixel point in the two-dimensional plane to the final structural mask. Its value is between [0, 1], the center point value is 1, and it approaches 0 closer to the edge. The Gaussian block mask is defined as: ; where, is the standard deviation of the Gaussian distribution in the horizontal direction, used to control the expansion range of the fitted ellipse in the lateral edge deformation direction, with a range of 1.0–3.0mm; is the standard deviation of the Gaussian distribution in the vertical direction, which is used to control the expansion range of the fitted ellipse in the longitudinal boundary extension direction, and the range is 0.5–2.0 mm; S130. Initialize the mask parameters, set the initial values of the mask parameters. The mask parameters include a radius of , a rotation angle of , and the Gaussian parameter . The initial radius is 10 mm, the initial rotation angle is 0, and the initial Gaussian parameter is (10, 10). Apply the initial values of the mask parameters to the joint disc image, perform local enhancement processing on the joint disc area to improve feature representation, and initialize the number of training rounds to 0.

[0078] S200. Optimize the mask parameters. Optimize the mask parameters based on reinforcement learning, perform mask prediction, obtain the optimal mask parameters, and save the set of joint disc boundary points; specifically including: S210. Adjust the mask parameters. Adjust the mask parameters according to the mask adjustment strategy. The straight pipe mask reduces the radius at a fixed step size, and the fixed step size is 0.02 mm, and the central area is refined until the minimum radius, and the minimum radius is 5 mm; The bent pipe mask increases the rotation angle at a fixed angle, and the fixed angle is 1 degree until the maximum angle, and the maximum angle is 90 degrees; the Gaussian block mask is enhanced with a fixed ellipsoidal flattening rate, and the fixed ellipsoidal flattening rate is 0.2 to strengthen the boundary envelope until the maximum change in the ellipsoidal flattening rate, and the maximum change in the ellipsoidal flattening rate is 2; S220. Optimize the mask parameters. Optimize the mask parameters based on reinforcement learning, perform mask prediction, obtain the set of joint disc boundary points and the predicted boundary of the joint disc. When the boundary IoU (Intersection over Union) is lower than the set boundary IoU threshold of 0.7, return to step S210 until the continuous 3 times of boundary IoU improvement rate is lower than the minimum change amount of 0.001 or reaches the maximum number of iterations of 500, complete the mask prediction, obtain the optimal mask parameters, and save the set of joint disc boundary points; specifically including: S221. Design the state space. The mask selection and mask parameter setting constitute the current state, denoted as: ; Among them, represents the mask, that is, the straight pipe mask, the bent pipe mask and the Gaussian block mask, represents the main axis rotation angle, which is used for the bent pipe mask, , respectively represent the standard deviation of the Gaussian distribution in the horizontal direction and the standard deviation of the Gaussian distribution in the vertical direction, which are collectively called the Gaussian distribution width. is the radius of the tubular mask. k is an integer representing the optional boundary curvature change rate. S222, Action space design, select an action according to the current state, including switching the mask T and adjusting the mask parameters, including , the action space is defined as: ; S223, Reward function design, define the reward function as the improvement in the detection of the boundary points of the joint disc improvement and the improvement in the overall detection accuracy of the joint disc improvement weighted combination, the reward function formula is as follows: ; where, , are empirical weight coefficients, , and , preferably .

[0079] S224, Mask parameter optimization, use the Actor-Critic framework based on policy gradient to optimize the mask parameters, complete mask prediction, obtain the optimal mask parameters, and save the set of joint disc boundary points. The policy gradient formula is as follows: ; where, represents the state and action distribution sampled from the current policy, represents, in the current state s, the probability distribution of generating an action from the action space according to the policy network with parameters a ; represents the logarithmic probability gradient of the action a under the current policy; is the advantage function, which measures how much better the execution of the action a is than the average policy. The advantage function is defined as: ; where, is the discount factor, preferably taken as 0.95, which controls the influence degree of the future state value; is the value estimation of the next state a after executing the action ; is the value estimation of the current state s.

[0080] S300. Determine the associated loss function, construct associated point pairs, including boundary associated point pairs and interior point associated point pairs, learn high-dimensional embedding features, determine the joint associated loss function, screen and optimize the set of articular disc boundary points, and obtain associated point pair associated learning to strengthen the boundary. Specifically, it includes: S310. Construct associated point pairs. Divide different regions of the articular disc formed by the straight pipe mask, bend pipe mask, and Gaussian block mask, assign a spatial role to each point, and the spatial roles include boundary points, interior points, and boundary adjacent points. Form associated point pairs, and classify the associated point pairs as: Boundary-Paired Points (BPP), including boundary points generated by the Gaussian block mask and the boundary point neighborhood; Interior-Paired Points (IPP), including the interior point regions covered by the straight pipe mask, bend pipe mask; S320. Learn high-dimensional embedding features. Use a multi-layer perceptron MLP embedding network to map the associated point pairs and perform high-dimensional embedding feature learning. The input dimension of the multi-layer perceptron MLP embedding network is the original features of the associated point pairs, including the position, intensity value, normal vector, and mask of the associated point pairs. The hidden layer is [128, 64], and the output dimension is the final embedding dimension 64. The activation function uses ReLU (Rectified Linear Unit) and BatchNorm (Batch Normalization); S330. Construct the associated point pair loss function. Construct the loss function for boundary associated point pairs (Boundary-Paired Points, BPP) and the loss function for interior point associated point pairs (Interior-Paired Points, IPP), and synthesize the loss function for boundary associated point pairs (BPP) and the loss function for interior point associated point pairs (IPP) to generate the joint associated point pair loss function. Specifically, it includes: S331. Construct the loss function for boundary associated point pairs (BPP). The loss of boundary associated point pairs (BPP) is the geometric continuity loss. Let any boundary associated point pair be , and the normal vectors are respectively . The loss function of boundary associated point pairs (BPP) is: ; where, The embedded features after mapping the boundary-associated point pairs through the MLP are called boundary-associated point pair embedded features. The weight for balancing the consistency between the normal vector and the boundary-associated point pair embedded features Represents the total number of boundary-associated point pairs, that is, the number of all boundary-associated point pairs participating in the calculation ; S332. Construction of the inlier-associated point pair (IPP) loss function. The inlier-associated point pair (IPP) loss is the feature consistency loss. Let any inlier-associated pair (IPP) be , and the inlier-associated point pair (IPP) loss function The formula is: ; Among them, Represents the total number of inlier-associated point pairs, that is, the number of all inlier-associated point pairs (IPP) participating in the calculation , Represents the pigment value or voxel value of the inlier-associated point pair, used to reflect its local color or density Represents the embedded features of the inlier-associated pair (IPP) Is a hyperparameter , adjusting the weight between the color and density difference term and the embedded feature difference term of the inlier-associated point pair (IPP); S333. Construction of the associated point pair loss function. The associated point pair loss function is the overall loss function, combining the boundary-associated point pair (BPP) loss function and the inlier-associated point pair (IPP) loss function. The associated point pair loss function L The formula is as follows: ; Among them, R Is the reward value of the self-masking strategy, dynamically adjusted according to the correct classification or recognition degree of the masked category area. The larger the reward value, the more accurate the correct classification or recognition degree of the masked category area, defined as: ; Among them, Is the true mask, Is the predicted mask, Is the indicator function, taking 1 when the predicted mask is correct, otherwise 0; Is a hyperparameter , and the value of the hyperparameter is obtained through cross-validation. Using 5-fold cross-validation, for each combination of hyperparameters in each fold of cross-validation Perform performance scoring and sorting, and select the one with the highest score to balance the contributions of the geometric continuity loss of the boundary pair points (BPP), the feature consistency loss of the inlier pair points (IPP), and the reward function of the mask category accuracy.

[0081] Further, in the joint disc modeling method based on the structure mask and the contrast learning mechanism in the above embodiments, cross-validation is performed.

[0082] S340. Screening of the boundary point set. Using the high-dimensional embedding features and the pair point loss function, screen the joint disc boundary point set to obtain an enhanced constrained joint disc boundary point set and pair point correlation learning to strengthen the boundary; specifically including: S341. Cosine similarity calculation. For each pair of associated points P (i,j) , including the boundary pair points BPP and the inlier pair points IPP, calculate the feature consistency score using the high-dimensional embedding features, that is, use the cosine similarity to evaluate whether the pair of points is in a consistent semantic relationship in the feature space. The cosine similarity s ij The formula is as follows: ; Among them, s ij The closer the i , j represents the numbers of the two associated points in the pair of associated points, the higher the similarity degree of the pair of associated points in the semantic relationship, P (i,j) The closer the s ij value is to 0, the less semantic association the pair of points has. If s ij the value is less than 0, it means that the semantic relationship between the two points is incorrect and may come from different categories; S342. Comprehensive confidence calculation. For each pair of associated points P (i,j) calculate the comprehensive confidence , the formula is as follows: ; Among them, represents the number of pairs of associated points, and the range is [0,1]; S343. Screening of the enhanced constrained joint disc boundary point set. Based on the cosine similarity and the comprehensive confidence, screen the joint disc boundary point set, and select the boundary points with the cosine similarity greater than or equal to the cosine similarity threshold of 0.9 and the comprehensive confidence greater than or equal to the comprehensive confidence threshold of 0.8 to obtain an enhanced constrained joint disc boundary point set and pair point correlation learning to strengthen the boundary.

[0083] S400, Boundary enhancement contrast learning, define the center of the center line of the articular disc, and divide the region of interest ROI For the left half region and the right half region , construct the area difference loss function, construct the contrast learning loss function, calculate the predicted boundary of the articular disc, construct the predicted boundary loss function of the articular disc, calculate the set of symmetric boundary points of the center line of the articular disc and the symmetric boundary of the center line of the articular disc, and complete the initial construction of the articular disc model; specifically including: S410, Definition of the center of the center line of the articular disc, and the set of boundary points of the articular disc is expressed as: ; Based on y the selection of the pair of points with the maximum spacing edge on the axis, select from all boundary points y points with approximately the same coordinates, that is, the difference is less than a predetermined threshold of the point pair, x the point pair with the largest coordinate difference as the left and right limit points, and the predetermined threshold is normalized according to the actual resolution of the image, specifically the product of the vertical resolution and 0.01: ; Among them, , are the left and right limit points respectively, and the abscissas are the smallest abscissa and the largest abscissa among the points with approximately the same ordinate (being ) in all boundary points; S420, Region permutation and area difference construction, define the region of interest ROI , take the center line of the articular disc as the axis of symmetry, and divide all boundary points and the mask region into the left half region and the right half region , perform region permutation operations through the region permutation function, and construct the area difference loss function; specifically including: S421, Define the region of interest, construct a rectangle covering the main boundary region, that is, the region of interest ROI, The formula is as follows: ; Among them, is the adjustment parameter, controlling the expansion range of the region of interest ROI in the y axis direction; S422, Define the region permutation function, and the formula of the region permutation function is as follows: ; Among them, That is, a mirror sample is constructed by reflecting through the center line of the articular disc, is the point The abscissa after mirroring along the center line of the articular disc; is the position of the center line of the articular disc, that is, the left half and the right half of the dividing line; S423. Construct an area difference loss function, using the boundary point mask area as the basis for area calculation. Let the original boundary mask area be , and the mask area after region replacement be . Define the area difference loss function as: ; The area difference loss function reflects the symmetry of the articular disc structure by comparing the spatial coincidence degree between the original boundary mask area and the mask area after region replacement. The larger the value of the area difference loss function , the greater the difference between the original structure in the left and right mirror states, and the worse the symmetry; S430. Construction of a contrast learning loss function. Extract the high-dimensional embedded features of the boundary points, define positive and negative sample pairs, and construct a contrast learning loss function based on cosine similarity; Specifically include: S431. Extraction of high-dimensional embedded features of boundary points. The formula for extracting high-dimensional embedded features of boundary points is as follows: ; Among them, p is the boundary point; S432. Definition of positive and negative sample pairs. Geometric transformation of the original boundary mask area gives positive samples. Geometric transformation includes rotation and translation, which are used to train the model to enhance boundary invariance and construct reasonable boundary variants close to the real structure as similar reference samples for the boundary structure. Mirror transformation of the original boundary mask area gives negative samples, which are used to identify asymmetry and improve the model's discrimination ability for inconsistent left and right structures, especially for identifying lesions or irregular morphologies. The original boundary mask area and the positive sample constitute a positive sample pair, and the original boundary mask area and the negative sample constitute a negative sample pair; S433. Definition of contrast learning loss. Construct a contrast learning loss function , defined as follows: ; Among them, represents the cosine similarity, is the temperature constant, is the feature representation of the original boundary mask area, is the feature representation of the positive samples, is the feature representation of the negative samples; S440. Boundary reconstruction and weighted fusion. Calculate the predicted boundary of the articular disc, construct the loss function of the predicted boundary of the articular disc, calculate the set of symmetric boundary points on the center line of the articular disc and the symmetric boundary on the center line of the articular disc, and complete the initial construction of the articular disc model. Specifically, it includes: S441. Boundary weighted fusion. Perform weighted fusion on the predicted boundary, the boundary strengthened by associated point pair association learning, and the symmetric boundary on the center line of the articular disc to obtain the predicted boundary of the articular disc as follows: ; ; wherein, is the predicted boundary, emphasizing the boundary position perception ability; is the boundary strengthened by associated learning, emphasizing the perception of associated point pairs and improving the local geometric continuity; is the symmetric boundary on the center line of the articular disc, considering the symmetry of the articular disc edge and the area reconstruction difference, emphasizing the repair of fuzzy areas or missing parts and reflecting the boundary complementation ability; are respectively weighted coefficients of, as learnable parameters, and perform normalization on through the gradient descent Softmax function to ensure that the following constraints are met: ; wherein, ; S442. Construction of the loss function of the predicted boundary of the articular disc. Construct the loss function of the predicted boundary of the articular disc as follows: ; wherein, are respectively the hyperparameters of the area difference loss function and the hyperparameters of the contrast learning loss function; S443. Calculate the set of symmetric boundary points on the center line of the articular disc and the symmetric boundary on the center line of the articular disc. For any point in the set of articular disc boundary points, obtain the point mapped through the center line of the articular disc, and construct the set of symmetric point pairs, which is specifically expressed as: ; The set of symmetric boundary points on the center line of the articular disc is expressed as: ; That is, the set of articular disc boundary points The union with the set of symmetric point pairs ; Using the curve reconstruction method α-Shape algorithm to perform boundary fitting on the set of symmetric boundary points of the joint disc center line is specifically expressed as: ; where Rec(.) represents using the curve reconstruction method to perform boundary fitting on the set of symmetric boundary points of the joint disc center line to obtain the symmetric boundary of the joint disc center line.

[0084] S500. Training of the joint disc model, constructing the overall loss function of the joint disc model, training the joint disc model through the Adam optimizer, incrementing the number of training epochs by one, and when it is determined that the number of training epochs is less than the maximum number of epochs, returning to step S200. When the number of training epochs is greater than or equal to the maximum number of epochs, ending the training of the joint disc model; specifically including: S510. Maintaining boundary structure symmetry. To ensure the continuous reinforcement of the joint disc structure symmetry, the consistency loss function formula is defined as follows: ; where is the number of boundary points, is the feature extraction function, is the boundary point, is the symmetric reflection point, that is, the joint disc model learns the structural symmetry attribute of the boundary point; S520. Definition of the overall loss function. The overall loss function formula of the joint disc model is as follows: ; where is the weight coefficient, ; S530. Training of the joint disc model. Training the joint disc model through the Adam optimizer, incrementing the number of training epochs by one, and when it is determined that the number of training epochs is less than the maximum number of training epochs T max = 2000, returning to step S200. When the number of training epochs is greater than or equal to the maximum number of epochs, ending the training of the joint disc model. The joint disc model is represented as follows: .

[0085] S600. Output of the joint disc model. Evaluating the joint disc model using model evaluation metrics. When the joint disc model meets the model evaluation metrics, saving the joint disc model. Otherwise, updating the initial value of the mask parameter according to the initial mask update strategy, setting the number of training epochs to 0, and returning to step S200; the model evaluation metrics include boundary accuracy evaluation metrics and structural symmetry evaluation metrics; specifically including: S610. Predictive boundary accuracy evaluation. Evaluate the accuracy of the predicted joint disc boundary of the joint disc model using boundary accuracy evaluation metrics, and evaluate the coincidence degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc. The boundary accuracy evaluation metrics include: Dice coefficient, with the formula as follows: ; Among them, P is the predicted boundary area of the joint disc, G is the true boundary area of the joint disc. The Dice coefficient reflects the overlapping degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc. The value range is [0, 1]. The closer the Dice coefficient value is to 1, the more consistent the predicted boundary area of the joint disc is with the true boundary area of the joint disc, the higher the coincidence degree, and the more accurate the joint disc model. The closer the Dice value is to 0, it indicates that the coincidence degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc is lower, indicating that the construction of the joint disc model fails; IoU (Jaccard), with the formula as follows: ; IoU (Jaccard) measures the overlapping degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc by calculating the intersection-over-union ratio of the predicted boundary area of the joint disc and the true boundary area of the joint disc, so as to reflect whether the joint disc model is overfitting or underfitting. IoU (Jaccard) has a value range of [0, 1]. IoU (Jaccard) The closer the value is to 1, the higher the coincidence degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc, and the more accurate the joint disc model. IoU (Jaccard) The closer the value is to 0, it indicates that the complete coincidence degree between the predicted boundary area of the joint disc and the true boundary area of the joint disc is lower, indicating that the construction of the joint disc model fails; S620. Predictive structural symmetry evaluation. Evaluate the symmetry of the predicted joint disc structure of the joint disc model using structural symmetry evaluation metrics. The structural symmetry evaluation metrics use the structural symmetry index (Symmetry Consistency Score, SCS) to quantify the error of the left-right symmetric structure. The formula is as follows: ; SCS Measures the left-right structural consistency of the predicted boundary structure of the joint disc model with respect to the axis. The predicted boundary point and its symmetric mapping point with respect to the center line of the joint disc. The average spatial deviation between them. When the predicted boundary structure is highly symmetric, SCS the value approaches 0. When the predicted boundary structure produces contour deformation,SCS The larger the value; S630, joint disc model evaluation and output. When the joint disc model meets the boundary accuracy evaluation index and the structural symmetry evaluation index, save and output the joint disc model. Otherwise, update the initial value of the mask parameter according to the initial mask update strategy, set the number of training rounds to 0, and the initial mask update strategy is: combining the evaluation indexes, when the Dice coefficient value is greater than the Dice threshold of 0.85 or IoU (Jaccard) is less than IoU the threshold of 0.75, the initial radius increases by a specified radius increment of 0.02, and the initial rotation angle increases by a specified angle increment of 2; when SCS the value is greater than SCS the threshold of 5%, the Gaussian parameter increases by a specified parameter increment of 1, and return to step S200.

[0086] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations according to the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning or limited experiments shall fall within the protection scope determined by the claims.

Claims

1. A method for modeling the articular disc based on a structural mask and a contrast learning mechanism, characterized in that It includes the following steps: S100, Mask generation and mask parameter initialization. Input the articular disc image into the conventional segmentation network, extract the center line of the articular disc, generate straight tube masks, bent tube masks, and Gaussian block masks, set the initial values of the mask parameters, and initialize the number of training rounds to 0; S200. Mask parameter optimization: Optimize the mask parameters based on reinforcement learning, perform mask prediction to obtain the optimal mask parameters, and save the set of joint disc boundary points; S300. Association loss function determination: Construct association point pairs, including boundary association point pairs and inlier association point pairs, learn high-dimensional embedding features, determine the joint association loss function, screen and optimize the set of joint disc boundary points to obtain the association point pair association learning enhanced boundary; S400, Boundary-enhanced contrastive learning, define the center of the center line of the articular disc, and divide the region of interest ROI For the left half region and the right half region , construct the area difference loss function, construct the contrastive learning loss function, calculate the predicted boundary of the articular disc, construct the predicted boundary loss function of the articular disc, calculate the set of symmetric boundary points of the center line of the articular disc and the symmetric boundary of the center line of the articular disc, and complete the initial construction of the articular disc model; S500. Joint disc model training: Construct the overall loss function of the joint disc model, train the joint disc model through the Adam optimizer, increment the number of training rounds, and when it is determined that the number of training rounds is less than the maximum number of rounds, return to step S200; when the number of training rounds is greater than or equal to the maximum number of rounds, end the joint disc model training; S600. Joint disc model output: Evaluate the joint disc model using model evaluation metrics. When the joint disc model meets the model evaluation metrics, save and output the joint disc model; otherwise, update the initial value of the mask parameters according to the initial mask update strategy, set the number of training rounds to 0, and return to step S200.

2. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 1, wherein The step S100 includes: S110. Extract the center line of the articular disc. Input the articular disc image into the conventional segmentation network to obtain the articular disc region mask and the set of articular disc boundary points. Use the skeleton extraction algorithm for the articular disc region mask to obtain the center line of the articular disc C ; S120. Mask generation, based on the center line of the articular disc C and the set of boundary points of the articular disc to generate masks, including straight pipe masks, bent pipe masks and Gaussian block masks, and generate corresponding mask regions; S130. Initialize the mask parameters, set the initial values of the mask parameters, where the mask parameters include a radius of , a rotation angle of , and Gaussian parameters . The initial radius is 10 mm, the initial rotation angle is 0, and the initial Gaussian parameters are (10, 10). Apply the initial values of the mask parameters to the joint disc image, and perform local enhancement processing on the joint disc area to improve feature representation.

3. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 2, wherein The step S120 includes: S121. Straight tube mask generation. The straight tube mask is used for the non-deformed area. With the center line of the articular disc as the spindle, a cylindrical area with a radius of C is generated along the natural direction of the articular disc. Let the center line of the articular disc be , and the direction is . Suppose the center line of the articular disc is , and the coordinate of each point is . Then the straight tube mask is defined as: ； S122, Bent pipe mask generation, the bent pipe mask is used for the mild to moderate deformation area, with the center line of the articular disc C rotated as the new main axis, and the rotation angle is set as , and the new main axis function is denoted as , and the deflection curve formula is: ； The said The elbow mask is defined as: ； S123. Generation of Gaussian block mask. The Gaussian block mask is used for large deformation regions. Perform local PCA on the set of boundary points of the articular disc to extract the principal curvature direction. , for the deformed region of the articular disc, generate an ellipsoidal mask based on boundary fitting and Gaussian distribution, that is, the Gaussian block mask. At the point , fit an ellipse. The mask intensity follows a two-dimensional Gaussian distribution, representing the response weight of each pixel point in the two-dimensional plane to the final structure mask. Its value is between [0, 1]. The value at the center point is 1, and it approaches 0 closer to the edge. The Gaussian block mask is defined as: 。 4. The method for modeling the articular disc based on the structure mask and the contrastive learning mechanism according to claim 3, wherein The step S200 includes: S210. Mask parameter adjustment. Adjust the mask parameters according to the adjustment strategy of the mask. The straight pipe mask reduces the radius at a fixed step to refine the central area until the minimum radius; the bent pipe mask increases the rotation angle at a fixed angle until the maximum angle; the Gaussian block mask is enhanced with a fixed ellipsoidal flattening to strengthen the boundary envelope until the maximum change in the ellipsoidal flattening; S220. Mask parameter optimization: optimize the mask parameters based on reinforcement learning, perform mask prediction to obtain the joint disc boundary point set and the predicted joint disc boundary. When the boundary IoU is lower than the set boundary IoU threshold, return to step S210 until the continuous N boundary IoU improvement rate is lower than the minimum change amount or the maximum number of iterations is reached, complete the mask prediction, obtain the optimal mask parameters, and save the joint disc boundary point set.

5. The method for modeling the articular disc based on the structural mask and the contrastive learning mechanism according to claim 4, wherein The step S220 includes: S221. State space design: The mask selection and the mask parameter setting constitute the current state, denoted as: ； S222. Action space design, selecting an action according to the current state, including switching masks T and adjusting the mask parameters, including , and the action space is defined as: ； S223. Reward function design, defining the reward function as the detection of the boundary points of the articular disc Improvement amount and the overall detection accuracy of the articular disc Improvement amount of the weighted combination, the said reward function The formula is as follows: ； S224. Mask parameter optimization: Use the Actor-Critic framework based on policy gradient to optimize the mask parameters, complete mask prediction to obtain the optimal mask parameters, and save the set of joint disc boundary points. The policy gradient formula is as follows: 。 6. The method for modeling the articular disc based on the structure mask and the contrastive learning mechanism according to claim 5, wherein The step S300 includes: S310. Association point pair construction. Divide different regions of the articular disc formed by the straight pipe mask, the bent pipe mask, and the Gaussian block mask, assign a spatial role to each point to form association point pairs, and classify the association point pairs into boundary association point pairs and interior point association point pairs; S320. High-dimensional embedding feature learning: Use a multi-layer perceptron (MLP) embedding network to map the association point pairs and perform high-dimensional embedding feature learning; S330. Association point pair loss function construction: Construct the boundary association point pair loss function and the inlier association point pair loss function, and synthesize the boundary association point pair loss function and the inlier association point pair loss function to generate the joint association point pair loss function; S340. Boundary point set screening: Use the high-dimensional embedding features and the association point pair loss function to screen the set of joint disc boundary points to obtain the enhanced constrained set of joint disc boundary points and the association point pair association learning enhanced boundary.

7. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 6, wherein The step S330 includes: S331. Construction of the boundary-associated point pair loss function. The boundary-associated point pair loss is the geometric continuity loss. Let any boundary-associated point pair be , with normal vectors respectively. The boundary-associated point pair loss function has the formula: ； S332. Construction of the inlier paired points (IPP) loss function. The inlier paired points loss is the feature consistency loss. Let any inlier pair be . The inlier paired points loss function has the formula: ； S333. Construction of the associated point pair loss function. The associated point pair loss function is the overall loss function, which combines the boundary associated point pair loss function and the inlier associated point pair loss function. The associated point pair loss function L is as follows: 。 8. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 7, wherein The step S400 includes: S410. Definition of the center of the joint disc center line: The set of joint disc boundary points is represented as: ； Based on y the selection of the pair of edge points with the maximum spacing along the axis, select among all boundary points y points with approximately the same coordinates, i.e., the difference is less than a predetermined threshold as the pair of points, x the pair of points with the largest coordinate difference as the left and right limit points: ； S420. Region replacement and area difference construction to define the region of interest ROI Using the center line of the articular disc as the axis of symmetry, divide all boundary points and the masked area into the left half and the right half Perform region replacement operations through the region replacement function to construct an area difference loss function; S430. Contrastive learning loss function construction: Extract the high-dimensional embedding features of the boundary points, define positive sample pairs and negative sample pairs, and construct a contrastive learning loss function based on cosine similarity; S440. Boundary reconstruction and weighted fusion: Calculate the predicted boundary of the joint disc, construct the predicted boundary loss function of the joint disc, calculate the set of symmetric boundary points of the joint disc center line and the symmetric boundary of the joint disc center line, and complete the initial construction of the joint disc model.

9. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 7 or 8, characterized in that The step S500 includes: S510. The symmetry of the boundary structure is maintained. To ensure the continuous enhancement of the symmetry of the articular disc structure, the formula for the consistency loss function is defined as follows: ； S520. Definition of the overall loss function. The formula for the overall loss function of the articular disc model is as follows: ； S530. Training of the articular disc model. The articular disc model is trained using the Adam optimizer. The number of training epochs is incremented by one. When it is determined that the number of training epochs is less than the maximum number of epochs, the process returns to step S200. When the number of training epochs is greater than or equal to the maximum number of epochs, the training of the articular disc model ends. The articular disc model is represented as follows: 。 10. The method for modeling the articular disc based on the structure mask and the contrast learning mechanism according to claim 9, wherein The step S600 includes: S610. Evaluation of the prediction boundary accuracy. The accuracy of the articular disc boundary predicted by the articular disc model is evaluated using the boundary accuracy evaluation metric, and the overlap between the predicted boundary region of the articular disc and the true boundary region of the articular disc is evaluated; S620. Evaluation of the prediction structure symmetry. The symmetry of the articular disc structure predicted by the articular disc model is evaluated using the structure symmetry evaluation metric; S630. Evaluation and output of the articular disc model. When the articular disc model meets the boundary accuracy evaluation metric and the structure symmetry evaluation metric, the articular disc model is saved and output. Otherwise, the initial value of the mask parameter is updated according to the initial mask update strategy, the number of training epochs is set to 0, and the process returns to step S200.

Images