A method, system, device, and medium for predicting implant stability status

Abstract

The application relates to an implant stability state prediction method, system, device and medium. The method comprises the following steps: acquiring implant monitoring data, performing feature extraction on the implant monitoring data, and generating a feature time sequence; calculating the distance relationship between feature points in the feature time sequence to generate a manifold adjacency graph; calculating the topological invariant of the manifold adjacency graph to generate a manifold topological feature vector; based on the manifold topological feature vector and the feature time sequence, calculating a manifold topological anomaly index to generate an anomaly index value; and inputting the anomaly index value into a pre-trained stability classification model to output an implant stability prediction result. The method can capture implicit topological distortion signals of high-dimensional continuous manifolds, and realize early and accurate prediction of the implant stability state.

Description

Technical Field

[0001] This invention belongs to the field of oral implant technology, and in particular relates to a method, system, device and medium for predicting implant stability. Background Technology

[0002] With the development of immediate loading surgery techniques in dental implantology, implant stability prediction technologies based on continuous monitoring data have emerged. This technology assesses osseointegration by analyzing dynamic signals during the post-operative healing process, leading to current machine learning-based prediction methods. Traditional techniques typically acquire time-series monitoring data such as implant micromotion and dynamic torque feedback. This data is then directly input into traditional regression networks or classification models as numerical features. The models fit the data based on the trends in numerical changes over time, thus outputting implant stability prediction results.

[0003] However, current prediction methods based on traditional regression networks suffer from prediction lag and insufficient accuracy. Because the bone remodeling process involving the interaction between implants and heterogeneous bone is highly nonlinear, and monitoring data is subject to multipath noise interference during acquisition, high-dimensional continuous data characterizing the mechanical state of the bone integration interface is prone to nonlinear abrupt changes and topological collapse in the feature space. Traditional regression networks can only capture numerical trend changes on the data surface and cannot sensitively perceive and capture these deep-seated implicit topological distortion signals before implant micro-loosening occurs. This makes it difficult for the model to provide early warning of abrupt changes in the mechanical state, thus hindering accurate prediction of stability. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, system, device, and medium for predicting implant stability status that can capture high-dimensional manifold implicit topological distortion signals to accurately predict implant stability in advance, in order to address the above-mentioned technical problems.

[0005] In a first aspect, this application provides a method for predicting implant stability, including:

[0006] S1. Obtain implant monitoring data, extract features from the implant monitoring data, and generate feature time series;

[0007] S2. Calculate the distance relationships between feature points in the feature time series and generate a manifold adjacency graph;

[0008] S3. Calculate the topological invariants of the adjacency graph of the manifold and generate the topological feature vector of the manifold;

[0009] S4. Based on the manifold topological feature vector and feature time series, calculate the manifold topological anomaly index and generate the anomaly index value;

[0010] S5. Input the mutation index values ​​into the pre-trained stability classification model and output the implant stability prediction results.

[0011] In one embodiment, S2 includes:

[0012] S21. Calculate the kernel distance between any two feature points in the feature time series and generate a distance matrix;

[0013] S22. The distance matrix is ​​filtered using the nearest neighbor algorithm to generate an adjacency matrix;

[0014] S23. Construct a manifold adjacency graph based on the adjacency matrix.

[0015] In one embodiment, S3 includes:

[0016] S31. Based on the adjacency graph of the manifold, construct a simple complex structure and generate the boundary matrix;

[0017] S32. Perform reduction calculations on the boundary matrix to generate a persistent graph;

[0018] S33. Based on the persistent graph, extract the Betti number sequence and the length of the persistent interval to generate the topological feature vector of the manifold.

[0019] In one embodiment, S4 includes:

[0020] S401. Calculate the difference in Betti numbers in the manifold topological eigenvectors at the current time and historical time, and generate the change value of Betti numbers;

[0021] S402. Calculate the relative rate of change of the persistence feature in the manifold topological feature vector at the current time and the historical time, and generate the persistence rate of change.

[0022] S403. A weighted sum of the changes in Betti number and the rate of change in durability is performed to generate a manifold topological intrinsic anomaly index. This manifold topological intrinsic anomaly index is then used as the anomaly index. The formula for calculating the manifold topological intrinsic anomaly index is as follows:

[0023]

[0024] in, The value represents the intrinsic anomaly index of the manifold topology. The change in Betty number, For the first The relative rate of change of each persistence feature The number of features with persistence higher than the baseline threshold. and These are the weighting coefficients.

[0025] In one embodiment, S4 includes:

[0026] S411. Concatenate the manifold topological feature vector and the feature vector corresponding to the feature time series to generate an augmented state vector;

[0027] S412. Perform dynamic prediction on the augmented state vector to generate the predicted augmented state vector;

[0028] S413. Based on the augmented state vector and the residual vector of the predicted augmented state vector, calculate the Mahalanobis distance of the residual vector to generate the manifold prediction residual variability index value. Use the manifold prediction residual variability index value as the variability index value. The calculation formula for the manifold prediction residual variability index value is as follows:

[0029]

[0030] in, The residual variation index value is used to predict the manifold. The residual vector between the augmented state vector and the predicted augmented state vector. This is the transpose of the residual vector between the augmented state vector and the predicted augmented state vector. It is the inverse of the covariance matrix of the residual vector.

[0031] In one embodiment, S5 includes:

[0032] S51. Input the variability index values ​​into the pre-trained stability classification model, perform nonlinear mapping on the variability index values, and generate hidden layer feature vectors.

[0033] S52. Input the hidden layer feature vector into the classification layer of the pre-trained stable classification model, calculate the activation probability of each stability level, and generate the classification probability distribution.

[0034] S53. Solve for the extreme values ​​of the classification probability distribution to generate implant stability prediction results. The implant stability prediction results include low-level warning signals, medium-level warning signals and high-level warning signals.

[0035] In one embodiment, the training method for the pre-trained stability classification model includes:

[0036] S61. Obtain historical implant monitoring data and the corresponding historical stability labels for the historical implant monitoring data;

[0037] S62. Based on historical implant monitoring data, calculate the manifold topological anomaly index and generate historical anomaly index values.

[0038] S63. Input the historical variation index values ​​into the initial classification network model, and perform feature encoding on the historical variation index values ​​to generate historical hidden layer feature vectors.

[0039] S64. Based on the historical hidden layer feature vector, calculate the historical activation probability of each stability level and generate the historical classification probability distribution;

[0040] S65. Based on the historical classification probability distribution and historical stability labels, calculate the cross-entropy loss and generate the loss function value;

[0041] S66. Based on the loss function value, the model parameters of the initial classification network model are updated by gradient backpropagation to generate a pre-trained stable classification model.

[0042] Secondly, this application also provides an implant stability prediction system, comprising:

[0043] The feature extraction module is used to acquire implant monitoring data, extract features from the implant monitoring data, and generate feature time series.

[0044] The manifold graph construction module is used to calculate the distance relationships between feature points in a feature time series and generate a manifold adjacency graph;

[0045] The topological feature calculation module is used to calculate the topological invariants of the adjacency graph of a manifold and generate the topological feature vector of the manifold.

[0046] The anomaly index calculation module is used to calculate the manifold topological anomaly index based on the manifold topological feature vector and feature time series, and generate anomaly index values.

[0047] The stability prediction module is used to input the anomaly index values ​​into the pre-trained stability classification model and output the implant stability prediction results.

[0048] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the implant stability prediction method as described in the first aspect.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the implant stability state prediction method of the first aspect.

[0050] The aforementioned implant stability prediction method, system, device, and medium extract features from implant monitoring data to generate feature time series. Based on the feature time series, a manifold adjacency graph reflecting the mechanical state of the osseointegration interface is constructed. Then, manifold topological invariants are calculated to extract deep topological features. Combined with the feature time series, manifold topological anomaly indices are calculated to quantify nonlinear mutations and topological collapses of the manifold. This allows for the keen detection of implicit topological distortion signals that are difficult to detect by traditional methods before microloosening occurs. Finally, the anomaly indices are input into a pre-trained stability classification model to output stability prediction results. This method overcomes the lag defect caused by traditional regression networks relying solely on numerical trend fitting, and achieves accurate early prediction of implant stability. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating a method for predicting implant stability provided by the present invention;

[0053] Figure 2 A flowchart illustrating a method for generating manifold topological feature vectors in an optional embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the structure of an implant stability prediction system provided by the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] In one embodiment, such as Figure 1 As shown, a method for predicting implant stability is provided. This embodiment illustrates the application of this method to a stability prediction terminal. It is understood that this method can also be applied to a stability prediction server, and further to a system including both a stability prediction terminal and a stability prediction server, and is implemented through the interaction between the two. In this embodiment, the method includes the following steps:

[0057] S1. Obtain implant monitoring data, extract features from the implant monitoring data, and generate feature time series.

[0058] Optionally, the stability prediction terminal can first acquire time-series monitoring data transmitted from the implant monitoring device. The implant monitoring data can include multi-source time-series signals such as implant micro-movement data, dynamic torque feedback data, and interface impedance monitoring data during the postoperative healing period.

[0059] Specifically, the stability prediction terminal can preprocess the implant monitoring data to remove baseline drift and high-frequency noise interference introduced during the acquisition process. Then, the stability prediction terminal can extract features from the preprocessed implant monitoring data. For each time sampling point, it can extract the time-domain statistical features and rate of change features of the signal. The extracted features are arranged in chronological order to generate a feature time series. Each element in the feature time series corresponds to a multi-dimensional feature vector at a single moment.

[0060] S2. Calculate the distance relationship between feature points in the feature time series and generate a manifold adjacency graph.

[0061] Optionally, the stability prediction terminal can treat each feature vector in the feature time series as an independent feature point in the high-dimensional manifold space. The stability prediction terminal can calculate the distance metric between any two feature points in the feature time series to quantify the similarity and proximity of the two feature points in the manifold space. Subsequently, the stability prediction terminal can filter out the neighboring feature points of each feature point based on the calculated distance relationship, construct the correspondence between nodes and edges, and generate a manifold adjacency graph. The manifold adjacency graph can effectively represent the local topological structure of high-dimensional feature points in the manifold space.

[0062] S3. Calculate the topological invariants of the adjacency graph of the manifold and generate the topological feature vector of the manifold.

[0063] Optionally, the stability prediction terminal can extract topological invariants from the topological space corresponding to the constructed manifold adjacency graph using methods based on algebraic topology. Topological invariants are quantitative features that remain unchanged during continuous deformation of the topological space, effectively representing the intrinsic topological structure of high-dimensional data and unaffected by data translation, rotation, and scaling transformation. Subsequently, the stability prediction terminal can arrange and integrate the extracted topological invariants according to preset dimensions to generate a manifold topological feature vector. The manifold topological feature vector can effectively capture changes in the topological structure of high-dimensional data that cannot be represented by traditional numerical features.

[0064] S4. Based on the manifold topological feature vector and feature time series, calculate the manifold topological anomaly index and generate the anomaly index value.

[0065] Optionally, the stability prediction terminal can acquire the manifold topological feature vector at the current moment and the manifold topological feature vector at historical moments, and simultaneously acquire the feature time series at the current moment and the feature time series at historical moments. The stability prediction terminal can quantify the degree of deviation of the manifold topological structure from the normal integration state at the current moment by comparing the differences in the manifold topological feature vectors at different moments and the dynamic prediction residuals of the feature time series, and calculate the manifold topological anomaly index in this way. The manifold topological anomaly index can effectively capture the implicit topological distortion signal that occurs in the high-dimensional feature space. The stability prediction terminal can use the quantized result of the calculated manifold topological anomaly index as the anomaly index value.

[0066] S5. Input the mutation index values ​​into the pre-trained stability classification model and output the implant stability prediction results.

[0067] Optionally, the stability prediction terminal can input the calculated anomaly index values ​​into a pre-trained stability classification model. The pre-trained stability classification model is a deep learning classification model trained in advance based on historical implant monitoring data and corresponding stability labels. It can perform non-linear mapping on the input anomaly index values, extract hidden layer features, and complete classification prediction. The stability prediction terminal can obtain the classification results output by the pre-trained stability classification model and use these results as the implant stability prediction results. The implant stability prediction results can characterize the current stability status of the implant and provide clinicians with early warning references for postoperative status.

[0068] In the aforementioned method for predicting implant stability, feature time series are extracted from implant monitoring data, a manifold adjacency graph is constructed to extract topological invariants, and then manifold topological anomaly indices are quantified. Finally, stability prediction is completed through a classification model. This method breaks through the limitation of traditional regression networks that can only capture numerical trends. It can keenly capture implicit topological distortion signals in high-dimensional feature space, realize early warning before implant microloosening occurs, effectively solve the problems of prediction lag and insufficient accuracy of traditional prediction methods, improve the accuracy and timeliness of implant stability prediction, and provide reliable technical support for postoperative monitoring of immediate loading surgery in oral implantology.

[0069] In one embodiment, S2 may include:

[0070] S21. Calculate the kernel distance between any two feature points in the feature time series and generate a distance matrix.

[0071] Optionally, the stability prediction terminal can traverse all feature points in the feature time series. For each pair of feature points, the stability prediction terminal can map the pair of feature points to a high-dimensional kernel space using a preset kernel function. The stability prediction terminal can calculate the kernel distance of the pair of feature points in the high-dimensional kernel space. Kernel distance is a distance metric that can effectively measure the similarity of samples in a nonlinear feature space. Compared with the traditional Euclidean distance, it can better adapt to the nonlinear structure of the manifold space. The formula for calculating the kernel distance is:

[0072]

[0073] in, For the first Feature points of feature points With feature points The distance between the nuclei For the first characteristic time series One feature point, For the first characteristic time series One feature point, This is the preset kernel function.

[0074] Optionally, the kernel function can be a Gaussian kernel function or a polynomial kernel function to achieve the mapping from low-dimensional features to a high-dimensional kernel space. The stability prediction terminal can arrange the kernel distance calculation results of all pairwise feature points into a two-dimensional matrix according to the order of the feature points. The rows and columns of the matrix correspond to the feature points in the feature time series, and the elements in the matrix are the kernel distances between two feature points at corresponding positions, thereby generating a distance matrix.

[0075] S22. Using the nearest neighbor algorithm, filter the distance matrix to generate an adjacency matrix.

[0076] Optionally, the stability prediction terminal can obtain a distance matrix, which stores the kernel distance calculation results for all pairwise feature points. For each feature point, the stability prediction terminal can filter out a predetermined number of nearest neighbor feature points based on the distance values ​​in the distance matrix. This filtering process is based on the nearest neighbor algorithm, a classic algorithm that filters neighboring samples using distance metrics. It effectively preserves local adjacency relationships in the manifold space while filtering out invalid connections at great distances, reducing the complexity of subsequent topological feature calculations. The stability prediction terminal can construct a two-dimensional adjacency matrix based on the filtered adjacency relationships. The rows and columns of the adjacency matrix correspond to feature points in the feature time series, and the elements in the adjacency matrix are used to indicate whether there is an adjacency connection between two corresponding feature points.

[0077] S23. Construct a manifold adjacency graph based on the adjacency matrix.

[0078] Optionally, the stability prediction terminal can obtain an adjacency matrix, which stores the adjacency connections between all feature points. The stability prediction terminal can use each feature point in the feature time series as a node in the manifold adjacency graph. Based on the adjacency connections in the adjacency matrix, the stability prediction terminal can add corresponding edges to two nodes that are adjacent, thereby converting the numerical relationships in the adjacency matrix into the correspondence between nodes and edges in the graph structure, thus completing the construction of the manifold adjacency graph. This construction method can represent the adjacency relationships of high-dimensional feature points with a standardized graph structure, providing a unified input structure for subsequent topological feature extraction and avoiding the computational inconvenience caused by high-dimensional unstructured data.

[0079] In the above embodiments, a distance matrix is ​​generated by calculating the kernel distance between feature points, and then an adjacency matrix is ​​obtained by filtering using the nearest neighbor algorithm, thus completing the construction of the manifold adjacency graph. This embodiment replaces the traditional Euclidean distance with kernel distance, which can better adapt to the distance metric of high-dimensional nonlinear feature space, effectively capture the true similarity of feature points in manifold space, and filter out invalid long-distance connections by filtering effective adjacency relationships through the nearest neighbor algorithm. While preserving the local topological structure of the manifold space, it reduces the complexity of subsequent calculations and effectively improves the accuracy and computational efficiency of topological feature extraction.

[0080] In an optional embodiment, such as Figure 2 As shown, S3 may include:

[0081] S31. Based on the adjacency graph of the manifold, construct a simple complex structure and generate the boundary matrix.

[0082] Optionally, the stability prediction terminal can obtain a manifold adjacency graph, which stores the relationships between nodes and adjacent edges of feature points. Based on the adjacency relationships of the manifold adjacency graph, the stability prediction terminal can construct a simplex complex. A simplex complex is a topological space structure composed of simplexes of different dimensions according to the closed subset rule. The 0-dimensional simplex corresponds to the nodes of the manifold adjacency graph, the 1-dimensional simplex corresponds to the edges of the manifold adjacency graph, and the higher-dimensional simplex is extended according to the pairwise adjacency rule. That is, when there are adjacent edges between every pair of k nodes, a k-1 dimensional simplex can be constructed.

[0083] Furthermore, the stability prediction terminal can generate a boundary matrix for the constructed simplex. The rows of the boundary matrix correspond to the simplex of the lower dimension, and the columns correspond to the simplex of the current dimension. The elements in the boundary matrix are used to mark whether the lower-dimensional simplex is the boundary surface of the simplex of the current dimension.

[0084] S32. Perform reduction calculations on the boundary matrix to generate a persistent graph.

[0085] Optionally, the stability prediction terminal can obtain the boundary matrix and perform standard Gaussian reduction calculation on the boundary matrix. This reduction calculation is the core of persistent homology calculation. Persistent homology is a classic topological data analysis method that can extract topological features at multiple scales and can effectively capture topological changes at different scales.

[0086] Specifically, during the reduction calculation, the stability prediction terminal can traverse each column of the boundary matrix, find the bottom non-zero pivot element of that column, and eliminate the non-zero elements at the corresponding pivot positions in other columns through column transformations until the pivot positions of all columns remain unique, thus completing the reduction of the boundary matrix. Based on the reduced boundary matrix, the stability prediction terminal can extract the birth and death scales of each topological feature, and organize these scale values ​​into a persistent graph. Each point in the persistent graph corresponds to a topological feature, and the horizontal and vertical coordinates of the point correspond to the birth and death scales of the topological feature, respectively.

[0087] S33. Based on the persistent graph, extract the Betti number sequence and the length of the persistent interval to generate the topological feature vector of the manifold.

[0088] Optionally, the stability prediction terminal can acquire a persistent graph. It can extract Betti numbers for each dimension for different filtering scales. Betti numbers are topological invariants representing the number of topological features in a corresponding dimension, where 0-dimensional Betti numbers represent the number of connected components, 1-dimensional Betti numbers represent the number of cycles, and 2-dimensional Betti numbers represent the number of cavities. The stability prediction terminal can arrange the Betti numbers at different scales in chronological order to generate a Betti number sequence. The stability prediction terminal can also extract the persistence interval length for each topological feature in the persistent graph. The formula for calculating the persistence interval length is:

[0089]

[0090] in, The length of the persistent interval of the topological feature. The death scale for topological features. The birth scale of the topological feature is used, and the persistence interval length is used to characterize the stability of the topological feature. The larger the persistence interval length, the more stable the topological feature is, and the more likely it is a real topological structure rather than a pseudo-feature caused by noise.

[0091] Furthermore, the stability prediction terminal can arrange and integrate the Betti number sequence and the persistence interval length according to a preset dimension to generate a manifold topological feature vector.

[0092] In the above embodiments, a simple complex and boundary matrix are constructed through a manifold adjacency graph. Then, a persistent graph is generated by reducing the boundary matrix. Finally, topological features are extracted to generate manifold topological feature vectors. This embodiment is based on persistent homology technology, which can effectively extract topological invariants at multiple scales, capture the intrinsic topological structure of high-dimensional data, and is not affected by data translation, rotation, and scaling transformation. At the same time, stable topological features are selected by persistent interval length, effectively filtering out false features caused by noise. This provides a reliable feature basis for topological anomaly detection, effectively solving the problem that traditional methods cannot capture implicit topological changes, and improving the accuracy and robustness of topological features.

[0093] In one embodiment, S4 may include:

[0094] S401. Calculate the difference in Betti numbers in the manifold topological eigenvectors at the current time and historical time, and generate the change value of Betti numbers.

[0095] Optionally, the stability prediction terminal can obtain the manifold topological feature vector at the current moment, and at the same time obtain the manifold topological feature vector at a preset historical reference moment. The preset historical reference moment can be the moment when the osseointegration state is stable in the early stage after implantation, or the moment corresponding to the average manifold topological feature vector within the sliding time window.

[0096] Specifically, the stability prediction terminal can extract 1D Betti numbers from the manifold topological feature vector at the current moment. The 1D Betti number characterizes the number of loop structures in the manifold space and effectively reflects changes in the connection structure of the implant-bone interface during osseointegration. The stability prediction terminal can also extract corresponding 1D Betti numbers from the manifold topological feature vector at a preset historical baseline moment, calculate the difference between the current 1D Betti number and the historical baseline 1D Betti number, and generate a Betti number change value. This change value directly quantifies the degree of change in the number of loops in the manifold topology, reflecting the overall change in the connection structure of the osseointegration interface.

[0097] S402. Calculate the relative rate of change of persistence features in the manifold topological feature vectors at the current time and historical time, and generate the persistence rate of change.

[0098] Optionally, the stability prediction terminal can extract all persistence features from the manifold topological feature vector at the current time. Each persistence feature corresponds to the persistence interval length of each topological feature stored in the manifold topological feature vector, and the persistence interval length characterizes the stability of the topological feature. The stability prediction terminal can obtain a predefined baseline threshold, which is statistically obtained in advance based on a large number of manifold topological feature vectors under normal fusion states. This baseline threshold is used to filter out short-persistence pseudo-features caused by noise interference. The stability prediction terminal can select all persistence features with persistence higher than the baseline threshold. For each selected persistence feature, it extracts the persistence value corresponding to the historical baseline time, calculates the relative rate of change between the persistence value at the current time and the persistence value at the historical baseline time, and generates a persistence change rate. The relative rate of change can eliminate the influence of differences in persistence levels among different topological features, more accurately reflecting the degree of stability change of individual topological features.

[0099] S403. A weighted sum of the changes in Betti number and the rate of change in durability is performed to generate a manifold topological intrinsic anomaly index. This manifold topological intrinsic anomaly index is then used as the anomaly index. The formula for calculating the manifold topological intrinsic anomaly index is as follows:

[0100]

[0101] in, The value represents the intrinsic anomaly index of the manifold topology. The change in Betty number, For the first The relative rate of change of each persistence feature The number of features with persistence higher than the baseline threshold. and These are the weighting coefficients.

[0102] Optionally, the stability prediction terminal can obtain the generated Betti number change value and the persistence change rate corresponding to all the filtered persistence features, take the absolute value of the Betti number change value, and take the absolute value of each persistence change rate. The absolute value operation can uniformly quantify the magnitude of the change and is not affected by the direction of the change.

[0103] Furthermore, the stability prediction terminal performs a weighted summation of the absolute values ​​of the processed Betti number change and all persistence change rates to generate a manifold topological intrinsic anomaly index, which is then used as the anomaly index. The formula for calculating the manifold topological intrinsic anomaly index is as follows:

[0104]

[0105] in, The value represents the intrinsic anomaly index of the manifold topology. The change in Betty number, For the first The relative rate of change of each persistence feature The number of features with persistence higher than the baseline threshold. and The weighting coefficients are determined in advance based on historical clinical data using cross-validation methods, and are used to balance the contributions of changes in Betty number and persistence to the overall degree of variability.

[0106] In the above embodiments, by extracting the core topological invariants from the manifold topological feature vectors, calculating the absolute change value of the Betti number and the relative change rate of the durability feature, and then comprehensively quantifying the intrinsic anomaly degree of the manifold topological structure through weighted summation, an intrinsic anomaly index value of the manifold topology is generated. This embodiment directly detects anomalies based on the changes in the topological invariants themselves. The calculation process is simple and efficient, and it can capture subtle changes in the manifold topology during osseointegration in real time. It can identify implicit topological distortion signals before implant microloosening occurs in advance, effectively improving the real-time performance and accuracy of anomaly detection, and providing a reliable quantitative basis for early warning of implant stability.

[0107] In one embodiment, S4 may include:

[0108] S411. Concatenate the manifold topological feature vector and the feature vector corresponding to the feature time series to generate an augmented state vector.

[0109] Optionally, the stability prediction terminal can acquire the manifold topological feature vector at the current moment, and simultaneously acquire the feature vector corresponding to the feature time series at the current moment. The feature vector corresponding to the feature time series is a multi-dimensional numerical feature vector corresponding to the current moment in the feature time series. The stability prediction terminal can concatenate all elements of the manifold topological feature vector with all elements of the feature vector corresponding to the feature time series in a preset order, that is, concatenate the elements of the two vectors end to end, and merge them into a higher-dimensional vector to generate an augmented state vector. The augmented state vector can simultaneously integrate the original time-series numerical features and the extracted topological features, more comprehensively representing the current state of the implant monitoring system, providing more complete input information for subsequent dynamic prediction, and avoiding information loss caused by single features.

[0110] S412. Perform dynamic prediction on the augmented state vector to generate the predicted augmented state vector.

[0111] Optionally, the stability prediction terminal can acquire augmented state vector sequences within a historical time period. The stability prediction terminal can pre-build a dedicated training dataset for the dynamic prediction model and complete model training, and input the historical augmented state vector sequences into the trained dynamic prediction model to output prediction results.

[0112] The training dataset only selects historical augmented state vector sequences containing data under normal osseointegration conditions, excluding all data samples labeled as abnormal states. This ensures the model learns only the evolution of augmented state vectors during normal osseointegration. The stability prediction terminal initializes the parameters of the dynamic prediction model, calculates the error between the predicted and actual augmented state vectors using the mean squared error loss function, and iteratively updates the model parameters using an adaptive moment estimation optimizer combined with gradient backpropagation. During training, a sliding window validation method is used to divide the training and validation sets. Training terminates when the validation set loss no longer decreases after several consecutive rounds, and the model parameters with the lowest validation loss are saved as the final dynamic prediction model parameters. This ensures the model's fitting accuracy to the evolution of normal states, enabling it to learn the evolution of augmented state vectors under normal osseointegration conditions and capture the temporal change patterns of state vectors under normal conditions.

[0113] Furthermore, the stability prediction terminal can obtain the prediction results output by the dynamic prediction model and use the prediction results as the prediction augmented state vector. The prediction augmented state vector represents the state that the system should reach at the current moment under normal and anomaly-free integration state.

[0114] S413. Based on the augmented state vector and the residual vector of the predicted augmented state vector, calculate the Mahalanobis distance of the residual vector to generate the manifold prediction residual variability index value. Use the manifold prediction residual variability index value as the variability index value. The calculation formula for the manifold prediction residual variability index value is as follows:

[0115]

[0116] in, The residual variation index value is used to predict the manifold. The residual vector between the augmented state vector and the predicted augmented state vector. This is the transpose of the residual vector between the augmented state vector and the predicted augmented state vector. It is the inverse of the covariance matrix of the residual vector.

[0117] Optionally, the stability prediction terminal can calculate the difference between the augmented state vector and the predicted augmented state vector to generate a residual vector. This residual vector characterizes the deviation between the actual state and the predicted state. The stability prediction terminal can also calculate the Mahalanobis distance of the residual vector. Mahalanobis distance is an effective measure of the distance between a vector and the sample distribution, eliminating the correlation and dimensionality effects between feature dimensions. Compared to Euclidean distance, it can more accurately identify anomalous deviations. The formula for calculating the manifold prediction residual variation index is:

[0118]

[0119] in, The residual variation index value is used to predict the manifold. The residual vector between the augmented state vector and the predicted augmented state vector. This is the transpose of the residual vector between the augmented state vector and the predicted augmented state vector. The covariance matrix is ​​the inverse of the residual vector. This covariance matrix is ​​pre-calculated based on the residual vectors under historical normal conditions and is used to characterize the distribution characteristics of the residuals under normal conditions. The stability prediction terminal can use the calculated Mahalanobis distance as the manifold prediction residual variation index value, and use the manifold prediction residual variation index value as the anomaly index value.

[0120] In the above embodiments, an augmented state vector is generated by splicing topological features and temporal features. Then, the predicted augmented state vector under normal conditions is obtained through dynamic prediction. Finally, the Mahalanobis distance of the residuals is calculated to obtain the manifold prediction residual variation index. This embodiment integrates topological features and original temporal features, which can more comprehensively characterize the state of the implant monitoring system. At the same time, it makes predictions based on the evolution law of the normal state and quantifies the deviation between the actual state and the normal state through residual distance. It effectively captures implicit anomalies outside the normal evolution and can identify the changes in the predicted residuals caused by topological distortion in advance. It provides a reliable anomaly quantification index for stability early warning and effectively improves the accuracy and timeliness of anomaly detection.

[0121] In one embodiment, S5 may include:

[0122] S51. Input the variability index values ​​into the pre-trained stability classification model, perform nonlinear mapping on the variability index values, and generate hidden layer feature vectors.

[0123] Optionally, the stability prediction terminal can input the calculated mutation index values ​​into a pre-trained stability classification model. The pre-trained stability classification model can employ a multilayer perceptron structure, including an input layer, multiple hidden layers, and a classification layer. The stability prediction terminal can convert the mutation index values ​​into a vector form that the model can process through the input layer. Then, the converted vectors are sequentially input into each hidden layer. Each hidden layer first performs a linear transformation on the input vector. This linear transformation is achieved through matrix multiplication of the weight matrix and the input vector, and addition of the bias term. The result of the linear transformation is then input into an activation function for nonlinear mapping. The activation function can be a ReLU (Rectified Linear Unit), which effectively solves the gradient vanishing problem and improves the model's training efficiency and expressive power. After continuous nonlinear mapping through multiple hidden layers, the stability prediction terminal obtains hidden layer feature vectors that characterize the deep abstract features of the mutation index values. Compared to the original mutation index values, these hidden layer feature vectors better capture the subtle differences between different stability levels and have stronger class discrimination.

[0124] S52. Input the hidden layer feature vector into the classification layer of the pre-trained stable classification model, calculate the activation probability of each stability level, and generate the classification probability distribution.

[0125] Optionally, the stability prediction terminal can input the generated hidden layer feature vector into the classification layer of a pre-trained stability classification model. The classification layer is a fully connected layer, and its output dimension matches the preset number of stability levels, corresponding to three levels: low-level warning signal, medium-level warning signal, and high-level warning signal. The stability prediction terminal first performs matrix multiplication between the hidden layer feature vector and the weight matrix of the classification layer, and then adds the bias term of the classification layer to obtain the original output value for each stability level.

[0126] Furthermore, the stability prediction terminal can input the raw output values ​​of all stability levels into the Softmax function for normalization. The Softmax function can convert any real number vector into a probability distribution with values ​​between 0 and 1, where the sum of all elements is 1. After processing by the Softmax function, the stability prediction terminal obtains the activation probability of each stability level. These activation probabilities are arranged according to a preset order of stability levels to generate a classification probability distribution. Each element in the classification probability distribution corresponds to a predicted probability of a stability level. The higher the probability value, the greater the likelihood that the implant's current state belongs to that stability level.

[0127] S53. Solve for the extreme values ​​of the classification probability distribution to generate implant stability prediction results. The implant stability prediction results include low-level warning signals, medium-level warning signals and high-level warning signals.

[0128] Optionally, the stability prediction terminal can obtain the generated classification probability distribution, and the stability prediction terminal can solve for the extreme values ​​of the classification probability distribution. The extreme value solution is achieved by taking the maximum value index operation.

[0129] Specifically, the stability prediction terminal can traverse all probability values ​​in the classification probability distribution, find the highest probability value, and record the index position corresponding to that highest probability value. Based on a preset mapping relationship between indices and stability levels, the terminal maps the index corresponding to the highest probability value to the corresponding implant stability prediction result. In this preset mapping relationship, different indices correspond to low-level, medium-level, and high-level warning signals, respectively. A low-level warning signal indicates a slight decrease in implant stability and mild abnormalities in the osseointegration process; a medium-level warning signal indicates a moderate decrease in implant stability and significant abnormalities in the osseointegration process; a high-level warning signal indicates a severe decrease in implant stability and a high risk of microloosening. The stability prediction terminal can output the mapped warning signal as the final implant stability prediction result.

[0130] In the above embodiments, a hierarchical nonlinear mapping of the mutation index values ​​is performed through a pre-trained multilayer perceptron structure to extract hidden layer feature vectors with strong class discrimination. Then, a normalized probability distribution for each stability level is generated through the Softmax function. Finally, the prediction results in the form of a three-level warning are obtained by solving the extreme values. This embodiment can effectively learn the complex nonlinear relationship between the mutation index values ​​and the implant stability level. The output in the form of probability distribution provides confidence information of the prediction results. The three-level warning mechanism can provide clinicians with clear references for different degrees of severity, improving the interpretability and clinical applicability of the prediction results. At the same time, the inference process of the pre-trained model has high computational efficiency and can quickly output stable and reliable prediction results, meeting the application needs of real-time monitoring after oral implant surgery.

[0131] In one embodiment, the training method for the pre-trained stability classification model may include:

[0132] S61. Obtain historical implant monitoring data and the corresponding historical stability labels for the historical implant monitoring data.

[0133] Optionally, the stability prediction terminal can obtain historical implant monitoring data from a pre-built clinical implant database. This historical data can include multi-source time-series signals such as implant micromotion data, dynamic torque feedback data, and interface impedance monitoring data collected during the osseointegration period of multiple oral implant surgeries. The historical implant monitoring data for each patient is standardized according to a uniform sampling frequency and acquisition duration. The stability prediction terminal can simultaneously obtain historical stability labels corresponding to the historical implant monitoring data. These labels are assigned by experienced dental clinicians based on postoperative follow-up results, imaging examination results, and the final osseointegration status of the implant. Historical stability labels are divided into three categories, corresponding to low-level, medium-level, and high-level warning signals, respectively. Each historical implant monitoring data point corresponds to a unique historical stability label, thereby constructing a supervised dataset for model training.

[0134] S62. Based on historical implant monitoring data, calculate the manifold topological anomaly index and generate historical anomaly index values.

[0135] Optionally, the stability prediction terminal can process each historical implant monitoring data in the same way as the real-time prediction stage. That is, firstly, feature extraction is performed on the historical implant monitoring data to generate a feature time series, then the distance relationship between feature points in the feature time series is calculated to generate a manifold adjacency graph, then the topological invariants of the manifold adjacency graph are calculated to generate a manifold topological feature vector, and finally the manifold topological anomaly index is calculated based on the manifold topological feature vector and the feature time series.

[0136] Optionally, the stability prediction terminal can use the quantification result of the manifold topological variation index corresponding to each historical implant monitoring data as the historical variation index value. The historical variation index values ​​corresponding to all historical implant monitoring data together constitute the feature dataset for model training, and the feature dataset corresponds one-to-one with the supervised dataset.

[0137] S63. Input the historical variation index values ​​into the initial classification network model, and perform feature encoding on the historical variation index values ​​to generate historical hidden layer feature vectors.

[0138] Optionally, the stability prediction terminal can initialize an initial classification network model. The initial classification network model adopts a multilayer perceptron structure, which includes an input layer, multiple hidden layers, and a classification layer. All parameters of the initial classification network model are set using a random initialization method.

[0139] Optionally, parameter initialization can employ Xavier initialization or He initialization methods to ensure a reasonable distribution of initial parameters and avoid gradient vanishing or exploding problems in the early stages of training. The stability prediction terminal can input historical variation index values ​​into the input layer of the initial classification network model. After format conversion, the values ​​are sequentially encoded through each hidden layer. The feature encoding process is achieved through the continuous action of linear transformations and nonlinear activation functions, ultimately generating historical hidden layer feature vectors that characterize the deep features of historical variation index values, completing the first step of the forward propagation process.

[0140] S64. Based on the historical hidden layer feature vector, calculate the historical activation probability of each stability level and generate the historical classification probability distribution.

[0141] Optionally, the stability prediction terminal can input the generated historical hidden layer feature vectors into the classification layer of the initial classification network model. The classification layer first performs a linear transformation on the historical hidden layer feature vectors to obtain the original output value for each stability level. Then, all the original output values ​​are input into the Softmax function for normalization. The Softmax function converts the original output values ​​into probability values ​​between 0 and 1, with the sum of all elements being 1. The stability prediction terminal can then arrange these probability values ​​according to a preset stability level order to generate a historical classification probability distribution. Each element in the historical classification probability distribution represents the probability that the initial classification network model, under the current parameters, predicts a certain stability level corresponding to that historical anomaly index value. The historical classification probability distribution reflects the model's current predictive ability.

[0142] S65. Based on the historical classification probability distribution and historical stability labels, calculate the cross-entropy loss and generate the loss function value.

[0143] Optionally, the stability prediction terminal can obtain historical classification probability distributions and corresponding historical stability labels. The terminal can then convert these historical stability labels into one-hot encoded forms. One-hot encoding is a method of converting discrete category labels into vector form, where the vector dimension matches the number of stability levels, with elements corresponding to the true label being 1 and all other elements being 0. Based on the historical classification probability distributions and the one-hot encoded historical stability labels, the stability prediction terminal can calculate the cross-entropy loss. The formula for calculating the cross-entropy loss is:

[0144]

[0145] in, This represents the cross-entropy loss value. The number of stability levels, The historical stability tag in one-hot encoding form The value of each element, The first in the historical classification probability distribution The value of each element. Cross-entropy loss can effectively measure the difference between the predicted probability distribution and the true label distribution. The smaller the loss value, the closer the model's prediction result is to the true label. The stability prediction terminal can use the calculated cross-entropy loss value as the loss function value.

[0146] S66. Based on the loss function value, the model parameters of the initial classification network model are updated by gradient backpropagation to generate a pre-trained stable classification model.

[0147] Optionally, the stability prediction terminal can calculate the gradient of the loss function with respect to all trainable parameters in the initial classification network model based on the loss function value using the gradient backpropagation algorithm. The gradient backpropagation algorithm calculates the gradient layer by layer from the output layer to the input layer using the chain rule, obtaining the contribution of each parameter to the loss function. The stability prediction terminal can then use a preset optimizer to update the model parameters based on the calculated gradients.

[0148] Alternatively, the optimizer can be an adaptive moment estimator (IME), which can adaptively adjust the learning rate of each parameter to accelerate the convergence of the model.

[0149] Furthermore, the stability prediction terminal can repeatedly execute the forward propagation to calculate the loss and the backpropagation to update the parameters until the preset training termination condition is met. The training termination condition can be reaching a preset number of iterations or the loss function value converging to a preset threshold range. At this point, the initial classification network model training is complete, and the stability prediction terminal can use the trained model as a pre-trained stability classification model.

[0150] In the above embodiments, a supervised dataset containing historical implant monitoring data and corresponding clinical labels is constructed. Historical anomaly index values ​​are calculated as training features according to a unified process. A multilayer perceptron structure is used for forward propagation to obtain the predicted probability distribution. The prediction error is quantified using cross-entropy loss, and the gradient backpropagation algorithm is used to iteratively optimize the model parameters, ultimately obtaining a pre-trained stability classification model. This model training uses real clinical data as supervision signals, enabling the model to accurately learn the complex nonlinear mapping relationship between manifold topological anomaly indices and implant stability levels. Cross-entropy loss adapts to the optimization objectives of multi-classification tasks, and the gradient backpropagation algorithm efficiently updates model parameters. The trained model has good generalization ability and prediction accuracy, providing reliable model support for real-time prediction of implant stability and effectively ensuring the credibility and clinical applicability of the prediction results.

[0151] In the aforementioned implant stability prediction methods, systems, devices, and media, this approach addresses the shortcomings of traditional implant stability prediction methods that rely solely on numerical trend fitting and cannot capture high-dimensional nonlinear topological distortions. It introduces manifold learning and persistent homology topological data analysis. First, implant monitoring data is used to generate feature time series and construct a manifold adjacency graph. By calculating the topological invariants of the manifold, manifold topological feature vectors characterizing the intrinsic connectivity structure of the osseointegration interface are extracted. Then, the degree of manifold topological anomaly is quantified from two dimensions: intrinsic topological changes and dynamic prediction residuals. Finally, a pre-trained stability classification model outputs graded early warning results. This technical solution overcomes the limitation of traditional regression networks that can only capture surface numerical trends. It can keenly perceive and capture implicit topological distortion signals in the high-dimensional feature space before implant micro-loosening occurs, achieving early warning of implant stability. This effectively solves the technical problems of prediction lag and insufficient accuracy in traditional prediction methods, providing reliable technical support for postoperative safety monitoring of immediate loading surgery in dental implants.

[0152] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0153] Based on the same inventive concept, this application also provides a system for implementing the implant stability prediction method described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more implant stability prediction system embodiments provided below can be found in the limitations of the implant stability prediction method described above, and will not be repeated here.

[0154] In one exemplary embodiment, such as Figure 3 As shown, an implant stability prediction system 10 is provided, comprising:

[0155] Feature extraction module 11 can be used to acquire implant monitoring data, extract features from the implant monitoring data, and generate feature time series.

[0156] The manifold construction module 12 can be used to calculate the distance relationship between feature points in a feature time series and generate a manifold adjacency graph;

[0157] The topological feature calculation module 13 can be used to calculate the topological invariants of the adjacency graph of a manifold and generate the topological feature vector of the manifold.

[0158] The anomaly index calculation module 14 can be used to calculate the manifold topology anomaly index based on the manifold topology feature vector and feature time series, and generate anomaly index values.

[0159] The stability prediction module 15 can be used to input the abnormality index value into the pre-trained stability classification model and output the implant stability prediction result.

[0160] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the implant stability prediction method as described above.

[0161] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the implant stability state prediction method as described above.

[0162] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0163] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting implant stability, characterized in that, The method includes: S1. Obtain implant monitoring data, extract features from the implant monitoring data, and generate a feature time series; S2. Calculate the distance relationship between feature points in the feature time series and generate a manifold adjacency graph; S3. Calculate the topological invariants of the adjacency graph of the manifold and generate the topological feature vector of the manifold; S4. Based on the manifold topological feature vector and the feature time series, calculate the manifold topological anomaly index and generate the anomaly index value; S5. Input the mutation index value into the pre-trained stability classification model and output the implant stability prediction result.

2. The method according to claim 1, characterized in that, S2 includes: S21. Calculate the kernel distance between any two feature points in the feature time series and generate a distance matrix; S22. The distance matrix is ​​filtered using the nearest neighbor algorithm to generate an adjacency matrix; S23. Based on the adjacency matrix, construct the manifold adjacency graph.

3. The method according to claim 1, characterized in that, The S3 includes: S31. Based on the adjacency graph of the manifold, construct a simple complex structure and generate a boundary matrix; S32. Perform reduction calculations on the boundary matrix to generate a persistent graph; S33. Based on the persistent graph, extract the Betti number sequence and the length of the persistent interval to generate the topological feature vector of the manifold.

4. The method according to claim 1, characterized in that, The S4 includes: S401. Calculate the difference in Betti numbers in the topological feature vectors of the manifold at the current time and at historical time, and generate the change value of Betti numbers; S402. Calculate the relative rate of change of the persistence feature in the manifold topological feature vector at the current time and the historical time, and generate the persistence rate of change. S403. A weighted sum is performed on the Betti number change value and the persistence change rate to generate a manifold topological intrinsic anomaly index value, and this manifold topological intrinsic anomaly index value is used as the anomaly index value. The calculation formula for the manifold topological intrinsic anomaly index value is: in, The value of the intrinsic anomaly index of the manifold topology. The change value of the Betty number, For the first The relative rate of change of each of the aforementioned persistence features The number of features whose persistence is higher than the baseline threshold. and These are the weighting coefficients.

5. The method according to claim 1, characterized in that, The S4 includes: S411. Concatenate the manifold topological feature vector and the feature vector corresponding to the feature time series to generate an augmented state vector; S412. Perform dynamic prediction on the augmented state vector to generate a predicted augmented state vector; S413. Based on the augmented state vector and the residual vector of the predicted augmented state vector, calculate the Mahalanobis distance of the residual vector to generate a manifold prediction residual anomaly index value. Use the manifold prediction residual anomaly index value as the anomaly index value. The calculation formula for the manifold prediction residual anomaly index value is as follows: in, For manifold prediction residual variation index values, The residual vector between the augmented state vector and the predicted augmented state vector. This is the transpose of the residual vector between the augmented state vector and the predicted augmented state vector. It is the inverse of the covariance matrix of the residual vector.

6. The method according to claim 1, characterized in that, The S5 includes: S51. Input the mutation index value into the pre-trained stability classification model, perform nonlinear mapping on the mutation index value, and generate hidden layer feature vectors. S52. Input the hidden layer feature vector into the classification layer of the pre-trained stability classification model, calculate the activation probability of each stability level, and generate a classification probability distribution. S53. Solve for the extreme values ​​of the classification probability distribution to generate the implant stability prediction result, which includes low-level warning signals, medium-level warning signals and high-level warning signals.

7. The method according to claim 6, characterized in that, The training method for the pre-trained stability classification model includes: S61. Obtain historical implant monitoring data and the historical stability labels corresponding to the historical implant monitoring data; S62. Based on the historical implant monitoring data, calculate the manifold topological anomaly index and generate historical anomaly index values. S63. Input the historical variation index values ​​into the initial classification network model, and perform feature encoding on the historical variation index values ​​to generate historical hidden layer feature vectors. S64. Based on the historical hidden layer feature vector, calculate the historical activation probability of each stability level and generate the historical classification probability distribution; S65. Based on the historical classification probability distribution and the historical stability label, calculate the cross-entropy loss and generate the loss function value; S66. Based on the loss function value, the model parameters of the initial classification network model are updated by gradient backpropagation to generate the pre-trained stable classification model.

8. A system for predicting implant stability, characterized in that, The system includes: The feature extraction module is used to acquire implant monitoring data, extract features from the implant monitoring data, and generate feature time series. The manifold graph construction module is used to calculate the distance relationship between feature points in the feature time series and generate a manifold adjacency graph; The topological feature calculation module is used to calculate the topological invariants of the adjacency graph of the manifold and generate the topological feature vector of the manifold. The anomaly index calculation module is used to calculate the manifold topology anomaly index based on the manifold topology feature vector and the feature time series, and generate anomaly index values. The stability prediction module is used to input the mutation index values ​​into a pre-trained stability classification model and output the implant stability prediction results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images