A joint disc modeling method based on structure masking and contrastive learning mechanism

Abstract

The present invention discloses a method for modeling an articular disc based on a structural mask and contrastive learning mechanism, which relates to the fields of artificial intelligence and medical image processing technology. The method comprises the following steps: S100, mask generation and mask parameter initialization; S200, mask parameter optimization; S300, association loss function determination; S400, boundary-enhanced contrastive learning; S500, articular disc model training; and S600, articular disc model output. The articular disc model constructed by the present invention solves the technical problems of blurred boundaries, incomplete morphological reconstruction, poor boundary point connectivity, and lack of structural symmetry that exist in the articular disc segmentation and reconstruction process.

Description

Technical Field

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

[0002] The articular disc, a key soft tissue structure in the temporomandibular joint (TMJ) system, lies between the mandibular condyle and the temporal bone. Its primary functions are to buffer mechanical loads, guide mandibular movement, and maintain joint stability. The integrity, morphology, and movement trajectory of the articular disc significantly impact the proper functioning of oral functions, including chewing, swallowing, and speech articulation.

[0003] Clinically, pathological changes in the disc are closely associated with various temporomandibular joint disorders (TMDs), including but not limited to disc anterior and posterior displacement, perforation, atrophy, and abnormal disc dislocation and reduction. Failure to promptly and accurately diagnose and intervene can lead to serious consequences such as joint clicking, pain, limited mouth opening, and even bone and joint degeneration. Therefore, morphological assessment and dynamic monitoring of the disc play an irreplaceable role in the imaging diagnosis and treatment planning of temporomandibular joint disorders.

[0004] Magnetic resonance imaging (MRI) is considered the "gold standard" for evaluating the position and morphology of articular discs, especially for evaluating their dynamic changes under different mouth opening states. However, current artificial intelligence image analysis technologies are mainly concentrated in the fields of lung CT images, breast cancer screening, cardiac MRI and brain MRI images. Research on MRI image analysis of the temporomandibular joint, especially articular disc anterior displacement, is still in its infancy. Due to the small size and complex structure of the articular disc, and the fact that lesions may be accompanied by significant deformation and signal blurring, the performance of traditional image segmentation models on this type of target has significantly decreased. In addition, the morphology of the articular discs of different patients varies greatly, which makes the traditional diagnostic process that relies on experience-based judgment prone to misjudgment when faced with complex or atypical cases, especially for young doctors with insufficient clinical experience. Traditional methods also include: doctors make a preliminary judgment by asking patients about their symptoms, such as pain in the preauricular area, joint noises, difficulty chewing, limited mouth opening, etc., combined with facial palpation and mouth opening observation. This is the most basic and common method; it includes mouth opening measurement (maximum mouth opening distance), deviation degree assessment, muscle tenderness point examination, and comprehensive analysis and functional assessment combined with the occlusal situation; it is used to observe the structure of the hard tissue of the temporomandibular joint, such as the position of the condyle and the joint space, but has poor display ability for soft tissue (especially the articular disc).

[0005] The main technical challenges faced by existing artificial intelligence image analysis of articular discs are:

[0006] 1. Highly fuzzy boundaries and low contrast. Because the grayscale contrast between the articular disc and surrounding tissues (such as the joint capsule and synovial cavity) in MRI images is extremely close, and is affected by factors such as imaging layer 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 segment using traditional image processing methods.

[0007] 2. Large structural variability. The anatomical differences between different patients are significant, and the morphological changes of the articular disc during opening and closing are complex, making it difficult for the model to capture its stable geometric features. Especially in pathological conditions, the articular disc morphology exhibits irregular changes such as atrophy, fracture, and displacement, which increases the complexity of segmentation and modeling.

[0008] 3. Lack of spatial structure modeling and prior guidance mechanisms. Although existing deep learning-based automatic segmentation methods have made some progress, most models ignore the anatomical direction and motion trajectory of the articular disc in three-dimensional space and fail to incorporate structural priors for constraints. This leads to morphological loss, false positive enhancement, or fractures in lesions or boundary areas.

[0009] 4. Weak point-to-point connectivity and poor boundary continuity. Especially in boundary point detection tasks, models trained solely on pixel- or point-level labels cannot effectively capture the geometric relationships, semantic associations, and structural coherence between points, making it difficult to achieve boundary integrity modeling.

[0010] 5. Lack of symmetry and geometric balance in modeling. The temporomandibular joint has bilaterally symmetrical anatomical features, but current models generally lack symmetry constraints. This results in asymmetric bias or geometric distortion in model output results in some severely affected areas, seriously affecting subsequent 3D reconstruction and clinical judgment.

[0011] In summary, existing technologies cannot meet the requirements of accurate identification, continuous boundary restoration, and structural integrity reconstruction of articular discs in complex clinical images, and lack a complete mechanism that integrates structure guidance, semantic enhancement, and boundary optimization.

[0012] Therefore, technicians in this field are committed to developing an articular disc modeling method based on structure masking and contrastive learning mechanism. Summary of the Invention

[0013] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is the blurred boundaries, incomplete morphological reconstruction, poor connectivity of boundary points and lack of structural symmetry in the process of articular disc segmentation and reconstruction.

[0014] This application introduces a structured mask mechanism, combines it with a point-pair semantic association loss function for deep supervision, and uses the symmetrical replacement area of ​​the articular disc centerline to construct a boundary contrast loss, thereby enhancing the global consistency of the articular disc morphology in the edge and central axis areas, solving the problems of boundary blur, structural occlusion, and boundary discontinuity, and significantly improving the reconstruction accuracy of boundary points while maintaining structural integrity.

[0015] Specifically, based on the structured mask mechanism, the straight tube mask, The curved tube mask and Gaussian block mask structures establish spatial priors in the articular disc area, and use reinforcement learning to dynamically optimize the mask, angle, and curvature parameters; construct an associated loss function for the joint of boundary points and internal points, and use the perceptual scale graph structure and multi-layer perceptron (MLP) to perform local pigment contrast, enhance edge geometric continuity, establish global dependencies between point pairs, and improve the ability to distinguish the edge and internal structure of articular disc tissue; based on the boundary reinforcement contrast learning strategy of symmetrical displacement of the articular disc centerline, the boundary point articular disc centerline definition and symmetrical area displacement are used to construct a geometric enhancement loss term, and adjustable parameters are introduced to achieve maximum overlap of boundary points and maximize structural differences, further enhancing the characteristic distinction of edge points and reconstruction of structural integrity.

[0016] In one embodiment of the present invention, a method for modeling an articular disc based on a structure mask and contrastive learning mechanism is provided, comprising the following steps:

[0017] S100, mask generation and mask parameter initialization, input the articular disc image into the conventional segmentation network, extract the articular disc centerline, generate the straight tube mask, For the elbow mask and Gaussian block mask, set the initial values ​​of the mask parameters and initialize the number of training rounds to 0;

[0018] S200, mask parameter optimization, optimizing mask parameters based on reinforcement learning, performing mask prediction, obtaining optimal mask parameters, and saving a set of articular disc boundary points;

[0019] S300, determining the association loss function, constructing association point pairs, including boundary association point pairs and interior point association point pairs, learning high-dimensional embedding features, determining the joint association loss function, screening and optimizing the set of articular disc boundary points, and obtaining association point pairs for association learning to strengthen boundaries;

[0020] S400, boundary enhancement contrast learning, defining the center of the articular disc centerline and dividing the region of interest ROI For the left half and right half , construct the area difference loss function, construct the contrastive learning loss function, calculate the articular disc prediction boundary, construct the articular disc prediction boundary loss function, calculate the articular disc centerline symmetric boundary point set and the articular disc centerline symmetric boundary, and complete the initial construction of the articular disc model;

[0021] S500, articular disc model training, constructing an overall loss function for the articular disc model, training the articular disc model using the Adam optimizer, increasing the number of training rounds by one, and returning to step S200 when the number of training rounds is greater than or equal to the maximum number of rounds, ending articular disc model training;

[0022] S600, output the articular disc model, use the model evaluation index to evaluate the articular disc model, save and output the articular disc model when the articular disc model meets the model evaluation index, 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.

[0023] Optionally, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the mask includes:

[0024] Straight tube mask, with the center line of the articular disc as the main axis, generates a radius of r , the direction is The cylindrical area of ​​​​the straight tube mask is obtained;

[0025] Bend mask, rotated about the disc centerline As the new main axis, the rotation angle is , the area of ​​the deflection curve around the new principal axis is elbow mask;

[0026] Gaussian block mask, fits an elliptical area at the boundary points of the articular disc to obtain an ellipsoidal mask based on boundary fitting and Gaussian distribution, namely, the Gaussian block mask.

[0027] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S100 includes:

[0028] S110, joint disc centerline extraction, input the joint disc image into the conventional segmentation network, obtain the joint disc area mask and the joint disc boundary point set, and use the skeleton extraction algorithm to obtain the joint disc centerline based on the joint disc area mask C ;

[0029] S120, mask generation, based on the centerline of the articular disc C and the articular disc boundary point set to generate masks, including straight tube masks, Bend pipe mask and Gaussian block mask, and generate corresponding mask areas;

[0030] S130, initialization of mask parameters, setting the initial values ​​of mask parameters, including the radius , the rotation angle is , and Gaussian parameters , initial radius 10mm, initial rotation angle is 0, the initial Gaussian parameter The initial value of the mask parameter is (10, 10), and the articular disc image is applied to perform local enhancement on the articular disc area to improve feature representation.

[0031] Optionally, in the articular disc modeling method based on structure mask and contrastive learning mechanism in any of the above embodiments, the conventional segmentation network uses U-Net.

[0032] Optionally, in the articular disc modeling method based on structure mask and contrastive learning mechanism in any of the above embodiments, the skeleton extraction algorithm includes the Zhang-Suen thinning algorithm and the Medial Axis Transform algorithm.

[0033] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the centerline of the articular disc C Represented as a set of discrete points:

[0034] ;

[0035] in, is the coordinate of the discrete point 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, which is the number of discrete points.

[0036] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, step S120 includes:

[0037] S121, straight tube mask generation, straight tube mask is used for non-deformation area, with the center line of the articular disc C As the main axis, the radius is generated along the natural direction of the articular disc (Y axis) , the direction is The cylindrical area; let the center line of the articular disc be , the coordinates of each point are, then the straight tube mask Defined as:

[0038] ;

[0039] in Represents a point in two-dimensional space A Boolean value indicating whether it belongs to the straight pipe mask area, with a value of 1 or 0. is the center and the radius is The mask value of the area within and is 1, and the others are 0;

[0040] S122, Bend mask generation, The elbow mask is used for areas with mild to moderate deformation, with the centerline of the articular disc as the centerline. C Rotation As the new principal axis, let the rotation angle be , the new principal axis function is denoted as , the deflection curve formula is:

[0041] ;

[0042] The elbow mask is defined as:

[0043] ;

[0044] in Represents a point in two-dimensional space Whether it belongs to The Boolean value of the elbow mask area, which can be 1 or 0. is the point on the new principal axis, As the radius, is the center and the radius is The mask value of the area within and is 1, and the others are 0;

[0045] S123, Gaussian block mask generation, Gaussian block mask is used for large deformation area, local PCA (‌Principal Component Analysis, principal component analysis) is performed on the set of articular disc boundary points to extract the main curvature direction , for the deformed area of ​​the articular disc, an ellipsoidal mask based on boundary fitting and Gaussian distribution, namely, a Gaussian block mask, is generated at point The mask strength obeys the two-dimensional Gaussian distribution, which represents the response weight of each pixel point on the two-dimensional plane to the final structure mask. Its value is between [0, 1], the center point value is 1, and the closer to the edge, the closer to 0. The Gaussian block mask Defined as:

[0046] ;

[0047] in, is the standard deviation of the Gaussian distribution in the horizontal direction, which is used to control the expansion range of the fitting ellipse in the lateral edge deformation direction, ranging from 1.0 to 3.0 mm; is the standard deviation of the Gaussian distribution in the vertical direction, which is used to control the expansion range of the fitting ellipse in the longitudinal boundary extension direction, ranging from 0.5 to 2.0 mm.

[0048] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the non-deformation area is the area where the center line or boundary curvature of the articular disc has no obvious change or the deformation area accounts for a small proportion, specifically: the boundary curvature The range is ; or the offset distance is less than 1mm; or the deformation area accounts for less than 10% of the total structural area of ​​the articular disc.

[0049] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the mild to moderate deformation area is the area where the center line or boundary curvature of the articular disc changes smoothly or the deformation area accounts for a medium proportion, specifically: the boundary curvature The range is ; or the offset distance is in the interval [1mm, 5mm]; or the deformation area accounts for the total structural area of ​​the articular disc in the interval [10%, 40%].

[0050] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the large deformation area is the area where the center line or boundary curvature of the articular disc changes drastically or the deformation area accounts for a large proportion, 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.

[0051] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, step S200 includes:

[0052] S210, adjusting mask parameters: adjusting mask parameters according to the mask adjustment strategy, reducing the radius of the straight pipe mask with a fixed step size, and refining the central area until the minimum radius is reached; The elbow mask increases the rotation angle by a fixed angle until the maximum angle; the Gaussian block mask increases the ellipsoidal flattening by a fixed angle, strengthening the boundary envelope until the ellipsoidal flattening changes to the maximum;

[0053] S220, mask parameter optimization, based on reinforcement learning to optimize the mask parameters, perform mask prediction, obtain the articular disc boundary point set and the articular disc prediction boundary, when the boundary IoU (Intersection over Union) is lower than the set limit IoU When the threshold is reached, return to step S210 until the number of consecutive N sub-border IoUWhen the improvement rate is lower than the minimum change or reaches the maximum number of iterations, the mask prediction is completed, the optimal mask parameters are obtained, and the set of articular disc boundary points is saved.

[0054] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the fixed step size is 0.02 mm.

[0055] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the minimum radius is 5 mm.

[0056] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the fixed angle is 1 degree.

[0057] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the maximum angle is 90 degrees.

[0058] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the ellipsoid flattening is fixed to 0.2.

[0059] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the maximum change of the ellipsoid flattening is 2.

[0060] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the boundary IoU The threshold is 0.7.

[0061] Optionally, in the articular disc modeling method based on structure mask and contrast learning mechanism in any of the above embodiments, N The range is greater than or equal to 3 and less than or equal to 5.

[0062] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, N =3.

[0063] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the minimum change amount is set to 0.001.

[0064] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the maximum number of iterations=500.

[0065] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, step S220 includes:

[0066] S221, state space design, mask selection and mask parameter setting constitute the current state, denoted as:

[0067] ,

[0068] in, Indicates the mask, that is, the straight pipe mask, elbow mask and Gaussian block mask, Indicates the spindle rotation angle, used for elbow mask, 、 They 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 rate of change of boundary curvature;

[0069] S222, action space design, select actions according to the current state, including switching masks T and adjust mask parameters, including , the action space is defined as:

[0070] ;

[0071] S223, reward function design, define the reward function as the joint disc boundary point detection Lift and overall detection accuracy of the articular disc Lift The reward function is a weighted combination of The formula is as follows:

[0072] ;

[0073] in, 、 is the empirical weight coefficient, ,and , preferably ;

[0074] S224, mask parameter optimization, using the Actor-Critic framework based on policy gradient to optimize mask parameters, complete mask prediction, obtain the optimal mask parameters, and save the set of articular disc boundary points. The policy gradient formula is as follows:

[0075] ;

[0076] in, represents the state and action distribution sampled from the current policy, Indicates that in the current state s, according to the parameter The policy network generates actions from the action space a The probability distribution of Indicates action a Log-probability gradient under the current policy; is the advantage function, which measures the execution of actions a How much better than the average policy, the advantage function is defined as:

[0077]

[0078] in, is the discount factor, preferably 0.95, to control the impact of future state value; To perform an action a After that, for the next state Estimated value of is the estimated value of the current state s.

[0079] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S300 includes:

[0080] S310, build the associated point pair, the straight pipe mask, The articular disc is divided into different regions formed by the elbow mask and Gaussian block mask, and a spatial role is assigned to each point to form associated point pairs. The associated point pairs are then classified into boundary-paired points (BPP) and interior-paired points (IPP).

[0081] S320, high-dimensional embedding feature learning, uses a multi-layer perceptron (MLP) embedding network to map associated point pairs and perform high-dimensional embedding feature learning;

[0082] S330, constructing a pairwise loss function, constructing a boundary-paired points (BPP) loss function and an interior-paired points (IPP) loss function, and combining the boundary-paired points (BPP) loss function and the interior-paired points (IPP) loss function to generate a joint pairwise loss function;

[0083] S340, boundary point set screening, using high-dimensional embedding features and associated point pair loss function to screen the articular disc boundary point set, and obtain the articular disc boundary point set with enhanced constraints and the associated point pair association learning to strengthen the boundary.

[0084] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the spatial roles include boundary points, interior points and boundary-adjacent points.

[0085] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, the associated point pairs are classified as follows:

[0086] Boundary-Paired Points (BPP), including boundary points and their neighborhoods generated by Gaussian block masks;

[0087] Interior-Paired Points (IPP), including straight pipe masks, The area of ​​inliers covered by the elbow mask.

[0088] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the input dimension of the multi-layer perceptron MLP embedding network is is the original feature of the associated point pair, the hidden layer is [128, 64], and the output dimension For the final embedding dimension, the activation functions used are ReLU (Rectified Linear Unit) and BatchNorm (‌Batch Normalization).

[0089] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the original features of the associated point pair include the position, intensity value, normal vector and mask of the associated point pair.

[0090] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the final embedding dimension is 32 or 64.

[0091] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S330 includes:

[0092] S331, boundary associated point pair (BPP) loss function construction, boundary associated point pair (BPP) loss is geometric continuity loss, let any boundary associated point pair be , the normal vectors are , Boundary Pair Point (BPP) loss function The formula is:

[0093] ;

[0094] in, The embedded features of the boundary-related point pairs after MLP mapping are called the boundary-related point pair embedded features. The weight for balancing the consistency of the normal vector and the boundary associated point pair embedding features Indicates the total number of boundary-related point pairs, that is, the number of points involved in the calculation The number of all boundary-related point pairs;

[0095] S332, Intrinsic Point Pair (IPP) loss function construction, Intrinsic Point Pair (IPP) loss is feature consistency loss, let any Intrinsic Point Pair (IPP) be , Intrinsic Point Pair (IPP) loss function The formula is:

[0096] ;

[0097] in, Indicates the total number of inner point-related point pairs, that is, the number of points involved in the calculation The number of all interior point pairs (IPPs) of Indicates the pigment value or voxel value of the internal point association point pair, which is used to reflect its local color or density. represents the inlier point pair (IPP) embedding feature, is a hyperparameter, , adjust the weights between the color and density difference terms of the inlier point pair (IPP) and the embedding feature difference terms;

[0098] S333, the construction of the associated point pair loss function, the associated point pair loss function is the overall loss function, the integrated boundary associated point pair (BPP) loss function and the interior point associated point pair (IPP) loss function, the associated point pair loss function L The formula is as follows:

[0099] ;

[0100] in, R is the reward value of the self-masking strategy, which is dynamically adjusted according to the degree of correct classification or recognition of the mask category area. The larger the reward value, the more accurate the correct classification or recognition of the mask category area. It is defined as:

[0101] ;

[0102] in, is the true mask, To predict the mask, is an indicator function, which takes 1 when the predicted mask is correct and 0 otherwise;

[0103] is a hyperparameter, ,The values ​​of the hyperparameters are determined through cross-validation to balance the contributions of the ,geometric continuity loss of boundary associated point pairs (BPP), the ,feature consistency loss of inlier point pairs (IPP), and the reward ,function of mask category accuracy.

[0104] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, cross validation uses 5-fold cross validation, and in each fold cross validation, each set of hyperparameter combinations The performance is scored and sorted, and the highest scoring ones are selected.

[0105] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, step S340 includes:

[0106] S341, cosine similarity calculation, for each associated point pair P (i,j) , including boundary correlation points BPP and internal correlation point pairs IPP, using high-dimensional embedded features to calculate feature consistency scores, that is, using cosine similarity to evaluate whether the point pairs are in a consistent semantic relationship in the feature space, cosine similarity s ij The formula is as follows:

[0107] ;

[0108] in, s ij The closer the value is to 1, i 、 j Indicates the numbers of the two associated points in the associated point pair, which means that the associated point pair P (i,j) The higher the similarity in semantic relations, s ij The closer the value is to 0, the more likely it is that the point pair has no semantic association. s ij The smaller the value is than 0, the more likely it is that the semantic relationship between the two points is incorrect and they may come from different categories.

[0109] S342, comprehensive confidence calculation, for each associated point pair P (i,j) Calculate the overall confidence , the formula is as follows:

[0110] ;

[0111] in, represents the number of associated point pairs, The range is [0,1];

[0112] S343. Screening of the articular disc boundary point set with enhanced constraints. Screening of the boundary point set based on cosine similarity and comprehensive confidence. Selecting boundary points whose cosine similarity is greater than or equal to the cosine similarity threshold and whose comprehensive confidence is greater than or equal to the comprehensive confidence threshold. Obtaining the articular disc boundary point set with enhanced constraints and the associated point pair for associated learning to enhance the boundary.

[0113] Furthermore, in the articular disc modeling method based on structure masking and contrastive learning mechanism in the above embodiment, the cosine similarity threshold is equal to 0.9.

[0114] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the comprehensive confidence threshold is equal to 0.8.

[0115] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S400 includes:

[0116] S410, the center of the articular disc centerline is defined, and the set of articular disc boundary points is expressed as:

[0117] ;

[0118] based on y The maximum spacing between the axis edge point pairs is selected, and all boundary points are selected. y The coordinates are approximately the same, i.e. the difference is less than a predetermined threshold The point is right, x The point pair with the largest coordinate difference is taken as the left and right limit points:

[0119] ;

[0120] in, 、 They are the left limit point and the right limit point respectively, and the horizontal coordinates are all the boundary points, and the vertical coordinates are approximately the same ( ) has the smallest horizontal coordinate among the points and the maximum horizontal coordinate ;

[0121] S420, region replacement and area difference construction, defining the region of interest ROI , taking the center line of the articular disc as the axis of symmetry, all boundary points and mask areas are divided into the left half and right half , perform region replacement operation through region replacement function and construct area difference loss function;

[0122] S430, constructing a contrastive learning loss function, extracting high-dimensional embedding features of boundary points, defining positive sample pairs and negative sample pairs, and constructing a contrastive learning loss function based on cosine similarity;

[0123] 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 symmetrical boundary points of the center line of the articular disc and the symmetrical boundary of the center line of the articular disc, and complete the initial construction of the articular disc model.

[0124] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the predetermined threshold Normalization is performed based on the actual image resolution, specifically the product of the vertical resolution and 0.01.

[0125] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S420 includes:

[0126] S421, define the region of interest, construct a rectangle covering the main boundary area, that is, the region of interest ROI, The formula is as follows:

[0127] ;

[0128] in, To adjust parameters, control the region of interest ROI exist y Axial expansion range;

[0129] S422. Define a region replacement function. The region replacement function formula is as follows:

[0130] ;

[0131] in, , that is, to construct a mirror sample by reflecting the center line of the articular disc, Yes The horizontal coordinate after mirroring along the center line of the articular disc; The centerline of the articular disc, i.e. the left half and right half the dividing line;

[0132] S423. Construct an area difference loss function, use the boundary point mask area as the basis for area calculation, and set the original boundary mask area as , the mask area after region replacement is , define the area difference loss function for:

[0133] ;

[0134] Area difference loss function By comparing the spatial overlap between the original boundary mask area and the mask area after region permutation, the symmetry of the articular disc structure is reflected, and the area difference loss function The larger the value, the greater the difference between the original structure in the left and right mirror states, and the worse the symmetry.

[0135] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S430 includes:

[0136] S431, boundary point high-dimensional embedding feature extraction, boundary point high-dimensional embedding feature extraction formula is as follows:

[0137] ;

[0138] in, p is the boundary point;

[0139] S432, the definition of positive sample pairs and negative sample pairs, the original boundary mask area is geometrically transformed 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, which are used as similar reference samples of the boundary structure; the original boundary mask area is mirrored to obtain negative samples, which are used to identify asymmetry and improve the model's ability to distinguish inconsistencies between left and right structures, especially to identify lesions or irregular shapes; the original boundary mask area and positive samples Constitute a positive sample pair, the original boundary mask area and negative samples Constitute a negative sample pair;

[0140] S433, contrastive learning loss definition, construction of contrastive learning loss function based on cosine similarity , defined as follows:

[0141] ;

[0142] in, 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 sample, is the feature representation of negative samples.

[0143] Optionally, in the articular disc modeling method based on structure mask and contrastive learning mechanism in any of the above embodiments, the geometric transformation includes rotation and translation.

[0144] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S440 includes:

[0145] S441, boundary weighted fusion, weighted fusion of the predicted boundary, the associated point pair associated learning enhanced boundary and the articular disc centerline symmetric boundary to obtain the articular disc predicted boundary as follows:

[0146] ;

[0147] ;

[0148] in, To predict boundaries, the ability to perceive boundary locations is emphasized; Strengthen boundaries for association learning, emphasize the perception of associated point pairs, and improve local geometric continuity; The symmetrical boundary of the disc centerline is considered, and the symmetry and area reconstruction differences of the disc edge are taken into account. It emphasizes the repair of fuzzy areas or missing parts, reflecting the boundary completion capability. They are The weight coefficient of is used as a learnable parameter, and the Softmax function is used to calculate the Normalize to ensure that the following constraints are met:

[0149] ;

[0150] in, ;

[0151] S442. Constructing the loss function for the prediction boundary of the articular disc. The loss function for the prediction boundary of the articular disc is constructed as follows:

[0152] ;

[0153] in, are the hyperparameters of the area difference loss function and the contrastive learning loss function respectively;

[0154] S443, calculate the set of symmetrical boundary points of the center line of the articular disc and the symmetrical boundary of the center line of the articular disc, for any midpoint of the set of boundary points of the articular disc, Points are obtained by mapping the centerline of the articular disc , construct a set of symmetric point pairs, specifically expressed as:

[0155] ;

[0156] The set of symmetrical boundary points of the articular disc centerline Expressed as:

[0157] ;

[0158] That is, the set of articular disc boundary points and the set of symmetric point pairs The union of

[0159] The curve reconstruction method is used to reconstruct the boundary points of the disc centerline symmetry Perform boundary fitting, specifically expressed as:

[0160] ;

[0161] Rec(.) represents the set of symmetrical boundary points of the disc centerline using the curve reconstruction method. Perform boundary fitting to obtain the symmetrical boundary of the articular disc centerline.

[0162] Optionally, in the articular disc modeling method based on structure mask and contrast learning mechanism in any of the above embodiments, the curve reconstruction method includes an α-Shape algorithm, a Delaunay algorithm, and a B-spline algorithm.

[0163] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S500 includes:

[0164] S510, Boundary structure symmetry maintenance, to ensure that the symmetry of the articular disc structure is continuously strengthened, the consistency loss function formula is defined as follows:

[0165] ;

[0166] in, is the number of boundary points, is the feature extraction function, is the boundary point, It is a symmetric reflection point, that is, the structural symmetry property of the boundary point of the articular disc model learning;

[0167] S520, overall loss function definition, the overall loss function formula of the articular disc model is as follows:

[0168] ;

[0169] in, is the weight coefficient, ;

[0170] S530, articular disc model training, the articular disc model is trained using the Adam optimizer, the number of training rounds is increased by one, and when it is determined that the number of training rounds is less than the maximum number of training rounds, the process returns to step S200, and when the number of training rounds is greater than or equal to the maximum number of rounds, the articular disc model training is terminated. The articular disc model is represented as follows:

[0171] .

[0172] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the maximum number of rounds is T max =2000.

[0173] Optionally, in the articular disc modeling method based on structure mask and contrast learning mechanism in any of the above embodiments, the model evaluation index includes a boundary accuracy evaluation index and a structure symmetry evaluation index.

[0174] Optionally, in the articular disc modeling method based on structure masking and contrastive learning mechanism in any of the above embodiments, step S600 includes:

[0175] S610, evaluating the accuracy of the predicted boundary, using a boundary accuracy evaluation index to evaluate the accuracy of the articular disc boundary predicted by the articular disc model, and evaluating the degree of overlap between the predicted articular disc boundary area and the actual articular disc boundary area;

[0176] S620, predictive structural symmetry assessment, using a structural symmetry assessment index to assess the symmetry of the articular disc structure predicted by the articular disc model;

[0177] 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.

[0178] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the boundary accuracy evaluation index includes:

[0179] The Dice coefficient is as follows:

[0180] ;

[0181] in, P Predicting the boundary region for the articular disc, G The Dice coefficient reflects the degree of overlap between the predicted and true boundary areas of the articular disc. Its value range is [0, 1]. The closer the Dice coefficient value is to 1, the more consistent the predicted and true boundary areas are, and the more accurate the articular disc model is. The closer the Dice coefficient value is to 0, the lower the degree of overlap between the predicted and true boundary areas, indicating that the articular disc model construction has failed.

[0182] IoU (Jaccard), the formula is as follows:

[0183] ;

[0184] IoU Jaccard measures the overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc by calculating the intersection-over-union ratio between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, thereby reflecting whether the articular disc model is overfitting or underfitting. IoU (Jaccard) value range is [0,1], IoU The closer the (Jaccard) value is to 1, the higher the degree of overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, and the more accurate the articular disc model is. IoU The closer the (Jaccard) value is to 0, the lower the degree of complete overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, indicating that the construction of the articular disc model has failed.

[0185] Furthermore, in the articular disc modeling method based on the structure mask and contrastive 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 bilaterally symmetrical structure. The formula is as follows:

[0186] ;

[0187] SCS Measures the consistency of the predicted boundary structure of the articular disc model relative to the left and right structures of the axis, and predicts the boundary points Its symmetrical mapping point about the center line of the articular disc The average of the spatial deviations between , 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.

[0188] Furthermore, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, the initial mask update strategy is: combined with the evaluation index, when the Dice coefficient value is greater than the Dice threshold or IoU (Jaccard) less than IoU Threshold, initial radius Increase the specified radius increment and the initial rotation angle Increase the specified angle increment; when SCS Value greater than SCS When the threshold is reached, the Gaussian parameter increases by the specified parameter increment.

[0189] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the Dice threshold is equal to 0.85.

[0190] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, IoU The threshold is equal to 0.75.

[0191] Preferably, in the articular disc modeling method based on structure mask and contrast learning mechanism in the above embodiment, SCS The threshold is equal to 5%.

[0192] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the range of the radius increment is specified to be [0.01, 0.03].

[0193] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the radius increment is specified to be 0.02.

[0194] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the range of the specified angle increment is [1, 3].

[0195] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the angle increment is specified as 2.

[0196] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the range of the parameter increment is specified to be [1, 2].

[0197] Preferably, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, the parameter increment is specified to be 1.

[0198] The present invention uses a disc model constructed by combining mask generation, point-pair association learning, and boundary-enhanced contrast learning to achieve reconstruction of the disc structure and enhancement of edge details. This improves the detection sensitivity and positioning accuracy of the disc region, enhances the structural consistency between boundary points and internal points, improves the ability to recover morphologically missing areas, enhances the reliability of boundary reconstruction, and improves the responsiveness in weak boundary signal areas. The disc model constructed by the present invention solves the technical problems of blurred boundaries, incomplete morphological reconstruction, poor boundary point connectivity, and lack of structural symmetry that exist in the disc segmentation and reconstruction process.

[0199] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0200] Figure 1 is a flow chart of an articular disc modeling method based on structure mask and contrastive learning mechanism according to an exemplary embodiment. DETAILED DESCRIPTION

[0201] The following describes several preferred embodiments of the present invention with reference to the accompanying drawings to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0202] In the drawings, components with identical structures 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 arbitrary and are not limited by the present invention. To enhance clarity, the thickness of components in some places in the drawings is schematically exaggerated.

[0203] This application provides a method for modeling articular discs based on structure masking and contrastive learning mechanisms, such as Figure 1 As shown, the following steps are included:

[0204] S100, mask generation and mask parameter initialization, input the articular disc image into the conventional segmentation network, extract the articular disc centerline, and generate a mask, including:

[0205] Straight tube mask, with the center line of the articular disc as the main axis, generates a radius of r , the direction is The cylindrical area of ​​​​the straight tube mask is obtained;

[0206] Bend mask, rotated about the disc centerline As the new main axis, the rotation angle is , the area of ​​the deflection curve around the new principal axis is elbow mask;

[0207] Gaussian block mask: fit an elliptical area at the boundary points of the articular disc to obtain an ellipsoidal mask based on boundary fitting and Gaussian distribution, namely, a Gaussian block mask; step S100 specifically includes:

[0208] S110, joint disc centerline extraction, the joint disc image is input into the conventional segmentation network U-Net, the joint disc area mask and the joint disc boundary point set are obtained, the joint disc centerline is obtained using the skeleton extraction algorithm for the joint disc area mask C， The skeleton extraction algorithm uses the Medial Axis Transform algorithm, the centerline of the articular disc C Represented as a set of discrete points:

[0209] ;

[0210] in, is the coordinate of the discrete point 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, which is the number of discrete points;

[0211] S120, mask generation, based on the centerline of the articular disc C and the articular disc boundary point set to generate masks, including straight tube masks, Bend pipe mask and Gaussian block mask, and generate corresponding mask areas; specifically including:

[0212] S121. Straight tube mask generation. Straight tube mask is used for non-deformation area. Non-deformation area is the area where the center line or boundary curvature of the articular disc has no obvious change or the deformation area accounts for a small proportion. Specifically: boundary curvature The range is ; or the offset distance is less than 1mm; or the deformation area accounts for less than 10% of the total structural area of ​​the articular disc, with the centerline of the articular disc as the C As the main axis, the radius is generated along the natural direction of the articular disc (Y axis) , the direction is The cylindrical area; let the center line of the articular disc be , the coordinates of each point are , then the straight pipe mask Defined as:

[0213] ;

[0214] in Represents a point in two-dimensional space A Boolean value indicating whether it belongs to the straight pipe mask area, with a value of 1 or 0. is the center and the radius is The mask value of the area within and is 1, and the others are 0;

[0215] S122, Bend mask generation, The elbow mask is used for mild to moderate deformation areas. Mild to moderate deformation areas are areas where the centerline or boundary curvature of the articular disc changes slowly or the deformation area accounts for a medium proportion. Specifically: boundary curvature The range is ; or the offset distance is in the interval [1mm, 5mm]; or the deformation area accounts for the total structural area of ​​the articular disc in the interval [10%, 40%], with the centerline of the articular disc C Rotation As the new principal axis, let the rotation angle be , the new principal axis function is denoted as , the deflection curve formula is:

[0216] ;

[0217] The elbow mask is defined as:

[0218] ;

[0219] in Represents a point in two-dimensional space Whether it belongs to The Boolean value of the elbow mask area, which can be 1 or 0. is the point on the new principal axis, As the radius, is the center and the radius is The mask value of the area within and is 1, and the others are 0;

[0220] S123. Generate Gaussian block mask. The Gaussian block mask is used for large deformation areas. Large deformation areas are areas where the center line or boundary curvature of the articular disc changes dramatically or the deformation area accounts for a large proportion. Specifically: 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, principal component analysis) on the set of articular disc boundary points to extract the main curvature direction , for the deformed area of ​​the articular disc, an ellipsoidal mask based on boundary fitting and Gaussian distribution, namely, a Gaussian block mask, is generated at point The mask strength obeys the two-dimensional Gaussian distribution, which represents the response weight of each pixel point on the two-dimensional plane to the final structure mask. Its value is between [0, 1], the center point value is 1, and the closer to the edge, the closer to 0. The Gaussian block mask Defined as:

[0221] ;

[0222] in, is the standard deviation of the Gaussian distribution in the horizontal direction, which is used to control the expansion range of the fitting ellipse in the lateral edge deformation direction, ranging from 1.0 to 3.0 mm; is the standard deviation of the Gaussian distribution in the vertical direction, which is used to control the expansion range of the fitting ellipse in the longitudinal boundary extension direction, ranging from 0.5 to 2.0 mm;

[0223] S130, initialization of mask parameters, setting the initial values ​​of mask parameters, including the radius , the rotation angle is , and Gaussian parameters , initial radius 10mm, initial rotation angle is 0, the initial Gaussian parameter The initial value of the mask parameter is (10, 10), and the initial value of the mask parameter is applied to the articular disc image to perform local enhancement on the articular disc area to improve the feature representation. The number of training rounds is initialized to 0.

[0224] S200, mask parameter optimization, optimizes mask parameters based on reinforcement learning, performs mask prediction, obtains optimal mask parameters, and saves the set of articular disc boundary points; specifically includes:

[0225] S210, adjusting mask parameters. Adjusting mask parameters according to the mask adjustment strategy. The radius of the straight tube mask is reduced with a fixed step size of 0.02 mm. The center area is refined until the minimum radius is 5 mm. The elbow mask increases the rotation angle by a fixed angle of 1 degree until the maximum angle is 90 degrees. The Gaussian block mask is improved by a fixed ellipsoidal flattening of 0.2, and the boundary envelope is strengthened until the maximum change of the ellipsoidal flattening is 2.

[0226] S220, mask parameter optimization, based on reinforcement learning to optimize the mask parameters, perform mask prediction, obtain the articular disc boundary point set and the articular disc prediction boundary, when the boundary IoU (Intersection over Union) is lower than the set limit IoU When the threshold is 0.7, return to step S210 until the continuous 3 sub-border IoU When the improvement rate is lower than the minimum change of 0.001 or the maximum number of iterations reaches 500, the mask prediction is completed, the optimal mask parameters are obtained, and the set of articular disc boundary points is saved. Specifically, the following steps are performed:

[0227] S221, state space design, mask selection and mask parameter setting constitute the current state, denoted as:

[0228] ;

[0229] in, Indicates the mask, that is, the straight pipe mask, elbow mask and Gaussian block mask, Indicates the spindle rotation angle, used for elbow mask, 、 They 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 rate of change of boundary curvature;

[0230] S222, action space design, select actions according to the current state, including switching masks T and adjust mask parameters, including , the action space is defined as:

[0231] ;

[0232] S223, reward function design, define the reward function as the joint disc boundary point detection Lift and overall detection accuracy of the articular disc Lift The reward function is a weighted combination of The formula is as follows:

[0233] ;

[0234] in, 、 is the empirical weight coefficient, ,and , preferably .

[0235] S224, mask parameter optimization, using the Actor-Critic framework based on policy gradient to optimize mask parameters, complete mask prediction, obtain the optimal mask parameters, and save the set of articular disc boundary points. The policy gradient formula is as follows:

[0236] ;

[0237] in, represents the state and action distribution sampled from the current policy, Indicates that in the current state s, according to the parameter The policy network generates actions from the action space a The probability distribution of Indicates action a Log-probability gradient under the current policy; is the advantage function, which measures the execution of actions a How much better than the average policy, the advantage function is defined as:

[0238] ;

[0239] in, is the discount factor, preferably 0.95, to control the impact of future state value; To perform an action a After that, for the next state Estimated value of is the estimated value of the current state s.

[0240] S300, determining the association loss function, constructing association point pairs, including boundary association point pairs and interior point association point pairs, learning high-dimensional embedding features, determining the joint association loss function, screening and optimizing the set of articular disc boundary points, and obtaining association point pairs for association learning to strengthen boundaries; specifically including:

[0241] S310, build the associated point pair, the straight pipe mask, The different regions of the articular disc formed by the elbow mask and Gaussian block mask are divided, and each point is assigned a spatial role, which includes boundary points, internal points, and boundary adjacent points. Associated point pairs are formed and classified into:

[0242] Boundary-Paired Points (BPP), including boundary points and their neighborhoods generated by Gaussian block masks;

[0243] Interior-Paired Points (IPP), including straight pipe masks, The inlier area covered by the elbow mask;

[0244] S320, high-dimensional embedding feature learning, using the 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 feature of the associated point pair, including the position, intensity value, normal vector and mask of the associated point pair. The hidden layer is [128, 64] and the output dimension is To achieve a final embedding dimension of 64, the activation function uses ReLU (Rectified Linear Unit) and BatchNorm (‌Batch Normalization)

[0245] S330: Constructing a pair-of-points loss function. Constructing a boundary-paired points (BPP) loss function and an interior-paired points (IPP) loss function. Combining the boundary-paired points (BPP) loss function and the interior-paired points (IPP) loss function to generate a joint pair-of-points loss function. Specifically, the following steps are performed:

[0246] S331, boundary associated point pair (BPP) loss function construction, boundary associated point pair (BPP) loss is geometric continuity loss, let any boundary associated point pair be , the normal vectors are , Boundary Pair Point (BPP) loss function The formula is:

[0247] ;

[0248] in, The embedded features of the boundary-related point pairs after MLP mapping are called the boundary-related point pair embedded features. The weight for balancing the consistency of the normal vector and the boundary associated point pair embedding features Indicates the total number of boundary-related point pairs, that is, the number of points involved in the calculation The number of all boundary-related point pairs;

[0249] S332, Intrinsic Point Pair (IPP) loss function construction, Intrinsic Point Pair (IPP) loss is feature consistency loss, let any Intrinsic Point Pair (IPP) be , Intrinsic Point Pair (IPP) loss function The formula is:

[0250] ;

[0251] in, Indicates the total number of inner point-related point pairs, that is, the number of points involved in the calculation The number of all interior point pairs (IPPs) of Indicates the pigment value or voxel value of the internal point association point pair, which is used to reflect its local color or density. represents the inlier point pair (IPP) embedding feature, is a hyperparameter, , adjust the weights between the color and density difference terms of the inlier point pair (IPP) and the embedding feature difference terms;

[0252] S333, the construction of the associated point pair loss function, the associated point pair loss function is the overall loss function, the integrated boundary associated point pair (BPP) loss function and the interior point associated point pair (IPP) loss function, the associated point pair loss function L The formula is as follows:

[0253] ;

[0254] in, R is the reward value of the self-masking strategy, which is dynamically adjusted according to the degree of correct classification or recognition of the mask category area. The larger the reward value, the more accurate the correct classification or recognition of the mask category area. It is defined as:

[0255] ;

[0256] in, is the true mask, To predict the mask, is an indicator function, which takes 1 when the predicted mask is correct and 0 otherwise;

[0257] is a hyperparameter, , the values ​​of the hyperparameters are cross-validated using 5-fold cross-validation, and each set of hyperparameter combinations is tested in each fold cross-validation. The performance is scored and ranked, and the highest score is selected to balance the contribution of the reward function of geometric continuity loss of boundary point pairs (BPP), feature consistency loss of interior point pairs (IPP), and mask category accuracy.

[0258] Furthermore, in the articular disc modeling method based on structure mask and contrastive learning mechanism in the above embodiment, cross-validation is performed.

[0259] S340, boundary point set screening, using high-dimensional embedding features and associated point pair loss function, screening the articular disc boundary point set to obtain an articular disc boundary point set with enhanced constraints and associated point pair associated learning to strengthen the boundary; specifically including:

[0260] S341, cosine similarity calculation, for each associated point pair P (i,j) , including boundary correlation points BPP and internal correlation point pairs IPP, using high-dimensional embedded features to calculate feature consistency scores, that is, using cosine similarity to evaluate whether the point pairs are in a consistent semantic relationship in the feature space, cosine similarity s ij The formula is as follows:

[0261] ;

[0262] in, s ij The closer the value is to 1, i 、 j Indicates the numbers of the two associated points in the associated point pair, which means that the associated point pair P (i,j) The higher the similarity in semantic relations, s ij The closer the value is to 0, the more likely it is that the point pair has no semantic association. s ij The smaller the value is than 0, the more likely it is that the semantic relationship between the two points is incorrect and they may come from different categories.

[0263] S342, comprehensive confidence calculation, for each associated point pair P (i,j) Calculate the overall confidence , the formula is as follows:

[0264] ;

[0265] in, represents the number of associated point pairs, The range is [0,1];

[0266] S343. Screening of the articular disc boundary point set with enhanced constraints. Screening of the articular disc boundary point set based on cosine similarity and comprehensive confidence. Selecting boundary points with cosine similarity greater than or equal to the cosine similarity threshold of 0.9 and comprehensive confidence greater than or equal to the comprehensive confidence threshold of 0.8 to obtain the articular disc boundary point set with enhanced constraints and the associated point pair association learning enhanced boundary.

[0267] S400, boundary enhancement contrast learning, defining the center of the articular disc centerline and dividing the region of interest ROI For the left half and right half , construct the area difference loss function, construct the contrastive learning loss function, calculate the articular disc prediction boundary, construct the articular disc prediction boundary loss function, calculate the articular disc centerline symmetric boundary point set and the articular disc centerline symmetric boundary, and complete the initial construction of the articular disc model; specifically including:

[0268] S410, the center of the articular disc centerline is defined, and the set of articular disc boundary points is expressed as:

[0269] ;

[0270] based on y The maximum spacing between the axis edge point pairs is selected, and all boundary points are selected. y The coordinates are approximately the same, i.e. the difference is less than a predetermined threshold The point is right, x The point pair with the largest coordinate difference is taken as the left and right limit points, and the predetermined threshold Normalize the image based on its actual resolution, specifically by multiplying the vertical resolution by 0.01:

[0271] ;

[0272] in, 、 They are the left limit point and the right limit point respectively, and the horizontal coordinates are all the boundary points, and the vertical coordinates are approximately the same ( ) has the smallest horizontal coordinate among the points and the maximum horizontal coordinate ;

[0273] S420, region replacement and area difference construction, defining the region of interest ROI , taking the center line of the articular disc as the axis of symmetry, all boundary points and mask areas are divided into the left half and right half , perform regional replacement operation through regional replacement function and construct area difference loss function; specifically including:

[0274] S421, define the region of interest, construct a rectangle covering the main boundary area, that is, the region of interest ROI, The formula is as follows:

[0275] ;

[0276] in, To adjust parameters, control the region of interest ROI exist y Axial expansion range;

[0277] S422. Define a region replacement function. The region replacement function formula is as follows:

[0278] ;

[0279] in, , that is, to construct a mirror sample by reflecting the center line of the articular disc, Yes The horizontal coordinate after mirroring along the center line of the articular disc; The centerline of the articular disc, i.e. the left half and right half the dividing line;

[0280] S423. Construct an area difference loss function, use the boundary point mask area as the basis for area calculation, and set the original boundary mask area as , the mask area after region replacement is , define the area difference loss function for:

[0281] ;

[0282] Area difference loss function By comparing the spatial overlap between the original boundary mask area and the mask area after region permutation, the symmetry of the articular disc structure is reflected, and the area difference loss function The larger the value, the greater the difference between the original structure and the left and right mirror images, and the worse the symmetry;

[0283] S430, constructing a contrastive learning loss function, extracting high-dimensional embedding features of boundary points, defining positive sample pairs and negative sample pairs, and constructing a contrastive learning loss function based on cosine similarity;

[0284] Specifically include:

[0285] S431, boundary point high-dimensional embedding feature extraction, boundary point high-dimensional embedding feature extraction formula is as follows:

[0286] ;

[0287] in, p is the boundary point;

[0288] S432, the definition of positive sample pairs and negative sample pairs, the original boundary mask area is geometrically transformed to obtain positive samples, which include rotation and translation. The geometric transformation is used to train the model to enhance boundary invariance and construct reasonable boundary variants close to the real structure, which are used as similar reference samples of the boundary structure; the original boundary mask area is mirrored to obtain negative samples, which are used to identify asymmetry and improve the model's ability to distinguish inconsistencies between left and right structures, especially to identify lesions or irregular shapes; the original boundary mask area and positive samples Constitute a positive sample pair, the original boundary mask area and negative samples Constitute a negative sample pair;

[0289] S433, contrastive learning loss definition, construction of contrastive learning loss function based on cosine similarity , defined as follows:

[0290] ;

[0291] in, 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 sample, is the feature representation of negative samples;

[0292] 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 of the centerline of the articular disc and the symmetric boundary of the centerline of the articular disc, and complete the initial construction of the articular disc model; specifically including:

[0293] S441, boundary weighted fusion, weighted fusion of the predicted boundary, the associated point pair associated learning enhanced boundary and the articular disc centerline symmetric boundary to obtain the articular disc predicted boundary as follows:

[0294] ;

[0295] ;

[0296] in, To predict boundaries, the ability to perceive boundary locations is emphasized; Strengthen boundaries for association learning, emphasize the perception of associated point pairs, and improve local geometric continuity; The symmetrical boundary of the disc centerline is considered, and the symmetry and area reconstruction differences of the disc edge are taken into account. It emphasizes the repair of fuzzy areas or missing parts, reflecting the boundary completion capability. They are The weight coefficient of is used as a learnable parameter, and the Softmax function is used to calculate the Normalize to ensure that the following constraints are met:

[0297] ;

[0298] in, ;

[0299] S442. Constructing the loss function for the prediction boundary of the articular disc. The loss function for the prediction boundary of the articular disc is constructed as follows:

[0300] ;

[0301] in, are the hyperparameters of the area difference loss function and the contrastive learning loss function respectively;

[0302] S443, calculate the set of symmetrical boundary points of the center line of the articular disc and the symmetrical boundary of the center line of the articular disc, for any midpoint of the set of boundary points of the articular disc, Points are obtained by mapping the centerline of the articular disc , construct a set of symmetric point pairs, specifically expressed as:

[0303] ;

[0304] The set of symmetrical boundary points of the articular disc centerline Expressed as:

[0305] ;

[0306] That is, the set of articular disc boundary points and the set of symmetric point pairs The union of

[0307] The curve reconstruction method α-Shape algorithm is used to reconstruct the set of symmetrical boundary points of the articular disc centerline Perform boundary fitting, which is specifically expressed as:

[0308] ;

[0309] Rec(.) represents the set of symmetrical boundary points of the disc centerline using the curve reconstruction method. Perform boundary fitting to obtain the symmetrical boundary of the articular disc centerline.

[0310] S500, articular disc model training, constructing the overall loss function of the articular disc model, training the articular disc model using the Adam optimizer, increasing the number of training rounds by one, and returning to step S200 when the number of training rounds is less than the maximum number of rounds. When the number of training rounds is greater than or equal to the maximum number of rounds, the articular disc model training is terminated. Specifically, the steps include:

[0311] S510, Boundary structure symmetry maintenance, to ensure that the symmetry of the articular disc structure is continuously strengthened, the consistency loss function formula is defined as follows:

[0312] ;

[0313] in, is the number of boundary points, is the feature extraction function, is the boundary point, It is a symmetric reflection point, that is, the structural symmetry property of the boundary point of the articular disc model learning;

[0314] S520, overall loss function definition, the overall loss function formula of the articular disc model is as follows:

[0315] ;

[0316] in, is the weight coefficient, ;

[0317] S530, articular disc model training, the articular disc model is trained using the Adam optimizer, the number of training rounds is increased by one, and the number of training rounds is judged to be less than the maximum number of training rounds T max =2000, return to step S200, and when the number of training rounds is greater than or equal to the maximum number of rounds, the articular disc model training is terminated. The articular disc model is represented as follows:

[0318] .

[0319] S600: Output the articular disc model and evaluate the articular disc model using the model evaluation index. If the articular disc model meets the model evaluation index, save 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. The model evaluation index includes a boundary accuracy evaluation index and a structural symmetry evaluation index. Specifically, it includes:

[0320] S610: Prediction boundary accuracy assessment: Use boundary accuracy assessment indicators to assess the accuracy of the articular disc boundary predicted by the articular disc model, and assess the degree of overlap between the predicted articular disc boundary area and the true articular disc boundary area. Boundary accuracy assessment indicators include:

[0321] The Dice coefficient is as follows:

[0322] ;

[0323] in, P Predicting the boundary region for the articular disc, G The Dice coefficient reflects the degree of overlap between the predicted and true boundary areas of the articular disc. Its value range is [0, 1]. The closer the Dice coefficient value is to 1, the more consistent the predicted and true boundary areas are, and the more accurate the articular disc model is. The closer the Dice coefficient value is to 0, the lower the degree of overlap between the predicted and true boundary areas, indicating that the articular disc model construction has failed.

[0324] IoU (Jaccard), the formula is as follows:

[0325] ;

[0326] IoU Jaccard measures the overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc by calculating the intersection-over-union ratio between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, thereby reflecting whether the articular disc model is overfitting or underfitting. IoU (Jaccard) value range is [0,1], IoU The closer the (Jaccard) value is to 1, the higher the degree of overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, and the more accurate the articular disc model is. IoU The closer the (Jaccard) value is to 0, the lower the degree of complete overlap between the predicted boundary area of ​​the articular disc and the true boundary area of ​​the articular disc, indicating that the construction of the articular disc model has failed;

[0327] S620: Predicted structural symmetry assessment. The structural symmetry assessment index is used to evaluate the symmetry of the articular disc structure predicted by the articular disc model. The structural symmetry assessment index uses the structural symmetry index (Symmetry Consistency Score, SCS) to quantify the error of the bilaterally symmetrical structure. The formula is as follows:

[0328] ;

[0329] SCS Measures the consistency of the predicted boundary structure of the articular disc model relative to the left and right structures of the axis, and predicts the boundary points Its symmetrical mapping point about the center line of the articular disc The average of the spatial deviations between , when the predicted boundary structure is highly symmetric, SCSThe value approaches 0, when the predicted boundary structure produces contour deformation, SCS The larger the value;

[0330] 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: combined with the evaluation index, when the Dice coefficient value is greater than the Dice threshold 0.85 or IoU (Jaccard) less than IoU When the threshold is 0.75, the initial radius Increase the specified radius increment by 0.02 and the initial rotation angle Increase the specified angle by 2; when SCS Value greater than SCS When the threshold is 5%, the Gaussian parameter is increased by a specified parameter increment of 1, and the process returns to step S200.

[0331] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for modeling articular discs based on structure masking and contrastive learning mechanism, characterized in that: The steps include: S100, mask generation and mask parameter initialization, input the articular disc image into the conventional segmentation network, extract the articular disc centerline, generate the straight tube mask, For the elbow mask and Gaussian block mask, set the initial values ​​of the mask parameters and initialize the number of training rounds to 0; S200, mask parameter optimization: optimizing the mask parameters based on reinforcement learning, performing mask prediction, obtaining optimal mask parameters, and saving a set of articular disc boundary points; S300, determining the association loss function, constructing association point pairs, including boundary association point pairs and interior point association point pairs, learning high-dimensional embedding features, determining the joint association loss function, screening and optimizing the set of articular disc boundary points, and obtaining association point pairs for association learning to strengthen boundaries; S400, boundary enhancement contrast learning, defining the center of the articular disc centerline and dividing the region of interest ROI For the left half and right half , construct the area difference loss function, construct the contrastive learning loss function, calculate the articular disc prediction boundary, construct the articular disc prediction boundary loss function, calculate the articular disc centerline symmetric boundary point set and the articular disc centerline symmetric boundary, and complete the initial construction of the articular disc model; S500, training the articular disc model, constructing an overall loss function for the articular disc model, training the articular disc model using the Adam optimizer, increasing the number of training rounds by one, and returning to step S200 when it is determined that the number of training rounds is less than the maximum number of rounds, and ending the articular disc model training when the number of training rounds is greater than or equal to the maximum number of rounds; S600, output the articular disc model, use the model evaluation index to evaluate the articular disc model, save and output the articular disc model when the articular disc model meets the model evaluation index, 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.

2. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 1, characterized in that: The step S100 includes: S110, extracting the centerline of the articular disc. Input the articular disc image into the conventional segmentation network to obtain an articular disc region mask and an articular disc boundary point set. Use a skeleton extraction algorithm to obtain the centerline of the articular disc based on the articular disc region mask. C ; S120, mask generation based on the centerline of the articular disc C and the articular disc boundary point set to generate a mask, including a straight tube mask, Bend pipe mask and Gaussian block mask, and generate corresponding mask areas; S130, initializing the mask parameters, setting the initial values ​​of the mask parameters, the mask parameters including the radius , the rotation angle is , and Gaussian parameters , initial radius 10mm, initial rotation angle is 0, the initial Gaussian parameter is (10, 10), the initial value of the mask parameter is applied to the articular disc image, and the articular disc area is locally enhanced to improve feature representation.

3. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 2, characterized in that: The step S120 includes: S121, generating a straight tube mask, wherein the straight tube mask is used for the non-deformation area, and the center line of the articular disc is used as the center line of the articular disc. C The main axis is the radius generated along the natural direction of the articular disc. , the direction is The cylindrical area; let the center line of the articular disc be , the coordinates of each point are , then the straight pipe mask Defined as: ； in Represents a point in two-dimensional space A Boolean value indicating whether it belongs to the straight pipe mask area, with a value of 1 or 0. is the center and the radius is The mask value of the area within and is 1, and the others are 0; S122, Bend mask generation, the The elbow mask is used for mild to moderate deformation areas, with the centerline of the articular disc C Rotation As the new principal axis, let the rotation angle be , the new principal axis function is denoted as , the deflection curve formula is: ； described The elbow mask is defined as: ； in Represents a point in two-dimensional space Whether it belongs to The Boolean value of the elbow mask area, which can be 1 or 0. is the point on the new principal axis, As the radius, is the center and the radius is The mask value of the area within and is 1, and the others are 0; S123, generating a Gaussian block mask, wherein the Gaussian block mask is used for the large deformation area, performing local PCA on the set of articular disc boundary points, and extracting the main curvature direction , for the deformed area of ​​the articular disc, an ellipsoidal mask based on boundary fitting and Gaussian distribution is generated, namely the Gaussian block mask, at point The mask strength obeys the two-dimensional Gaussian distribution, which represents the response weight of each pixel point on the two-dimensional plane to the final structure mask. Its value is between [0, 1], the center point value is 1, and the closer to the edge, the closer to 0. The Gaussian block mask Defined as: ； in, is the standard deviation of the Gaussian distribution in the horizontal direction, which is used to control the expansion range of the fitting ellipse in the lateral edge deformation direction, ranging from 1.0 to 3.0 mm; is the standard deviation of the Gaussian distribution in the vertical direction, which is used to control the expansion range of the fitting ellipse in the longitudinal boundary extension direction, ranging from 0.5 to 2.0 mm.

4. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 3, characterized in that: The step S200 includes: S210, adjusting the mask parameters, adjusting the mask parameters according to the mask adjustment strategy, reducing the radius of the straight pipe mask with a fixed step size, and refining the central area until the minimum radius; The elbow mask increases the rotation angle by a fixed angle until the maximum angle is reached; the Gaussian block mask is improved by a fixed ellipsoidal flattening, strengthening the boundary envelope until the ellipsoidal flattening changes to the maximum; S220, mask parameter optimization, based on reinforcement learning to optimize the mask parameters, perform mask prediction, obtain the articular disc boundary point set and the articular disc prediction boundary, when the boundary IoU Below the set limit IoU When the threshold is reached, return to step S210 until the number of consecutive N sub-border IoU When the improvement rate is lower than the minimum change or reaches the maximum number of iterations, the mask prediction is completed, the optimal mask parameters are obtained, and the set of articular disc boundary points is saved.

5. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 4, characterized in that: The step S220 includes: S221, state space design, mask selection and the mask parameter setting constitute the current state, denoted as: ； in, Indicates the mask, that is, the straight pipe mask, elbow mask and Gaussian block mask, Indicates the spindle rotation angle, used for elbow mask, 、 They 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 rate of change of boundary curvature; S222, action space design, select action according to the current state, including switching mask T and adjust the mask parameters, including , the action space is defined as: ； S223, reward function design, define the reward function as the joint disc boundary point detection Lift and overall detection accuracy of the articular disc Lift The reward function is a weighted combination of The formula is as follows: ； in, 、 is the empirical weight coefficient, ,and ; S224: Optimize the mask parameters using 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 articular disc boundary points. The policy gradient formula is as follows: ； in, represents the state and action distribution sampled from the current policy, Indicates that in the current state s, according to the parameter The policy network generates actions from the action space a The probability distribution of Indicates action a Log-probability gradient under the current policy; is the advantage function, which measures the execution of actions a How much better than the average policy, the advantage function is defined as: ； in, is the discount factor, which controls the impact of future state value; To perform an action a After that, for the next state Estimated value of is the estimated value of the current state s.

6. The articular disc modeling method based on structure mask and contrast learning mechanism according to claim 5, characterized in that: The step S300 includes: S310, constructing a pair of associated points, and performing the operations on the straight pipe mask and the The different regions of the articular disc formed by the elbow mask and the Gaussian block mask are divided, a spatial role is assigned to each point, and associated point pairs are formed, and the associated point pairs are classified into boundary associated point pairs and interior point associated point pairs; S320, high-dimensional embedding feature learning, mapping the associated point pairs using a multi-layer perceptron (MLP) embedding network to perform high-dimensional embedding feature learning; S330, constructing a correlation point pair loss function, constructing a boundary correlation point pair loss function and an interior point correlation point pair loss function, and combining the boundary correlation point pair loss function and the interior point correlation point pair loss function to generate a joint correlation point pair loss function; S340, boundary point set screening, using high-dimensional embedding features and the associated point pair loss function to screen the articular disc boundary point set to obtain an articular disc boundary point set with enhanced constraints and associated point pair associated learning to strengthen the boundary.

7. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 6, characterized in that: The step S330 includes: S331, constructing the BPP loss function of boundary association point pairs, wherein the boundary association point pair loss is the geometric continuity loss. Assume that any boundary association point pair is , the normal vectors are , the boundary association point loss function The formula is: ； in, The embedded features of the boundary-related point pairs after MLP mapping are called the boundary-related point pair embedded features. To balance the weight of the normal vector and the boundary associated points on the embedded feature consistency, Indicates the total number of boundary-related point pairs, that is, the number of points involved in the calculation The number of all boundary-related point pairs; S332, construct the IPP loss function of the inner point association point pair, the inner point association point pair loss is the feature consistency loss, and assume that any inner point association pair is , the inner point association point pair loss function The formula is: ； in, Indicates the total number of inner point-related point pairs, that is, the number of points involved in the calculation The number of all internal point-connected point pairs IPP, Indicates the pigment value or voxel value of the internal point association point pair, which is used to reflect its local color or density. Indicates the IPP embedding feature of the inner point association pair, is a hyperparameter, , adjust the weights between the inlier point-association point pair IPP color and density difference terms and the embedded feature difference terms; S333, constructing the associated point pair loss function, which is the overall loss function, by integrating the boundary associated point pair loss function and the interior point associated point pair loss function. L The formula is as follows: ； in, R is the reward value of the self-masking strategy, which is dynamically adjusted according to the degree of correct classification or recognition of the mask category area. The larger the reward value, the more accurate the correct classification or recognition of the mask category area. It is defined as: ； in, is the true mask, To predict the mask, is an indicator function, which takes 1 when the predicted mask is correct and 0 otherwise; is a hyperparameter, ,The values ​​of the hyperparameters are determined through cross-validation to balance the contributions of the ,geometric continuity loss of boundary associated points to BPP, the feature consistency loss of ,inlier associated points to IPP, and the reward function of mask category accuracy.

8. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 7, characterized in that: The step S400 includes: S410, the center of the articular disc centerline is defined, and the articular disc boundary point set is expressed as: ； based on y The maximum spacing between the axis edge point pairs is selected, and all boundary points are selected. y The coordinates are approximately the same, i.e. the difference is less than a predetermined threshold The point is right, x The point pair with the largest coordinate difference is taken as the left and right limit points: ； in, 、 They are the left limit point and the right limit point respectively. The horizontal coordinates are all the boundary points, and the vertical coordinates are approximately the same as The smallest horizontal coordinate among the points and the maximum horizontal coordinate ; S420, region replacement and area difference construction, defining the region of interest ROI , taking the center line of the articular disc as the axis of symmetry, all boundary points and mask areas are divided into the left half and right half , perform region replacement operation through region replacement function and construct area difference loss function; S430, constructing a contrastive learning loss function, extracting high-dimensional embedding features of boundary points, defining positive sample pairs and negative sample pairs, and constructing a contrastive learning loss function based on cosine similarity; 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 symmetrical boundary points of the center line of the articular disc and the symmetrical boundary of the center line of the articular disc, and complete the initial construction of the articular disc model.

9. The articular disc modeling method based on structure masking and contrastive learning mechanism according to claim 7 or 8, characterized in that: The step S500 includes: S510, Boundary structure symmetry maintenance, to ensure that the symmetry of the articular disc structure is continuously strengthened, the consistency loss function formula is defined as follows: ； in, is the number of boundary points, is the feature extraction function, is the boundary point, It is a symmetric reflection point, that is, the structural symmetry property of the boundary point of the articular disc model learning; S520, overall loss function definition, the overall loss function formula of the articular disc model is as follows: ； in, is the boundary loss function for disc prediction, is the weight coefficient, ; S530, articular disc model training, the articular disc model is trained using the Adam optimizer, the number of training rounds is increased by one, and when it is determined that the number of training rounds is less than the maximum number of rounds, the process returns to step S200, and when the number of training rounds is greater than or equal to the maximum number of rounds, the articular disc model training is terminated. The articular disc model is represented as follows: ； in, is the set of symmetrical boundary points about the center line of the articular disc.

10. The articular disc modeling method based on structure mask and contrastive learning mechanism according to claim 9, characterized in that: The step S600 includes: S610, evaluating the accuracy of the predicted boundary, using a boundary accuracy evaluation index to evaluate the accuracy of the boundary of the articular disc predicted by the articular disc model, and evaluating the degree of overlap between the predicted boundary area of ​​the articular disc and the actual boundary area of ​​the articular disc; S620, predicting structural symmetry evaluation, using a structural symmetry evaluation index to evaluate the symmetry of the articular disc structure predicted by the articular disc model; 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 return to step S200.

Images