Intelligent evaluation method and system for implant bone integration state

Abstract

The application relates to the technical field of computer-aided diagnosis, and discloses an intelligent evaluation method and system for implant bone integration state, wherein discrete physical measurement data and cone-beam computed tomography images are cross-modally fused, the problems of metal artifact interference and sparse clinical observation data in the prior art are effectively solved, a calibration mechanism based on material mechanics constitutive consistency is established by using a physical stability distribution tensor as a true value benchmark, non-physiological artifact noise caused by high-density metal can be distinguished and removed from the physical nature, misjudgment caused by image visual fraud is avoided, and meanwhile, an anisotropic spatial distance weight matrix is constructed, the model is forced to focus on an effective bone combination interface, and the interference of far-field background noise is shielded.

Description

Technical Field

[0001] This application relates to the field of computer-aided diagnostic technology, and in particular to a method and system for intelligent assessment of implant osseointegration status. Background Technology

[0002] Dental implant technology has become the preferred solution for modern restoration of missing teeth. The direct and stable structural and functional connection between the implant and the surrounding bone tissue—osseointegration—is the biological basis for the long-term success of implant restoration. During the healing period and long-term maintenance after implantation, to comprehensively and accurately monitor the dynamic evolution of osseointegration, clinical settings typically require the collection and comprehensive analysis of monitoring data from multiple sources. Imaging techniques such as cone-beam computed tomography (CBCT) or periapical radiographs are primarily used to obtain two-dimensional or three-dimensional images of the bone tissue surrounding the implant, directly reflecting the anatomical changes in trabecular bone microstructure, bone density distribution, and marginal bone height. Resonance frequency analysis technology sends magnetic pulses to the implant and measures its echo frequency, outputting a stability coefficient representing the implant's micromovement, objectively quantifying the implant's stability from a physical and mechanical perspective. Furthermore, clinical examination indicators such as periodontal probing depth and bleeding index are key discrete data reflecting the inflammatory state of the soft tissues surrounding the implant and the integrity of its biological seal. This multi-dimensional monitoring model, encompassing anatomical morphology, physical and mechanical properties, and biological inflammatory responses, forms the data foundation for achieving proactive health management of implants throughout their entire lifecycle.

[0003] However, despite the widespread collection of multimodal data in clinical practice, existing computer-aided assessment techniques primarily focus on the independent processing and analysis of single-modal data. Current image analysis software often concentrates solely on lesion identification or density measurement at the image level, while stability measurement devices are merely used as independent numerical display tools, resulting in significant information barriers between data streams from different modalities. In the absence of intelligent fusion analysis methods, the final assessment of bone integration status still heavily relies on clinicians' subjective summarization of fragmented data based on personal experience, making it highly susceptible to human error and misjudgment. From a deeper technical perspective of data processing, these three types of data possess completely different physical properties and data structures. Image data is represented as a high-dimensional matrix containing millions of pixels, resonance frequency values ​​are represented as a single one-dimensional scalar, while clinical examination records are mostly discrete text symbols or categorical variables. These three types exhibit significant dimensionality curse and semantic gaps in their feature spaces. If conventional feature stitching methods are used to directly input the raw data into a neural network model, the massive amount of image feature information often overwhelms the low-dimensional physical numerical features, causing the algorithm model to lose sensitivity to key mechanical stability indicators. Current pattern recognition technologies are still unable to construct a unified feature embedding space to achieve semantic alignment and joint representation of heterogeneous data. They cannot effectively reveal the complex nonlinear causal logic between microscopic changes in bone density and the decrease in physical stability values ​​shown in images. As a result, when faced with complex and hidden cases such as poor bone quality but relatively good initial stability values, the system is unable to make accurate objective quantitative assessments and timely risk warnings. Summary of the Invention

[0004] This application proposes an intelligent assessment method and system for the osseointegration status of implants to address the problems raised in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: an intelligent assessment method for implant osseointegration status, comprising the following steps:

[0006] Step S1: Obtain postoperative cone-beam CT image data and discrete physical measurement data of the subject, perform signal normalization on the cone-beam CT image data, extract the anatomical structure feature tensor representing the bone microstructure through a deep convolutional network, establish coordinate mapping between discrete physical measurement data and image space, convert the discrete physical measurement data into physical state vectors and project them onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space.

[0007] Step S2: Calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds the preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise.

[0008] Step S3: Calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface.

[0009] Step S4: Construct a feature evolution sequence based on multi-time point physiological-anatomical fusion features, input a continuous-time dynamic model to fit the biological healing rate function of bone integration, solve the manifold coordinate position of the current node in the healing trajectory and map it to the probability of successful bone integration.

[0010] Furthermore, in step S1, the specific operation of performing signal intensity normalization processing on the cone-beam CT image data is as follows:

[0011] In cone-beam CT image data, a region of interest containing the implant and surrounding bone tissue is defined. The statistical distribution characteristics of all voxel gray values ​​within the region of interest are calculated to determine the first and third quartile values. Based on the first and third quartile values, the interquartile range is calculated. Extremely bright gray values ​​exceeding a preset multiple of the interquartile range are identified as non-physiological artifact signals generated by the metal beam hardening effect. Nonlinear compression processing is performed on the non-physiological artifact signals to limit their interference with the overall statistical distribution characteristics. Subsequently, the overall voxel gray values ​​after nonlinear compression are linearly mapped to a dimensionless range of zero to one to generate normalized image data that has removed the baseline drift of the acquisition device.

[0012] The specific operation of extracting anatomical structural feature tensors representing the microstructure of trabecular bone through a deep convolutional feature extraction network is as follows: normalized image data is input into a fully convolutional neural network with the fully connected layers removed. An encoder containing dilated convolution operations is used to expand the receptive field while maintaining the spatial resolution of deep features. Multi-channel feature maps representing the micro-connectivity of trabecular bone and the thickness gradient of cortical bone are extracted through multi-layer convolution operations, generating anatomical structural feature tensors that encode micro-anatomical morphological information in the channel dimension.

[0013] Furthermore, in step S1, the specific operation of converting discrete physical measurement data into a continuous physical state vector and projecting it onto the spatial coordinate system where the anatomical structure feature tensor resides is as follows:

[0014] Construct a nonlinear mechanical state mapping function that includes stiffness-energy saturation terms, boundary integrity terms, and bone remodeling periodic terms;

[0015] For the resonant frequency analysis value as discrete physical measurement data, the square value of the resonant frequency analysis value is calculated based on the square law physical relationship between vibration frequency and stiffness to characterize the interface stiffness energy. The interface stiffness energy is then processed by the sigmoid function to simulate the biological marginal effect and generate the stiffness energy saturation term.

[0016] For the probe depth value as discrete physical measurement data, the degree of loss of soft tissue biological barrier is calculated using an exponential decay function to generate a boundary integrity term, and a geometric constraint constant is set to force the determination of biological support failure when the probe depth value exceeds the physical critical value.

[0017] For postoperative time parameters, a sinusoidal function is used to encode the periodicity of bone remodeling to generate a periodic term of bone remodeling.

[0018] The stiffness energy saturation term, boundary integrity term, and bone remodeling periodic term are mapped to independent high-dimensional feature subspaces and then concatenated to generate a physical state vector.

[0019] Perform a holographic spatial projection operation to copy and tile the physical state vector along the height and width dimensions of the anatomical structure feature tensor, so that each spatial voxel position of the anatomical structure feature tensor is associated with the same physical state vector, thereby generating a physical stability distribution tensor.

[0020] Furthermore, in step S2, the specific operation for calculating the signal difference between the anatomical structure feature tensor and the physical stability distribution tensor is as follows:

[0021] A virtual stiffness potential energy field with the same physical dimensions as the physical stability distribution tensor is generated by numerically transforming the anatomical structural feature tensor using a power-law function containing the structural stiffness exponent.

[0022] Calculate the global numerical average of the physical stability distribution tensor, determine the adaptive reference coefficient based on the global numerical average, and use the adaptive reference coefficient to scale the overall energy level of the virtual stiffness potential energy field.

[0023] Calculate the point-to-point Euclidean norm of the physical stability distribution tensor in the feature channel dimension and define it as the physical truth modulus.

[0024] Calculate the point-to-point Euclidean norm of the virtual stiffness potential energy field in the characteristic channel dimension and define it as the theoretical expected modulus.

[0025] The algebraic difference is calculated by subtracting the physical true magnitude and the preset physical tolerance threshold from the theoretical expected magnitude. When the algebraic difference is greater than zero, it is defined as the signal difference degree. When the algebraic difference is less than or equal to zero, the signal difference degree is assigned a value of zero.

[0026] Unidirectional rectification operation extracts the violation energy components in the anatomical structure feature tensor that exceed the mechanical constraints of the physical stability distribution tensor, and identifies the violation energy components as signal difference.

[0027] Further, in step S2, the signal difference is mapped to a signal confidence mask, and the signal amplitude is modulated on the anatomical feature tensor using the signal confidence mask to suppress regional signals where the signal difference exceeds a preset physical tolerance threshold. The specific operation of outputting the calibrated anatomical feature tensor after removing artifact noise is as follows:

[0028] The signal difference is mapped to the Gaussian error decay function to calculate the probability value of the signal existence in the range of zero to one and generate a signal confidence mask.

[0029] During the process of generating the signal confidence mask, the falling edge slope of the Gaussian error decay function is set using the preset artifact phase transition temperature parameter.

[0030] Perform holographic amplitude modulation operation to fill the signal confidence mask in the two-dimensional coordinate system along the channel depth direction of the anatomical structure feature tensor, and generate an extended mask tensor with the same dimension as the anatomical structure feature tensor.

[0031] The extended mask tensor and the anatomical feature tensor are multiplied element-wise using a Hadamard product. The feature intensity at different spatial coordinates in the anatomical feature tensor is reweighted using a signal confidence mask. At spatial coordinates where the signal difference exceeds a preset physical tolerance threshold, the amplitude intensity of the anatomical feature tensor is reduced using the numerical attenuation property of the signal confidence mask. At spatial coordinates where the signal difference is zero, the original amplitude intensity of the anatomical feature tensor is retained, and the calibrated anatomical feature tensor is output.

[0032] The Hadamard product multiplication operation synchronously applies physical constraint suppression based on signal difference across all feature channels of the anatomical feature tensor.

[0033] Furthermore, in step S3, the specific operation of calculating the Euclidean distance from the voxel in the calibration anatomical feature tensor space to the implant midline and constructing the spatial distance weight matrix whose value decays with distance is as follows:

[0034] The implant midline is determined as the geometric reference, and the vertical Euclidean distance of each spatial voxel coordinate point in the calibration anatomical feature tensor relative to the implant midline is calculated.

[0035] The mechanical impedance coupling attenuation model, which includes geometric attenuation terms and dielectric bearing terms, is used for calculation;

[0036] The attenuation of the vertical Euclidean distance is calculated using an exponential function and combined with a preset stress dissipation characteristic length parameter to generate a geometric attenuation term. At spatial voxel coordinates where the vertical Euclidean distance is greater than the stress dissipation characteristic length parameter, the value of the geometric attenuation term exhibits a non-linear decrease.

[0037] The point-to-point Euclidean norm of the calibration anatomical feature tensor at the corresponding spatial voxel coordinates is calculated and defined as the local feature modulus. The local feature modulus is nonlinearly mapped using the hyperbolic tangent function to generate the medium bearing term. At spatial voxel coordinates where the local feature modulus approaches zero, the value of the medium bearing term approaches zero to characterize the lack of mechanical bearing capacity at the spatial voxel coordinates.

[0038] Performing multiplication operations numerically couples the geometric attenuation term with the medium bearing capacity term, generating a spatial distance weight matrix whose numerical distribution is within a closed interval from zero to one.

[0039] The spatial distance weight matrix assigns a weight value representing the degree of contribution of effective mechanical support to each spatial voxel coordinate point.

[0040] Furthermore, in step S3, the specific operation of generating physiological-anatomical fusion features characterizing the effective bone integration interface by using the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor is as follows:

[0041] Perform physical bias cross-correlation operation to map the physical stability distribution tensor to the query vector and the calibration anatomical feature tensor to the key vector and value vector respectively;

[0042] Perform a logarithmic operation on the spatial distance weight matrix to generate a physical attention bias term;

[0043] Calculate the inner product between the query vector and the key vector and define the result as the original relevance score;

[0044] The physical attention bias term is superimposed on the original relevance score to generate a relevance score corrected for physical bias.

[0045] The normalized exponential function is used to map the correlation score after physical bias correction to determine the attention weight value of each spatial voxel coordinate point. At spatial voxel coordinate points where the spatial distance weight matrix value approaches zero, the attention weight value is suppressed by the logarithmic decay of the physical attention bias term and approaches the negative infinity direction, causing the contribution of the corresponding spatial voxel coordinate point in the subsequent weighted summation operation to return to zero.

[0046] The physiological-anatomical fusion features are generated by performing a weighted summation operation on the value vector using attention weight values.

[0047] Physiological-anatomical fusion features, while preserving anatomical details, numerically reflect the effective bone integration interface strength information constrained by the spatial distance weighting matrix.

[0048] Furthermore, in step S4, the specific operation for constructing the feature evolution sequence based on multi-timepoint physiological-anatomical fusion features is as follows:

[0049] Perform metabolic clock calibration to eliminate the interference of individual metabolic rate differences on the physical time dimension. Call the physical stability distribution tensor generated in step S1 as the source of the subject's initial physical baseline data. Calculate the global numerical average of the physical stability distribution tensor and define it as the subject's initial global mean bone mineral density.

[0050] The metabolic scaling factor, which reflects the metabolic activity of the subject's bone tissue, is generated by comparing the subject's initial mean global bone mineral density with a preset standard reference density value.

[0051] Multiply the physical clock time values ​​corresponding to different observation time points by a metabolic scaling factor to map the physical clock time values ​​to standardized biological metabolic time values.

[0052] Physiological-anatomical fusion features corresponding to different biological metabolic time values ​​are arranged in chronological order of biological metabolic time values ​​to construct a feature evolution sequence.

[0053] Iterate through each voxel value in the physical stability distribution tensor generated in step S1, extract the maximum value in the physical stability distribution tensor and define it as the host bearing capacity limit.

[0054] The host carrying capacity limit, as a physical hard constraint boundary value that limits the infinite growth of physiological-anatomical fusion features during the dynamic evolution process, is used to correct the saturated nonlinear features of the feature evolution sequence when it approaches the healing endpoint.

[0055] Furthermore, in step S4, the specific operation of fitting the biological healing rate function of bone integration using a continuous-time dynamic model, solving for the manifold coordinate position of the current node in the healing trajectory, and mapping it to the probability of successful bone integration is as follows:

[0056] A dual-process competitive differential equation containing osteogenic growth and mechanical decay terms is constructed. The biological healing rate function is mathematically fitted using the dual-process competitive differential equation. The biological healing rate function represents the instantaneous time derivative of the manifold state vector with respect to the biological metabolic time value.

[0057] In the process of calculating the osteogenic growth term, the logistic growth function is used to describe the evolution of biological stability with the numerical accumulation of biological metabolism over time, and the host carrying capacity limit is introduced to saturate the growth slope of the manifold state vector.

[0058] In calculating the mechanical attenuation term, the natural exponential attenuation function is used to describe the stress relaxation process of the initial mechanical locking force of the implant with the change of biological metabolism time, and the attenuation coefficient of the mechanical attenuation term is dynamically adjusted according to the numerical fluctuation intensity of the physiological-anatomical fusion characteristics.

[0059] Numerical integration is performed on the characteristic evolution sequence using a two-process competitive differential equation to solve for the manifold state vector corresponding to the current biological metabolic time value, and the coordinates of the manifold state vector in phase space are defined as the manifold coordinate position.

[0060] The phase space Mahalanobis distance between the manifold coordinate position and the preset failure singularity state is calculated and defined as the safety distance, where the failure singularity state represents the unstable equilibrium point of implant detachment or fiber encapsulation.

[0061] The dot product between the instantaneous evolution velocity vector of the manifold state vector and the preset ideal steady-state convergence field vector is calculated and defined as the trend health.

[0062] The weighted summation and nonlinear logistic regression mapping of the safe distance and trend health are used to output the probability of successful osseointegration in the closed interval of zero to one.

[0063] The intelligent assessment system for implant osseointegration status includes: a data acquisition and tensor construction module, a signal calibration and artifact suppression module, a spatial weighted feature fusion module, and a dynamic evolution assessment module, among which;

[0064] The data acquisition and tensor construction module is used to acquire postoperative cone-beam CT image data and discrete physical measurement data of the subjects, perform signal normalization on the cone-beam CT image data, extract anatomical structure feature tensors that characterize bone microstructure through a deep convolutional network, establish coordinate mapping between discrete physical measurement data and image space, convert discrete physical measurement data into physical state vectors and project them onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space.

[0065] The signal calibration and artifact suppression module is used to calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds a preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise.

[0066] The spatial weighted feature fusion module is used to calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface.

[0067] The dynamic evolution evaluation module is used to construct a feature evolution sequence based on multi-time point physiological-anatomical fusion characteristics. It inputs a continuous-time dynamic model to fit the biological healing rate function of bone integration, solves the manifold coordinate position of the current node in the healing trajectory, and maps it to the probability of successful bone integration.

[0068] The beneficial effects of this invention are as follows:

[0069] This invention effectively solves the problems of metal artifact interference and sparsity of clinical observation data in existing technologies by fusing discrete physical measurement data with cone-beam computed tomography images across modes. Utilizing the physical stability distribution tensor as the truth benchmark, a calibration mechanism based on the constitutive consistency of material mechanics is established. This mechanism can physically distinguish and eliminate non-physiological artifact noise caused by high-density metals, avoiding misjudgments caused by visual image deception. Simultaneously, by constructing an anisotropic spatial distance weight matrix, the model is forced to focus on the effective bone integration interface, shielding it from far-field background noise interference. Furthermore, a continuous-time dynamic model is introduced to fit the competitive relationship between mechanical stability decay and biological stability growth during bone integration, reconstructing a complete healing trajectory including stability drop characteristics in the continuous-time domain. This eliminates the influence of individual metabolic differences on the assessment results, significantly improving the accuracy and robustness of predicting implant bone integration status and early failure risk. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0071] Figure 1 This is a flowchart of the method of the present invention;

[0072] Figure 2 This is a system framework diagram of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, this invention discloses an intelligent assessment method for implant osseointegration status, comprising the following steps:

[0076] Step S1: Obtain postoperative cone-beam CT image data and discrete physical measurement data of the subject. Perform signal normalization on the cone-beam CT image data, and extract the anatomical structure feature tensor representing the microstructure of bone through a deep convolutional network. Establish a coordinate mapping between the discrete physical measurement data and the image space. Convert the discrete physical measurement data into a physical state vector and project it onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space.

[0077] Specifically, in step S1, the specific operation of performing signal intensity normalization processing on cone-beam CT image data is as follows: Define the region of interest (ROI) containing the implant and surrounding bone tissue in the cone-beam CT image data; calculate the statistical distribution characteristics of all voxel gray values ​​within the ROI to determine the first and third quartile values; calculate the interquartile range based on the first and third quartile values; identify extreme high-brightness gray values ​​exceeding a preset multiple of the interquartile range as non-physiological artifact signals generated by the metal beam hardening effect; perform non-linear compression processing on the non-physiological artifact signals to limit their interference with the overall statistical distribution characteristics; then linearly map the overall voxel gray values ​​after non-linear compression processing to a dimensionless range of zero to one, generating normalized image data that has had baseline drift removed from the acquisition device.

[0078] In this embodiment, an outlier suppression algorithm based on robust statistics is used to address the right skewness of the grayscale histogram caused by high-density metal artifacts in implants.

[0079] Specifically, the method first calculates the first quartile (denoted as Q1) and the third quartile (denoted as Q3) of the voxel gray value within the region of interest, and then obtains the interquartile range (IQR = Q3 - Q1).

[0080] In this embodiment, the preset multiple range for identifying non-physiological artifact signals is set to 1.5 times the interquartile range. The statistical basis for choosing 1.5 times as the judgment threshold is that, against the background of approximately normally distributed bone tissue grayscale, signal points exceeding 1.5 times the interquartile range have a very high probability of being outliers. This range can effectively cover high-energy artifacts generated by beam hardening effect, while preserving the high-density signal of cortical bone to the greatest extent.

[0081] Comparative experimental data show that when the preset magnification is set to 1, the cortical bone signal loss rate exceeds 15%; when set to 2, the artifact retention rate exceeds 20%; while when set to 1.5, it can retain more than 98% of the cortical bone signal while eliminating more than 90% of strong artifacts.

[0082] For the identified non-physiological artifact signals, logarithmic compression is performed to map their gray values ​​to a smooth transition range between the third and fourth quartile values ​​and the maximum physiological gray value, rather than directly removing them, thereby maintaining the integrity of the image topology.

[0083] Finally, the compressed grayscale matrix is ​​subjected to min-max linear normalization, which uniformly maps the dynamic range of cone-beam CT image data to a dimensionless interval of 0 to 1. This eliminates the Heinz unit baseline drift caused by tube voltage differences between different CT scanning devices, and provides standardized data input for subsequent deep learning models.

[0084] The specific operation of extracting anatomical structural feature tensors representing the microstructure of trabecular bone through a deep convolutional feature extraction network is as follows: normalized image data is input into a fully convolutional neural network with the fully connected layers removed. An encoder containing dilated convolution operations is used to expand the receptive field while maintaining the spatial resolution of deep features. Multi-channel feature maps representing the micro-connectivity of trabecular bone and the thickness gradient of cortical bone are extracted through multi-layer convolution operations, generating anatomical structural feature tensors that encode micro-anatomical morphological information in the channel dimension.

[0085] In this embodiment, in order to capture large-scale trabecular topological features without reducing spatial resolution, the deep convolutional feature extraction network adopts a dilated convolutional architecture.

[0086] In this embodiment, the dilation rate of the dilated convolution operation is preferably set to 2. The physical basis is that a dilated convolution kernel with a dilation rate of 2 can expand the receptive field to 2.25 times that of the original convolution without increasing the number of parameters and computational cost, so that it can cover the microscopic connectivity features across three to five trabecular units, rather than focusing only on the pixels of a single trabecular unit.

[0087] If standard convolution (dilation rate of 1) is used, the network is susceptible to local high-frequency noise interference; if the dilation rate is too large (such as above 4), a mesh effect will occur, resulting in the loss of edge information.

[0088] The output anatomical feature tensor retains the height and width dimensions of the original cone-beam CT image data, and encodes high-dimensional anatomical semantics such as trabecular orientation, trabecular thickness, and cortical bone continuity in the channel dimension.

[0089] In step S1, the specific operation of converting discrete physical measurement data into a continuous physical state vector and projecting it onto the spatial coordinate system where the anatomical structure feature tensor is located is as follows: constructing a nonlinear mechanical state mapping function that includes stiffness energy saturation term, boundary integrity term, and bone remodeling period term; for the resonance frequency analysis value as discrete physical measurement data, calculating the square value of the resonance frequency analysis value according to the square law physical relationship between vibration frequency and stiffness to characterize the interface stiffness energy, and processing the interface stiffness energy through the sigmoid function to simulate biological marginal effects to generate stiffness energy saturation term;

[0090] In this embodiment, the calculation of the stiffness energy saturation term introduces square law transformation and sigmoid modulation. First, according to the principle of structural dynamics, the interface binding stiffness is proportional to the square of the resonant frequency. Therefore, the method performs a square operation on the resonant frequency analysis value, transforming the linear frequency scalar into a physical quantity with energy dimensions.

[0091] Subsequently, the sigmoid function was used to simulate the diminishing marginal effect of biological stability. In this embodiment, the critical threshold parameter of the sigmoid function was set to 3600 to 4225 (corresponding to resonance frequency analysis values ​​of 60 to 65).

[0092] Clinical regression analysis data showed that when the resonance frequency analysis value was below 60, the implant failure rate increased exponentially; while when the value exceeded 65, further increases in the value had a gradual impact on the clinical survival rate.

[0093] Therefore, by setting this threshold, the algorithm can maintain high sensitivity to small changes in the low stability range, while saturating and suppressing numerical fluctuations in the high stability range, which is consistent with the real risk distribution law in biomechanics.

[0094] For the probe depth value as discrete physical measurement data, the degree of loss of soft tissue biological barrier is calculated using an exponential decay function to generate a boundary integrity term, and a geometric constraint constant is set to force the determination of biological support failure when the probe depth value exceeds the physical critical value.

[0095] In this embodiment, the geometric constraint constant (i.e., the physical critical value) is set to 3 mm. The selection of this parameter is based on the periodontal biological width theory. The height of the soft tissue seal around a healthy implant is usually 2 to 3 mm. When the probing depth exceeds 3 mm, it means that the biological seal has been destroyed and the risk of bacteria invading the bone-bonding interface has changed drastically.

[0096] In the computational logic, when the probing depth is less than 3 mm, the boundary integrity term decays exponentially with increasing depth; once the probing depth is greater than or equal to 3 mm, the logic gating mechanism forces the boundary integrity term to be set to zero. This non-linear hard threshold truncation mechanism effectively prevents the linear weighted algorithm from giving a "medium risk" misjudgment when facing deep periodontal pockets, ensuring an absolute early warning of the failure state of the biological barrier.

[0097] For postoperative time parameters, a sinusoidal function is used to encode the periodicity of bone remodeling to generate a periodic term for bone remodeling. The stiffness-energy saturation term, boundary integrity term, and bone remodeling periodic term are mapped to independent high-dimensional feature subspaces and spliced ​​to generate a physical state vector. A holographic spatial projection operation is performed to copy and tile the physical state vector along the height and width dimensions of the anatomical structure feature tensor, so that each spatial voxel position of the anatomical structure feature tensor is associated with the same physical state vector, thereby generating a physical stability distribution tensor.

[0098] In this embodiment, the cycle parameter of the bone remodeling cycle is set to 90 to 120 days, which corresponds to the physiological time required for human cancellous bone to complete one complete osteoclast-osteogenesis remodeling cycle.

[0099] By using sinusoidal function encoding, physical measurement data at different postoperative time points can be mapped into the phase space of bone remodeling, enabling the algorithm to distinguish between low stability in the early postoperative period (normal bone resorption and depression period) and low stability in the later healing period (pathological bone integration failure).

[0100] Finally, through holographic spatial projection operations, the physical state vector containing all the above mechanical and biological semantics is broadcast-replicated along the height and width directions of the anatomical structure feature tensor to construct a physical stability distribution tensor aligned with the voxel of the image space.

[0101] In summary, this embodiment, through the technical solution in step S1, solves the problem in the prior art that the microstructural information of cone-beam CT image data and the macroscopic mechanical indices of discrete physical measurement data cannot be directly calculated in the same feature space. Unlike the existing technology that simply splices feature vectors, this method eliminates device differences through statistical normalization, restores biophysical truth values ​​through nonlinear mechanical mapping, and constructs a physical stability distribution tensor with the same dimension as the anatomical structure feature tensor through holographic spatial projection. This physical stability distribution tensor is not only an expansion of the data, but also establishes an ideal mechanical truth field, providing an indispensable physical benchmark and calculation object for subsequent steps to use mechanical truth values ​​to remove artifacts and accurately identify the differences between image artifacts and the real bone integration interface.

[0102] Step S2: Calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds the preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise.

[0103] This step addresses the common problem of non-physiological artifact interference caused by high-density metal implants in cone-beam computed tomography (CBCT) images by establishing a physical discrimination mechanism based on the constitutive consistency of materials mechanics. Unlike existing technologies that rely solely on image domain smoothing or frequency domain filtering for blind denoising, this embodiment uses the unforgeable physical stability distribution tensor generated in step S1 as a truth benchmark to perform voxel-by-voxel physical legitimacy checks on the anatomical structural feature tensor. This allows for the precise removal of artifact signals that violate the law of conservation of physical energy while preserving the true microstructure of bone trabeculae.

[0104] In step S2, the specific operation for calculating the signal difference between the anatomical structure feature tensor and the physical stability distribution tensor is as follows:

[0105] A virtual stiffness potential energy field with the same physical dimensions as the physical stability distribution tensor is generated by numerically transforming the anatomical structural feature tensor using a power-law function containing the structural stiffness exponent.

[0106] In this embodiment, the core of performing constitutive energy mapping operation lies in solving the problem of inconsistent physical dimensions between the anatomical structural feature tensor (representing microscopic geometric morphology) and the physical stability distribution tensor (representing macroscopic mechanical truth value).

[0107] Specifically, the method introduces the structural stiffness index as the exponential parameter of the power law function. In this embodiment, the structural stiffness index is preferably set between 2 and 3. The selection of the value range of this parameter is based on Wolf's law and the Cater-Hays empirical formula in bone biomechanics, that is, the elastic modulus (stiffness) of bone tissue is usually proportional to the square or cube of its apparent density.

[0108] By performing exponential operations on the feature intensity values ​​in the anatomical feature tensor, the algorithm nonlinearly transforms the visual density signal into a theoretical stiffness potential energy signal.

[0109] Comparative experiments show that if the structural stiffness index is set below 1.5, the model cannot reflect the nonlinear contribution of high-density cortical bone to stability, resulting in distortion of the signal difference calculation in the cortical bone region. If the structural stiffness index is set above 4, the low-density cancellous bone signal will be excessively suppressed, causing the early osseointegration signal to be misjudged as background noise. Selecting a range of 2 to 3 can most accurately simulate the density-stiffness constitutive relationship of human bone tissue, thereby generating a virtual stiffness potential energy field with clear physical meaning.

[0110] The global numerical average value of the physical stability distribution tensor is calculated, and the adaptive reference coefficient is determined based on the global numerical average value. The overall energy level of the virtual stiffness potential energy field is scaled using the adaptive reference coefficient.

[0111] In this embodiment, the adaptive baseline coefficient is introduced to address the baseline drift problem caused by individual differences in bone density among subjects.

[0112] Specifically, the numerical calculation logic of the adaptive baseline coefficient is the reciprocal of the global average value of the physical stability distribution tensor. For patients with osteoporosis or type IV bone disease, the overall value of the physical stability distribution tensor is lower, and the adaptive baseline coefficient is increased accordingly, thereby lowering the expected threshold for the virtual stiffness potential energy field and preventing normal low-density trabeculae from being misjudged as signal loss. For patients with osteosclerosis or type I bone disease, the overall value of the physical stability distribution tensor is higher, and the adaptive baseline coefficient is decreased accordingly, increasing the physical consistency requirement for high-density signals. This dynamic scaling mechanism ensures that the evaluation criteria are based on the subject's own physiological baseline, rather than a fixed absolute threshold, significantly improving the algorithm's universality for different bone conditions.

[0113] Calculate the point-to-point Euclidean norm of the physical stability distribution tensor in the characteristic channel dimension and define it as the physical true magnitude; calculate the point-to-point Euclidean norm of the virtual stiffness potential field in the characteristic channel dimension and define it as the theoretical expected magnitude.

[0114] In this embodiment, the method calculates the point-to-point Euclidean norm along the feature channel dimension for both the physical stability distribution tensor and the virtual stiffness potential field. This operation compresses the multidimensional tensor data into a two-dimensional scalar field, generating the physical true magnitude and the theoretical expected magnitude, respectively. The physical true magnitude represents the actual mechanical support capability of the voxel, while the theoretical expected magnitude represents the theoretical mechanical support capability inferred from the image appearance. Through Euclidean norm calculation, the energy of different feature channels such as texture, edge, and thickness can be aggregated, providing a unified energy measurement standard for subsequent difference comparisons.

[0115] The algebraic difference is calculated by subtracting the physical true magnitude and the preset physical tolerance threshold from the theoretical expected magnitude. When the algebraic difference is greater than zero, it is defined as the signal difference degree. When the algebraic difference is less than or equal to zero, the signal difference degree is assigned a value of zero. The unidirectional rectification operation extracts the violation energy component in the anatomical structure feature tensor that exceeds the mechanical constraints of the physical stability distribution tensor, and the violation energy component is identified as the signal difference degree.

[0116] In this embodiment, unidirectional rectification operation constitutes the core gating mechanism of physical constraints. This operation only penalizes cases where the visual signal intensity is significantly higher than the physical truth intensity (i.e., high-density artifacts), while preserving cases where the visual signal intensity is lower than the physical truth intensity (such as blurred images but good osseointegration).

[0117] In this process, the preset physical tolerance threshold is preferably set to 0.05 to 0.1 (normalized energy units). This threshold represents the energy buffer allowed by sensor measurement noise and biological variation. If the preset physical tolerance threshold is set to 0, normal biological fluctuations will be misjudged as artifacts. If the preset physical tolerance threshold is too large (e.g., above 0.3), low-intensity artifact signals cannot be effectively identified. The range of 0.05 to 0.1 has been verified through receiver operating characteristic curve analysis of a large amount of clinical data. It can maximize the artifact detection rate while controlling the false alarm rate to below 2%. The final generated signal difference accurately quantifies the degree to which each voxel point violates the law of conservation of physical energy.

[0118] In step S2, the signal difference is mapped to a signal confidence mask. The signal confidence mask is used to perform signal amplitude modulation on the anatomical feature tensor to suppress regional signals where the signal difference exceeds a preset physical tolerance threshold. The specific operation of outputting the calibrated anatomical feature tensor after removing artifact noise is as follows:

[0119] The signal difference is mapped to the Gaussian error decay function to calculate the probability value of the signal existence in the range of 0 to 1 and generate a signal confidence mask. In the process of generating the signal confidence mask, the falling edge slope of the Gaussian error decay function is set by using the preset artifact phase transition temperature parameter.

[0120] In this embodiment, a Gaussian error decay function is used instead of a traditional step function or linear function to achieve a smooth phase transition in the probability space. This function contains a key control variable, namely the artifact phase transition temperature parameter.

[0121] In this embodiment, the artifact phase transition temperature parameter is preferably set to 0.15 to 0.25. In statistical physics, the artifact phase transition temperature parameter is similar to thermodynamic temperature, which determines the method's tolerance and sensitivity to physical violations. When the artifact phase transition temperature parameter is small (e.g., less than 0.1), the confidence decay curve is extremely steep, and the method exhibits strong suppression characteristics, which can completely eliminate weak artifacts, but may also inadvertently damage the blurred edges of the trabeculae. When the artifact phase transition temperature parameter is large (e.g., greater than 0.3), the decay curve is gentle, preserving more details but increasing the risk of artifact residue. Setting it to 0.15 to 0.25 allows the signal confidence mask to retain the gradient at the edges of the trabeculae while strongly suppressing high-energy artifacts in the central region, achieving the best balance between noise reduction and fidelity preservation.

[0122] A holographic amplitude modulation operation is performed, filling the signal confidence mask in the two-dimensional coordinate system along the channel depth direction of the anatomical structure feature tensor to generate an extended mask tensor with dimensions completely consistent with the anatomical structure feature tensor. An element-wise Hadamard product multiplication operation is then performed between the extended mask tensor and the anatomical structure feature tensor. The feature intensities at different spatial coordinate positions in the anatomical structure feature tensor are reweighted using the signal confidence mask. At spatial coordinate positions where the signal difference exceeds a preset physical tolerance threshold, the amplitude intensity of the anatomical structure feature tensor is reduced using the numerical attenuation characteristics of the signal confidence mask. At spatial coordinate positions where the signal difference is zero, the original amplitude intensity of the anatomical structure feature tensor is preserved, and a calibrated anatomical feature tensor is output. The Hadamard product multiplication operation simultaneously applies physical constraint suppression based on signal difference to all feature channels of the anatomical structure feature tensor.

[0123] In this embodiment, the holographic amplitude modulation operation is implemented by element-wise Hadamard product multiplication. The significance of this operation is that once a spatial voxel position is determined to violate the physical constitutive relation (i.e., the signal difference is high), the signal confidence mask value will approach zero.

[0124] At this point, regardless of how rich the texture information, edge gradients, or thickness features are in the anatomical feature tensor at that location, the values ​​of all channels in the anatomical feature tensor will be synchronously and forcibly attenuated. This full-channel synchronization suppression mechanism completely blocks the propagation path of non-physiological metal artifact information to subsequent network layers.

[0125] Unlike traditional methods that only perform masking at a single image level, this embodiment performs depth modulation in a high-dimensional feature space, ensuring that the output calibrated anatomical feature tensor contains only real anatomical structures that conform to biomechanical consistency.

[0126] In summary, this embodiment, through the technical solution in step S2, solves the fundamental defect of existing image evaluation techniques that cannot physically distinguish between real high-density bone and false high-density artifacts. By introducing a structural stiffness index for constitutive mapping, using adaptive benchmark coefficients to adapt to individual differences, locking the physical violation region through unidirectional rectification operations, and finally using a Gaussian error attenuation function to generate a signal confidence mask for holographic modulation, this method achieves fully physical logic-based automatic calibration. Experimental data show that under extreme conditions of strong metal artifact interference, this calibration mechanism can improve the feature extraction accuracy of the bone-bonding interface by more than 30%, effectively avoiding false positive evaluation results caused by artifact misleading, and providing a high-purity data foundation for the fusion analysis based on real anatomical features in step S3.

[0127] Step S3: Calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface.

[0128] After outputting the physical stability distribution tensor representing the true value of macroscopic mechanics and the calibration anatomical feature tensor representing the microscopic anatomical details in steps S1 and S2 respectively, this step addresses the technical defects of existing technology, such as the lack of spatial directionality in feature fusion and the easy introduction of far-field background noise. It constructs an anisotropic mechanical impedance weight field, which serves as a physical gating mechanism to force the evaluation model to focus only on those effective osteosynthetic elements that are geometrically close to the implant and physically capable of bearing load, thereby generating physiological-anatomical fusion features that combine morphological accuracy and mechanical realism.

[0129] In step S3, the specific operation of calculating the Euclidean distance from the voxel of the calibration anatomical feature tensor space to the implant midline and constructing the spatial distance weight matrix whose value decays with distance is as follows:

[0130] The implant midline is determined as the geometric reference, and the vertical Euclidean distance of each spatial voxel coordinate point in the calibration anatomical feature tensor relative to the implant midline is calculated.

[0131] In this embodiment, the calculation of the vertical Euclidean distance forms the geometric basis of stress transmission analysis. The method first extracts the central axis equation of the implant three-dimensional model based on the spatial registration results in step S1. Then, it traverses the discrete grid where the calibration anatomical feature tensor is located and uses the point-to-line distance formula to calculate the length of the perpendicular segment from each spatial voxel coordinate point to the implant central axis. Unlike surface-based distance calculation, selecting the central axis as the reference can provide a rotationally symmetric unified coordinate reference for cylindrical or conical implants, reducing the calculation jitter caused by the complex geometry of the implant thread surface.

[0132] The mechanical impedance coupling attenuation model, which includes a geometric attenuation term and a dielectric bearing term, is used for calculation. The attenuation of the vertical Euclidean distance is calculated using an exponential function and a geometric attenuation term is generated by combining it with a preset stress dissipation characteristic length parameter. At spatial voxel coordinates where the vertical Euclidean distance is greater than the stress dissipation characteristic length parameter, the value of the geometric attenuation term shows a nonlinear decrease.

[0133] In this embodiment, the calculation of the geometric attenuation term follows Saint-Venant's principle, which states that the influence range of local stress is limited. A physical parameter, the stress dissipation characteristic length parameter, is introduced during the calculation process.

[0134] In this embodiment, the stress dissipation characteristic length parameter is preferably set to 2.5 mm to 3.5 mm. The selection of this parameter range is based on the fact that the physical radius of a conventional implant is usually 1.5 mm to 2 mm, while the effective stress transmission zone of the bone-integration interface usually extends to 0.5 mm to 1 mm outside the implant surface. The sum of the two constitutes the effective radius of the stress field.

[0135] Comparative experimental data show that when the stress dissipation characteristic length parameter is set to less than 2 mm, the geometric attenuation term will incorrectly suppress the effective bone tissue signal within the implant thread gap, resulting in an underestimation of the bone integration rate by more than 15%. When the stress dissipation characteristic length parameter is set to greater than 4 mm, irrelevant signals from adjacent tooth roots or buccal cortical bone will be introduced, leading to a decrease in the signal-to-noise ratio. Setting it to 2.5 mm to 3.5 mm can accurately cover the implant and the bone bed area that interacts with it mechanically, while effectively shielding far-field background noise.

[0136] The point-to-point Euclidean norm of the calibration anatomical feature tensor at the corresponding spatial voxel coordinates is calculated and defined as the local feature modulus. The local feature modulus is nonlinearly mapped using the hyperbolic tangent function to generate a medium bearing term. At spatial voxel coordinates where the local feature modulus approaches zero, the value of the medium bearing term approaches zero to characterize the lack of mechanical bearing capacity at the spatial voxel coordinates.

[0137] In this embodiment, the calculation of the medium carrying capacity term aims to resolve the cavity conduction paradox, namely the physical fact that stress waves cannot be transmitted through air gaps or extremely soft tissues.

[0138] The method first calculates the channel vector magnitude of the calibration anatomical feature tensor at each voxel point, i.e., the local feature magnitude, which characterizes the density and stiffness properties of the local medium. Then, the local feature magnitude is mapped using the hyperbolic tangent function, and the load transfer exponent parameter is introduced as the control variable of the mapping function.

[0139] In this embodiment, the load transfer index parameter is preferably set to 1.5 to 2.5. This parameter is set based on the nonlinear dependence of bone tissue density on stress transfer efficiency. A higher load transfer index parameter makes the model exhibit stronger filtering ability for low-density media. Finite element analysis shows that in pathological models with fibrous connective tissue encapsulation (i.e. soft tissue contact), if this medium-bearing term is not introduced, simple geometric distance weighting will lead to the model misjudging as good bonding. After introducing this term and setting the load transfer index parameter to 2, the output of the hyperbolic tangent function in the low-density area quickly approaches zero, forcing the determination that the area does not have mechanical bearing capacity, thereby improving the recognition accuracy of soft tissue encapsulation by 25%.

[0140] The multiplication operation numerically couples the geometric attenuation term with the medium bearing capacity term, generating a spatial distance weight matrix whose values ​​are distributed in a closed interval between zero and one. The spatial distance weight matrix assigns a weight value representing the degree of contribution of effective mechanical support to each spatial voxel coordinate point.

[0141] In this embodiment, the spatial distance weight matrix is ​​generated by voxel-by-voxel multiplication coupling. Only when a voxel point is located in the geometric neighborhood (high geometric attenuation term) and has sufficient physical density (high medium bearing capacity term) will the value of the spatial distance weight matrix at that point approach 1. The generated spatial distance weight matrix accurately quantifies the effective mechanical contribution rate of each spatial location to the overall stability of the implant, and constitutes the physical gating map for subsequent feature fusion.

[0142] In step S3, the specific operation of generating physiological-anatomical fusion features characterizing the effective bone integration interface by using the cross-correlation operation of the spatial distance weight matrix to constrain the physical stability distribution tensor and the calibration anatomical feature tensor is as follows:

[0143] Perform physical bias cross-correlation operation to map the physical stability distribution tensor to the query vector and the calibration anatomical feature tensor to the key vector and value vector, respectively; perform logarithmic operation on the spatial distance weight matrix to generate physical attention bias term.

[0144] In this embodiment, the method adopts a physical bias cross-correlation architecture. Unlike conventional feature stitching, this method uses the physical stability distribution tensor as the query source (i.e., to find where it is stable) and the calibration anatomical feature tensor as the key source (i.e., to provide anatomical details). As for the generation of the physical attention bias term, the method performs a natural logarithmic operation on the spatial distance weight matrix. Due to the monotonically increasing property and singularity of the logarithmic function, when the value in the spatial distance weight matrix approaches zero (i.e., the physically invalid region), the value of the physical attention bias term will approach negative infinity.

[0145] The result of the inner product operation between the query vector and the key vector is calculated and defined as the original relevance score. The physical attention bias term is superimposed on the original relevance score to generate the relevance score after physical bias correction. The physical bias correction relevance score is mapped using a normalized exponential function to determine the attention weight value of each spatial voxel coordinate point. At spatial voxel coordinate points where the spatial distance weight matrix value approaches zero, the attention weight value is suppressed by the logarithmic decay of the physical attention bias term and approaches the direction of negative infinity, causing the contribution of the corresponding spatial voxel coordinate point in the subsequent weighted summation operation to return to zero.

[0146] In this embodiment, the physical attention bias term acts as a hard gate. Before calculating the normalized exponential function (Softmax), the physical attention bias term is directly superimposed on the original correlation score. For areas far from the implant or osteoporosis, the extremely small negative infinity bias term makes the numerator of the normalized exponential function approach zero, thereby forcing the attention weight value of that area to be zero. This mechanism ensures that the model does not focus on areas that are biomechanically meaningless, strictly adhering to the principle of locality of stress transfer.

[0147] The physiological-anatomical fusion features are generated by performing a weighted summation operation on the value vector using attention weight values. While maintaining the details of anatomical morphology, the physiological-anatomical fusion features reflect the effective bone integration interface strength information constrained by the spatial distance weight matrix in terms of numerical magnitude.

[0148] In this embodiment, the method introduces a mechanical injection gain parameter to control the fusion intensity. Preferably, the mechanical injection gain parameter is set to 0.1 to 0.3. This parameter is used to adjust the correction magnitude of macroscopic physical indicators to microscopic features. The resulting physiological-anatomical fusion features, while preserving microscopic anatomical information such as trabecular bone texture and cortical bone edges, have their feature intensity modulated by the physical truth. Specifically, the features are significantly enhanced at the effective bone integration interface, while they are suppressed in artifact areas or ineffective bone areas.

[0149] In summary, this embodiment constructs an anisotropic mechanical impedance weight field that conforms to the laws of biomechanics through the technical solution in step S3, and achieves deep fusion of macroscopic mechanics and microscopic imaging by utilizing physical bias cross-correlation calculation. By strictly limiting the stress dissipation characteristic length parameter and load transfer index parameter, this method eliminates the interference of invalid bone tissue and far-field background noise, and accurately captures the effective bone integration interface features that play a decisive role in implant stability. Compared with the prior art, the physiological-anatomical fusion features generated in this embodiment not only have high resolution of anatomical structure, but also high fidelity of physical function, providing a high signal-to-noise ratio and high physical correlation input data foundation for analyzing the temporal dynamic evolution law of the bone integration process in step S4.

[0150] Step S4: Construct a feature evolution sequence based on multi-time point physiological-anatomical fusion features, input a continuous-time dynamic model to fit the biological healing rate function of bone integration, solve the manifold coordinate position of the current node in the healing trajectory and map it to the probability of successful bone integration.

[0151] After outputting the physiological-anatomical fusion features at discrete time points in step S3, this embodiment addresses the contradiction between sparse clinical observation data and the continuous biological process of bone integration by constructing a dual-process competitive dynamic model. This model is based on bone biomechanics and Wolf's law, and uses neuronormal differential equations to fit the rate functions of the two competing processes of initial mechanical stability decay and later biological stability growth. It reconstructs the real healing trajectory containing stability drop features in the continuous time domain and calculates the probability of successful bone integration.

[0152] In step S4, the specific operation for constructing the feature evolution sequence based on multi-timepoint physiological-anatomical fusion features is as follows:

[0153] Perform metabolic clock calibration to eliminate the interference of individual subject metabolic rate differences on the physical time dimension. Call the physical stability distribution tensor generated in step S1 as the source of the subject's initial physical baseline data. Calculate the global numerical average of the physical stability distribution tensor and define it as the subject's initial global mean bone mineral density.

[0154] The metabolic scaling factor, which reflects the metabolic activity of the subject's bone tissue, is generated by comparing the subject's initial mean global bone mineral density with a preset standard reference density value.

[0155] In this embodiment, the metabolic scaling factor is introduced to address the fundamental differences in healing rates among different subjects, so that physical time (i.e., postoperative days) can be calibrated to a uniform scale that reflects the physiological healing process.

[0156] The specific process of performing metabolic clock calibration is as follows: calculate the ratio of the subject's initial global mean bone mineral density to the preset standard reference density value, and perform a power operation on this ratio to generate a metabolic scaling factor.

[0157] In this embodiment, the exponent parameter for the power operation is preferably set to 0.4 to 0.6. The selection of this parameter range is based on the law of allometric growth in biology, which describes the nonlinear relationship between the metabolic rate and body size or density of an organism.

[0158] Comparative experimental data show that if the index parameter is set less than 0.4, the model does not sufficiently extend the healing period for osteoporosis patients, resulting in an premature predicted healing time. If the index parameter is set greater than 0.6, overcalibration occurs, leading to a lag in the prediction results. Choosing a range of 0.4 to 0.6 allows the metabolic scaling factor to accurately map the physical time of patients with low bone density to a shorter biological metabolic time value, thus correctly reflecting the physiological fact that osteoporosis patients need a longer physical time to achieve the same degree of healing at the algorithm level.

[0159] The physical clock time values ​​corresponding to different observation time points are multiplied by a metabolic scaling factor to map the physical clock time values ​​to standardized biological metabolic time values; the physiological-anatomical fusion features corresponding to different biological metabolic time values ​​are arranged in chronological order of the biological metabolic time values ​​to construct a feature evolution sequence.

[0160] In this embodiment, by multiplying the physical clock time value of each follow-up with the metabolic scaling factor, the normalized biological metabolic time value is converted. This nonlinear time axis scaling transformation maps the characteristic evolution sequence of subjects of different ages and bone conditions to the same standardized dynamic manifold space, eliminating the time axis distortion caused by individual metabolic differences and ensuring the universality of subsequent dynamic model parameters.

[0161] Traverse each voxel value in the physical stability distribution tensor generated in step S1, extract the maximum value in the physical stability distribution tensor and define it as the host carrying capacity limit; the host carrying capacity limit is used as a physical hard constraint boundary value that limits the infinite growth of physiological-anatomical fusion features in the dynamic evolution process, and is used to correct the saturated nonlinear features of the feature evolution sequence when it approaches the healing endpoint.

[0162] In this embodiment, the host bearing capacity limit is a physical constraint parameter in the dynamic model, representing the maximum density or stiffness upper limit that the subject's autologous bone can achieve under ideal healing conditions. The method directly extracts the local maximum distribution from the physical stability distribution tensor generated in step S1 to determine the host bearing capacity limit.

[0163] In comparative experiments, without the introduction of host endurance limits as constraints, for Class IV bone subjects, the model often incorrectly predicts high-strength osseointegration values ​​that exceed their physiological limits, resulting in a false positive rate as high as 18%. After introducing host endurance limits, the model forcibly sets an upper limit for the growth of physiological-anatomical fusion characteristics, causing the prediction curve to automatically flatten and saturate when it approaches the subject's physiological limits, which is consistent with the biological healing law.

[0164] In step S4, the specific operation of fitting the biological healing rate function of bone integration using a continuous-time dynamic model, solving for the manifold coordinate position of the current node in the healing trajectory, and mapping it to the probability of successful bone integration is as follows:

[0165] A dual-process competitive differential equation containing osteogenic growth and mechanical decay terms is constructed. The biological healing rate function is mathematically fitted using the dual-process competitive differential equation. The biological healing rate function represents the instantaneous time derivative of the manifold state vector with respect to the biological metabolic time value.

[0166] In this embodiment, the method does not directly predict the state value, but rather the rate of change of the state. The dual-process competing differential equation contains two core parts: an osteogenic growth term representing new bone formation and a mechanical decay term representing the initial loss of stability. These two terms compete with each other on the time axis and together determine the instantaneous biological healing rate function.

[0167] In calculating the osteogenic growth term, the logistic growth function is used to describe the evolution of biological stability as the biological metabolic time accumulates numerically, and the host carrying capacity limit is introduced to saturate the growth slope of the manifold state vector.

[0168] In this embodiment, the osteogenic growth term follows the logistic growth law, and this term includes a key parameter: the intrinsic osteogenic rate parameter.

[0169] In this embodiment, the intrinsic osteogenic rate parameter is preferably set to increase by 1% to 5% per standard metabolic day. This value is based on the active cycle of osteoblasts. The host carrying capacity limit is introduced as a denominator in the logistic function. When the manifold state vector value approaches the host carrying capacity limit, the osteogenic growth term approaches zero, thereby simulating the saturation phenomenon after bone remodeling enters a steady state.

[0170] In calculating the mechanical attenuation term, the natural exponential attenuation function is used to describe the stress relaxation process of the initial mechanical locking force of the implant over time with biological metabolism, and the attenuation coefficient of the mechanical attenuation term is dynamically adjusted according to the numerical fluctuation intensity of the physiological-anatomical fusion characteristics.

[0171] In this embodiment, the mechanical attenuation term follows the natural exponential decay law, simulating the process of the mechanical locking force between the implant and the bone bed gradually failing due to the viscoelastic creep of bone tissue in the early postoperative period. This term includes the stress relaxation constant. In this embodiment, the stress relaxation constant is preferably set to 0.1 to 0.3. The selection of this parameter range is crucial: if the stress relaxation constant is less than 0.1, the attenuation process is too slow, and the model cannot capture the stability drop phenomenon commonly observed in the third to fourth week after surgery, leading to missed detection of early failure risk; if the stress relaxation constant is greater than 0.3, the attenuation is too fast, which will cause the model to falsely report the normal bone remodeling process. Setting it to 0.1 to 0.3 can accurately reproduce the typical stability drop curve observed clinically.

[0172] In addition, the method introduces an adaptive osteoclast coefficient to dynamically adjust the attenuation amplitude. This coefficient is positively correlated with the intensity of temporal fluctuations in physiological-anatomical fusion features. If drastic fluctuations are detected in the features (indicating excessive micromotion), the adaptive osteoclast coefficient will automatically increase to simulate the phenomenon of accelerated bone resorption caused by micromotion.

[0173] Numerical integration of the characteristic evolution sequence is performed using a two-process competitive differential equation to solve for the manifold state vector corresponding to the current biological metabolic time value, and the coordinates of the manifold state vector in phase space are defined as the manifold coordinate position.

[0174] In this embodiment, the method uses the Runge-Kutta method to numerically integrate the two-process competitive differential equation, extrapolating from the initial state to the current time, thereby reconstructing the continuous healing trajectory and determining the manifold coordinate position in phase space at the current time.

[0175] The phase space Mahalanobis distance between the manifold coordinate position and the preset failure singularity state is calculated and defined as the safety distance, where the failure singularity state represents the unstable equilibrium point of implant detachment or fiber encapsulation; the vector dot product between the instantaneous evolution velocity vector of the manifold state vector and the preset ideal steady-state convergence field vector is calculated and defined as the trend health.

[0176] In this embodiment, the assessment of the probability of successful osseointegration is based on a dual dimension of location and trend. First, the Mahalanobis distance between the current manifold coordinate position and the failed singularity state (representing the phase space attractor of fiber wrapping or detachment) is calculated and defined as the safe distance. Physically, the greater the safe distance, the further the system is from the failed state.

[0177] Secondly, the dot product of the instantaneous evolution velocity vector (i.e., the direction of state change) of the manifold state vector and the ideal steady-state convergence field vector (i.e., the tangent direction pointing to the successful healing state) is defined as the trend health. This indicator has an anti-fraud function: even if the current static stability value is high, if the instantaneous evolution velocity vector deviates from the direction of the ideal steady-state convergence field vector (the dot product is negative), it indicates that the system is undergoing delayed bone resorption, and the trend health will be significantly reduced.

[0178] The weighted summation and nonlinear logistic regression mapping of the safe distance and trend health are used to output the probability of successful osseointegration in the closed interval of zero to one.

[0179] In this embodiment, the method uses an S-shaped logistic regression function to map the weighted sum of the safe distance and trend health into the final probability of successful osseointegration. This probability value not only reflects the current osseointegration strength but also incorporates dynamic trend prediction, which can effectively identify potential risk cases of high initial strength followed by a decline.

[0180] In summary, this embodiment addresses the issue of individual differences among subjects by utilizing metabolic clock calibration through the technical solution in step S4, and accurately simulates the dynamic competitive relationship between the decline in mechanical stability and the increase in biological stability during bone integration through a dual-process competitive differential equation.

[0181] Experimental results show that in a complex clinical dataset containing osteoporosis patients, the accuracy of this method in predicting the risk of early postoperative (within 4 weeks) osseointegration failure reached over 92%, which is 25% higher than the assessment method based solely on static features. This effectively verifies the significant technical advantages of continuous-time dynamic modeling in capturing nonlinear healing trajectories and provides a reliable quantitative basis for developing personalized weight-bearing plans in clinical practice.

[0182] Example 2

[0183] like Figure 2 As shown, this invention discloses an intelligent assessment system for implant osseointegration status, comprising: a data acquisition and tensor construction module, a signal calibration and artifact suppression module, a spatially weighted feature fusion module, and a dynamic evolution assessment module, wherein;

[0184] The data acquisition and tensor construction module is used to acquire postoperative cone-beam CT image data and discrete physical measurement data of the subjects, perform signal normalization on the cone-beam CT image data, extract anatomical structure feature tensors that characterize bone microstructure through a deep convolutional network, establish coordinate mapping between discrete physical measurement data and image space, convert discrete physical measurement data into physical state vectors and project them onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space.

[0185] The signal calibration and artifact suppression module is used to calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds a preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise.

[0186] The spatial weighted feature fusion module is used to calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface.

[0187] The dynamic evolution evaluation module is used to construct a feature evolution sequence based on multi-time point physiological-anatomical fusion characteristics. It inputs a continuous-time dynamic model to fit the biological healing rate function of bone integration, solves the manifold coordinate position of the current node in the healing trajectory, and maps it to the probability of successful bone integration.

[0188] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent assessment method for implant osseointegration status, characterized in that, Includes the following steps: Step S1: Obtain postoperative cone-beam CT image data and discrete physical measurement data of the subject, perform signal normalization on the cone-beam CT image data, extract the anatomical structure feature tensor representing the bone microstructure through a deep convolutional network, establish coordinate mapping between discrete physical measurement data and image space, convert the discrete physical measurement data into physical state vectors and project them onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space. Step S2: Calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds the preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise. Step S3: Calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface. Step S4: Construct a feature evolution sequence based on multi-time point physiological-anatomical fusion features, input a continuous-time dynamic model to fit the biological healing rate function of bone integration, solve the manifold coordinate position of the current node in the healing trajectory and map it to the probability of successful bone integration.

2. The intelligent assessment method for implant osseointegration status according to claim 1, characterized in that, In step S1, the specific operation of performing signal intensity normalization processing on cone-beam CT image data is as follows: In cone-beam CT image data, a region of interest containing the implant and surrounding bone tissue is defined. The statistical distribution characteristics of all voxel gray values ​​within the region of interest are calculated to determine the first and third quartile values. Based on the first and third quartile values, the interquartile range is calculated. Extremely bright gray values ​​exceeding a preset multiple of the interquartile range are identified as non-physiological artifact signals generated by the metal beam hardening effect. Nonlinear compression processing is performed on the non-physiological artifact signals to limit their interference with the overall statistical distribution characteristics. Subsequently, the overall voxel gray values ​​after nonlinear compression are linearly mapped to a dimensionless range of zero to one to generate normalized image data that has removed the baseline drift of the acquisition device. The specific operation of extracting anatomical structural feature tensors representing the microstructure of trabecular bone through a deep convolutional feature extraction network is as follows: normalized image data is input into a fully convolutional neural network with the fully connected layers removed. An encoder containing dilated convolution operations is used to expand the receptive field while maintaining the spatial resolution of deep features. Multi-channel feature maps representing the micro-connectivity of trabecular bone and the thickness gradient of cortical bone are extracted through multi-layer convolution operations, generating anatomical structural feature tensors that encode micro-anatomical morphological information in the channel dimension.

3. The intelligent assessment method for implant osseointegration status according to claim 2, characterized in that, In step S1, the specific operation of converting discrete physical measurement data into continuous physical state vectors and projecting them onto the spatial coordinate system where the anatomical structure feature tensor resides is as follows: Construct a nonlinear mechanical state mapping function that includes stiffness-energy saturation terms, boundary integrity terms, and bone remodeling periodic terms; For the resonant frequency analysis value as discrete physical measurement data, the square value of the resonant frequency analysis value is calculated based on the square law physical relationship between vibration frequency and stiffness to characterize the interface stiffness energy. The interface stiffness energy is then processed by the sigmoid function to simulate the biological marginal effect and generate the stiffness energy saturation term. For the probe depth value, which is a discrete physical measurement data, the degree of loss of soft tissue biological barrier is calculated using an exponential decay function to generate a boundary integrity term, and a geometric constraint constant is set to force the determination of biological support failure when the probe depth value exceeds the physical critical value. For postoperative time parameters, a sinusoidal function is used to encode the periodicity of bone remodeling to generate a periodic term of bone remodeling. The stiffness energy saturation term, boundary integrity term, and bone remodeling periodic term are mapped to independent high-dimensional feature subspaces and then concatenated to generate a physical state vector. Perform a holographic spatial projection operation to copy and tile the physical state vector along the height and width dimensions of the anatomical structure feature tensor, so that each spatial voxel position of the anatomical structure feature tensor is associated with the same physical state vector, thereby generating a physical stability distribution tensor.

4. The intelligent assessment method for implant osseointegration status according to claim 1, characterized in that, In step S2, the specific operation for calculating the signal difference between the anatomical structure feature tensor and the physical stability distribution tensor is as follows: A virtual stiffness potential energy field with the same physical dimensions as the physical stability distribution tensor is generated by numerically transforming the anatomical structural feature tensor using a power-law function containing the structural stiffness exponent. Calculate the global numerical average of the physical stability distribution tensor, determine the adaptive reference coefficient based on the global numerical average, and use the adaptive reference coefficient to scale the overall energy level of the virtual stiffness potential energy field. Calculate the point-to-point Euclidean norm of the physical stability distribution tensor in the feature channel dimension and define it as the physical truth modulus. Calculate the point-to-point Euclidean norm of the virtual stiffness potential energy field in the characteristic channel dimension and define it as the theoretical expected modulus. The algebraic difference is calculated by subtracting the physical true magnitude and the preset physical tolerance threshold from the theoretical expected magnitude. When the algebraic difference is greater than zero, it is defined as the signal difference degree. When the algebraic difference is less than or equal to zero, the signal difference degree is assigned a value of zero. Unidirectional rectification operation extracts the violation energy components in the anatomical structure feature tensor that exceed the mechanical constraints of the physical stability distribution tensor, and identifies the violation energy components as signal difference.

5. The intelligent assessment method for implant osseointegration status according to claim 4, characterized in that, In step S2, the signal difference is mapped to a signal confidence mask. The signal confidence mask is used to perform signal amplitude modulation on the anatomical feature tensor to suppress regional signals where the signal difference exceeds a preset physical tolerance threshold. The specific operation of outputting the calibrated anatomical feature tensor after removing artifact noise is as follows: The signal difference is mapped to the Gaussian error decay function to calculate the probability value of the signal existence in the range of zero to one and generate a signal confidence mask. During the process of generating the signal confidence mask, the falling edge slope of the Gaussian error decay function is set using the preset artifact phase transition temperature parameter. Perform holographic amplitude modulation operation to fill the signal confidence mask in the two-dimensional coordinate system along the channel depth direction of the anatomical structure feature tensor, and generate an extended mask tensor with the same dimension as the anatomical structure feature tensor. The extended mask tensor and the anatomical feature tensor are multiplied element-wise using a Hadamard product. The feature intensity at different spatial coordinates in the anatomical feature tensor is reweighted using a signal confidence mask. At spatial coordinates where the signal difference exceeds a preset physical tolerance threshold, the amplitude intensity of the anatomical feature tensor is reduced using the numerical attenuation property of the signal confidence mask. At spatial coordinates where the signal difference is zero, the original amplitude intensity of the anatomical feature tensor is retained, and the calibrated anatomical feature tensor is output. The Hadamard product multiplication operation synchronously applies physical constraint suppression based on signal difference across all feature channels of the anatomical feature tensor.

6. The intelligent assessment method for implant osseointegration status according to claim 1, characterized in that, In step S3, the specific operation of calculating the Euclidean distance from the voxel of the calibration anatomical feature tensor space to the implant midline and constructing the spatial distance weight matrix whose value decays with distance is as follows: The implant midline is determined as the geometric reference, and the vertical Euclidean distance of each spatial voxel coordinate point in the calibration anatomical feature tensor relative to the implant midline is calculated. The mechanical impedance coupling attenuation model, which includes geometric attenuation terms and dielectric bearing terms, is used for calculation; The attenuation of the vertical Euclidean distance is calculated using an exponential function and combined with a preset stress dissipation characteristic length parameter to generate a geometric attenuation term. At spatial voxel coordinates where the vertical Euclidean distance is greater than the stress dissipation characteristic length parameter, the value of the geometric attenuation term exhibits a non-linear decrease. The point-to-point Euclidean norm of the calibration anatomical feature tensor at the corresponding spatial voxel coordinates is calculated and defined as the local feature modulus. The local feature modulus is nonlinearly mapped using the hyperbolic tangent function to generate the medium bearing term. At spatial voxel coordinates where the local feature modulus approaches zero, the value of the medium bearing term approaches zero to characterize the lack of mechanical bearing capacity at the spatial voxel coordinates. Performing multiplication operations numerically couples the geometric attenuation term with the medium bearing capacity term, generating a spatial distance weight matrix whose numerical distribution is within a closed interval from zero to one. The spatial distance weight matrix assigns a weight value representing the contribution of effective mechanical support to each spatial voxel coordinate point.

7. The intelligent assessment method for implant osseointegration status according to claim 6, characterized in that, In step S3, the specific operation of generating physiological-anatomical fusion features characterizing the effective bone integration interface by using the cross-correlation operation of the spatial distance weight matrix to constrain the physical stability distribution tensor and the calibration anatomical feature tensor is as follows: Perform physical bias cross-correlation operation to map the physical stability distribution tensor to the query vector and the calibration anatomical feature tensor to the key vector and value vector respectively; Perform a logarithmic operation on the spatial distance weight matrix to generate a physical attention bias term; Calculate the inner product between the query vector and the key vector and define the result as the original relevance score; The physical attention bias term is superimposed on the original relevance score to generate a relevance score corrected for physical bias. The normalized exponential function is used to map the correlation score after physical bias correction to determine the attention weight value of each spatial voxel coordinate point. At spatial voxel coordinate points where the spatial distance weight matrix value approaches zero, the attention weight value is suppressed by the logarithmic decay of the physical attention bias term and approaches the negative infinity direction, causing the contribution of the corresponding spatial voxel coordinate point in the subsequent weighted summation operation to return to zero. The physiological-anatomical fusion features are generated by performing a weighted summation operation on the value vector using attention weight values. Physiological-anatomical fusion features, while preserving anatomical details, numerically reflect the effective bone integration interface strength information constrained by the spatial distance weighting matrix.

8. The intelligent assessment method for implant osseointegration status according to claim 1, characterized in that, In step S4, the specific operation for constructing the feature evolution sequence based on multi-timepoint physiological-anatomical fusion features is as follows: Perform metabolic clock calibration to eliminate the interference of individual metabolic rate differences on the physical time dimension. Call the physical stability distribution tensor generated in step S1 as the source of the subject's initial physical baseline data. Calculate the global numerical average of the physical stability distribution tensor and define it as the subject's initial global mean bone mineral density. The metabolic scaling factor, which reflects the metabolic activity of the subject's bone tissue, is generated by comparing the subject's initial mean global bone mineral density with a preset standard reference density value. Multiply the physical clock time values ​​corresponding to different observation time points by a metabolic scaling factor to map the physical clock time values ​​to standardized biological metabolic time values. Physiological-anatomical fusion features corresponding to different biological metabolic time values ​​are arranged in chronological order of biological metabolic time values ​​to construct a feature evolution sequence. Iterate through each voxel value in the physical stability distribution tensor generated in step S1, extract the maximum value in the physical stability distribution tensor and define it as the host bearing capacity limit. The host carrying capacity limit, as a physical hard constraint boundary value that limits the infinite growth of physiological-anatomical fusion features during the dynamic evolution process, is used to correct the saturated nonlinear features of the feature evolution sequence when it approaches the healing endpoint.

9. The intelligent assessment method for implant osseointegration status according to claim 8, characterized in that, In step S4, the specific operation of fitting the biological healing rate function of bone integration using a continuous-time dynamic model, solving for the manifold coordinate position of the current node in the healing trajectory, and mapping it to the probability of successful bone integration is as follows: A dual-process competitive differential equation containing osteogenic growth and mechanical decay terms is constructed. The biological healing rate function is mathematically fitted using the dual-process competitive differential equation. The biological healing rate function represents the instantaneous time derivative of the manifold state vector with respect to the biological metabolic time value. In the process of calculating the osteogenic growth term, the logistic growth function is used to describe the evolution of biological stability with the numerical accumulation of biological metabolism over time, and the host carrying capacity limit is introduced to saturate the growth slope of the manifold state vector. In calculating the mechanical attenuation term, the natural exponential attenuation function is used to describe the stress relaxation process of the initial mechanical locking force of the implant with the change of biological metabolism time, and the attenuation coefficient of the mechanical attenuation term is dynamically adjusted according to the numerical fluctuation intensity of the physiological-anatomical fusion characteristics. Numerical integration is performed on the characteristic evolution sequence using a two-process competitive differential equation to solve for the manifold state vector corresponding to the current biological metabolic time value, and the coordinates of the manifold state vector in phase space are defined as the manifold coordinate position. The phase space Mahalanobis distance between the manifold coordinate position and the preset failure singularity state is calculated and defined as the safety distance, where the failure singularity state represents the unstable equilibrium point of implant detachment or fiber encapsulation. The dot product between the instantaneous evolution velocity vector of the manifold state vector and the preset ideal steady-state convergence field vector is calculated and defined as the trend health. The weighted summation and nonlinear logistic regression mapping of the safe distance and trend health are used to output the probability of successful osseointegration in the closed interval of zero to one.

10. An intelligent assessment system for implant osseointegration status, employing the intelligent assessment method for implant osseointegration status as described in any one of claims 1-9, characterized in that, include: The system includes a data acquisition and tensor construction module, a signal calibration and artifact suppression module, a spatially weighted feature fusion module, and a dynamic evolution evaluation module. The data acquisition and tensor construction module is used to acquire postoperative cone-beam CT image data and discrete physical measurement data of the subjects, perform signal normalization on the cone-beam CT image data, extract anatomical structure feature tensors that characterize bone microstructure through a deep convolutional network, establish coordinate mapping between discrete physical measurement data and image space, convert discrete physical measurement data into physical state vectors and project them onto the coordinate system where the anatomical structure feature tensor is located, and generate a physical stability distribution tensor aligned with the anatomical structure feature tensor space. The signal calibration and artifact suppression module is used to calculate the signal difference between the anatomical feature tensor and the physical stability distribution tensor, map the signal difference to a signal confidence mask, use the signal confidence mask to perform signal amplitude modulation on the anatomical feature tensor, suppress regional signals where the signal difference exceeds a preset physical tolerance threshold, and output the calibrated anatomical feature tensor after removing artifact noise. The spatial weighted feature fusion module is used to calculate the Euclidean distance from the spatial voxel of the calibration anatomical feature tensor to the implant midline, construct a spatial distance weight matrix whose value decays with distance, and use the spatial distance weight matrix to constrain the cross-correlation operation between the physical stability distribution tensor and the calibration anatomical feature tensor to generate physiological-anatomical fusion features that characterize the effective bone integration interface. The dynamic evolution evaluation module is used to construct a feature evolution sequence based on multi-time point physiological-anatomical fusion characteristics. It inputs a continuous-time dynamic model to fit the biological healing rate function of bone integration, solves the manifold coordinate position of the current node in the healing trajectory, and maps it to the probability of successful bone integration.

Images