A method and system for detecting marginal microleakage of dental restorations under multi-modal images

Abstract

The present application belongs to the technical field of dental restoration detection, and relates to a method and system for detecting marginal micro-leakage of dental restoration under multi-modal images. The present application aims to solve the problems of strong subjectivity, time lag and one-sided information of existing detection methods. The present application acquires a composite dynamic excitation signal, synchronously collects mechanical and thermal response signals, constructs a space-time correlation tensor field, and uses a parameter collaborative optimization model for fusion inversion calculation to decouple the mechanical state parameters representing the integrity of the edge interface structure and the thermal transfer state parameters representing the material transport activity. The system generates a graded detection result accordingly, clearly indicating the leakage state and risk level. The present application realizes the transition from static morphological evaluation to dynamic functional early warning, improves the early and forward nature of the detection, provides accurate diagnostic basis and differentiated management strategies for clinicians, and helps to maintain the oral health of patients.

Description

Technical Field

[0001] This invention belongs to the field of dental restoration testing technology, and relates to a method and system for detecting microleakage at the edge of dental restorations under multimodal imaging. Background Technology

[0002] Dental restorations, such as crowns, inlays, or veneers, are commonly used in dentistry to restore the shape and function of teeth. The key to their long-term success lies in maintaining a durable and tight seal between the restoration's margins and the tooth structure—that is, marginal seal. Once the marginal seal fails, microleakage occurs, allowing oral fluids, ions, and bacteria to infiltrate through the tiny gaps between the restoration and the tooth structure. This is a major cause of complications such as secondary caries, pulpitis, and restoration loss, seriously affecting the patient's oral health. Therefore, early and accurate non-destructive testing for marginal microleakage of dental restorations is of crucial clinical significance for the prevention and treatment of oral diseases.

[0003] In current clinical practice, the methods for detecting microleakage at the margins of dental restorations are relatively limited. Conventional methods mainly rely on visual examination by the dentist and physical probing along the restoration margins using sharp probes. These methods are highly subjective and can only detect mid-to-late stage leaks that have formed obvious gaps or staining. In terms of imaging, X-rays are a commonly used auxiliary examination tool, but they are not sensitive to tiny gaps parallel to the projection direction, and the limitations of two-dimensional imaging make it difficult to accurately assess the complex three-dimensional structure of tooth margins. Some studies have attempted to use a single physical modality for detection, such as using acoustic reflection or thermal imaging techniques to assess the margin status, but these methods provide only a single dimension of information and cannot comprehensively reflect the complex physical processes of microleakage.

[0004] Current detection technologies generally suffer from drawbacks such as delayed detection timing, incomplete information, and lack of functional assessment. Visual and probe examinations cannot provide early warning. Single imaging or physical detection methods can only reflect one aspect of the interface properties, such as structural continuity or thermal conductivity, and cannot capture the intrinsic relationship between structural damage and material transport, the two core elements of microleakage. Therefore, in the face of complex clinical situations, their diagnostic specificity and sensitivity are limited, making it difficult to meet the clinical need for early and accurate diagnosis of microleakage activity. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background art, a method and system for detecting microleakage at the margins of dental restorations under multimodal imaging is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: The first aspect of this invention provides a method for detecting microleakage at the edge of a dental restoration under multimodal imaging, comprising: S1, acquiring a composite dynamic excitation signal for stimulating a coupling response at the edge interface of the target dental restoration.

[0007] S2. Based on the composite dynamic excitation signal, the first response signal and the second response signal generated by the coupling response are acquired simultaneously to generate a dual-mode response signal set.

[0008] S3. Based on the dual-mode response signal set, construct a spatiotemporal correlation tensor field containing the first response signal, the second response signal, and intermodal correlation information representing the correlation information between the two.

[0009] S4. Input the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization for fusion inversion calculation, and decouple the mechanical state parameters that characterize the structural integrity of the edge interface and the heat transfer state parameters that characterize the material transport activity.

[0010] S5. Based on the combined relationship between mechanical state parameters and heat transfer state parameters, generate detection results characterizing the leakage state of the edge interface.

[0011] A second aspect of the present invention provides a multimodal imaging system for detecting microleakage at the edge of a dental restoration, comprising: a composite dynamic excitation signal acquisition module for acquiring a composite dynamic excitation signal used to excite a coupling response at the edge interface of the target dental restoration.

[0012] The dual-mode response signal set generation module, based on the composite dynamic excitation signal, synchronously acquires the first response signal and the second response signal generated by the coupled response to generate a dual-mode response signal set.

[0013] The spatiotemporal correlation tensor field construction module constructs a spatiotemporal correlation tensor field based on a dual-mode response signal set. This field includes a first response signal, a second response signal, and intermodal correlation information representing the correlation between the two.

[0014] The neural network model decoupling module inputs the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization for fusion and inversion calculation, decoupling the mechanical state parameters characterizing the structural integrity of the edge interface and the heat transfer state parameters characterizing the material transport activity.

[0015] The edge leakage detection result generation module generates detection results characterizing the leakage state of the edge interface based on the combination relationship between mechanical state parameters and heat transfer state parameters.

[0016] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) This invention, through a composite dynamic excitation and dual-mode synchronous acquisition mechanism, can simulate the functional load of the real oral environment under non-invasive conditions, thereby stimulating a coupled physical response directly related to early microleakage. This detection method eliminates the dependence on existing macroscopic structural defects, enabling the detection of subtle signs of deterioration in the sealing performance of the edge interface before the appearance of clinical symptoms. It realizes the transformation from static morphological assessment to dynamic functional early warning, improving the earlyness and prospectiveness of detection.

[0017] (2) This invention establishes a data processing framework for constructing a spatiotemporally correlated tensor field and using a parameter collaborative optimization model for fusion and inversion. This framework can deeply decouple intertwined multimodal signals into mutually independent mechanical state parameters and heat transfer state parameters. This intelligent decoupling based on strong physical correlation constraints has significantly better accuracy than traditional signal processing or simple feature fusion, effectively suppresses interference from single-modal noise and artifacts, ensures high reliability and accuracy of detection results, and greatly reduces the risk of clinical misjudgment and missed diagnosis.

[0018] (3) This invention generates multi-dimensional, graded detection results by combining and analyzing the decoupled mechanical and thermal transfer state parameters. These results not only clearly indicate whether leakage exists, but also distinguish between the "risk period" where only structural defects exist and the "active period" where material exchange has already occurred. This makes it possible to adopt differentiated management strategies for restorations in different states, optimizes the allocation of medical resources, and helps to achieve long-term health maintenance of patients' restorations. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0021] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0022] 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.

[0023] Please see Figure 1 The first aspect of the present invention provides a method for detecting microleakage at the edge of a dental restoration under multimodal imaging, comprising: S1, acquiring a composite dynamic excitation signal for stimulating a coupling response at the edge interface of the target dental restoration.

[0024] In a specific embodiment of the present invention, obtaining a composite dynamic excitation signal for stimulating a coupling response at the edge interface of the target tooth restoration includes: generating a dynamic mechanical excitation signal simulating alternating load.

[0025] Generate a transient thermal excitation signal that is synchronously triggered at a specific phase point of the dynamic mechanical excitation signal.

[0026] The dynamic mechanical excitation signal and the transient thermal excitation signal are coupled in a programmed manner to generate a composite dynamic excitation signal.

[0027] In one specific embodiment of the present invention, the device for applying the composite dynamic excitation signal is an integrated probe. The integrated probe includes a piezoelectric ceramic actuator (PZT) for generating micron-level mechanical vibration (dynamic mechanical excitation), and a miniature laser diode or miniature resistance heating coil integrated on or coaxially positioned at the front end of the piezoelectric ceramic for generating pulsed thermal radiation (transient thermal excitation). The points of application of both are focused at the same detection point through an optical lens or mechanical conduit, ensuring spatial overlap between the mechanical and thermal excitations.

[0028] Specifically, the engineering objective is to generate a composite dynamic excitation signal that highly simulates the combined effect of chewing force and food temperature changes in a real oral environment, serving as the input source for subsequent detection steps. This process is executed by the system's excitation generation unit and consists of the following engineering steps.

[0029] First, the system's control core aims to generate a dynamic mechanical excitation signal that simulates physiological chewing loads. This process begins by retrieving waveform definition parameters from a pre-set parameter library to construct a time-varying mechanical excitation function. In engineering, this dynamic mechanical excitation signal... This can be represented by the following mathematical model: ; Here, Representative at The amplitude of the drive signal constantly output to the mechanical brake; The amplitude of the signal, after calibration, corresponds to the force applied to the surface of the dental restoration. This force is typically set between 1 and 10 Newtons to simulate normal chewing force. The excitation frequency was set within the physiological range of 0.5 Hz to 5 Hz to reproduce the rate of chewing. This is the initial phase angle, used for synchronization control; This represents the DC bias value, indicating the applied preload to ensure continuous contact between the probe and the tooth surface. These parameters are obtained from the system's configuration file.

[0030] Next, the system generates a transient thermal excitation signal that is synchronously triggered at a specific phase point of the dynamic mechanical excitation signal. Its engineering objective is to apply a brief thermal shock at a critical moment in the mechanical excitation cycle. The system monitors the dynamic mechanical excitation signal in real time. phase When phase Achieve the preset trigger phase angle When, it is usually set to At the peak point corresponding to the mechanical loading, the system immediately triggers the generation of a thermal excitation signal. This transient thermal excitation signal... In engineering, it is represented as an impulse function.

[0031] ; Here, It is to satisfy Each discrete time point; yes The driving signal power constantly output to the thermal actuator; It is the pulse power amplitude, the magnitude of which is set to produce an instantaneous temperature rise of 2 to 5 degrees Celsius at the edge of the restoration; It is the duration of the heat pulse, which is usually controlled between 10 milliseconds and 100 milliseconds to simulate the brief contact between food heat and teeth.

[0032] In one specific embodiment of the present invention, thermal excitation is applied at the peak of mechanical loading. The technical reason for this is that at this point, the interface between the restoration and the tooth structure is under maximum compressive stress, and any potential micro-gaps or poor contact areas will be instantly compressed to a minimum under this pressure. Applying a thermal pulse at this moment can maximize the highlighting of local thermal response differences caused by the presence of fluid within the gap (whose thermal conductivity differs greatly from that of solid materials), thereby significantly enhancing the detection sensitivity of microleakage signals.

[0033] Finally, the system programmatically couples the dynamic mechanical excitation signal and the transient thermal excitation signal in the time domain to generate a composite dynamic excitation signal. This coupling is not a direct superposition of physical quantities, but rather the integration of two independent control signals into a synchronized time-series vector. This composite dynamic excitation signal... In engineering, it is represented as a two-dimensional vector.

[0034] ; Here, It is the final generated composite excitation signal vector, and its first component The mechanical actuator sent to the excitation probe, the second component The thermal actuator is synchronously sent to the probe. Through this programmed coupling method, the system ensures that at each peak of mechanical loading, an accurately synchronized micro-thermal pulse is applied to the same detection area, thereby generating a highly realistic composite physical excitation to stimulate a coupled response at the edge interface of the dental restoration.

[0035] S2. Based on the composite dynamic excitation signal, the first response signal and the second response signal generated by the coupling response are acquired simultaneously to generate a dual-mode response signal set.

[0036] In a specific embodiment of the present invention, the first response signal and the second response signal generated by the coupling response are acquired synchronously to generate a dual-mode response signal set, including: using the periodic events of the dynamic mechanical excitation component in the composite dynamic excitation signal as the main triggering reference.

[0037] Based on the main trigger reference, the timing sequence for acquiring the first response signal and the second response signal is started synchronously.

[0038] Based on the shape of the first acquired response signal, the timing of subsequent acquisitions is adaptively fine-tuned to generate an aligned dual-mode response signal set.

[0039] Specifically, the project aims to establish an adaptive synchronous acquisition mechanism capable of resisting interference from factors such as physiological micro-movements in patients, ensuring high alignment accuracy of the data required for subsequent fusion analysis in the time domain. This process is executed by the system's synchronous acquisition and control unit, and consists of the following engineering steps.

[0040] First, the system uses the periodic events of the dynamic mechanical excitation component in the composite dynamic excitation signal as the primary triggering reference. This aims to replace the external, easily disturbed absolute clock with stable physical events generated internally by the system. In engineering, the control unit continuously monitors the dynamic force... stimulus signal The output waveform. When When the excitation signal waveform meets the preset triggering conditions, such as crossing its DC bias point from a negative value. At the zero-crossing point, the system generates a high-precision digital pulse signal. This pulse sequence is defined as the master trigger reference. ,in Representing the One incentive cycle.

[0041] Next, based on the main trigger reference, the system synchronously initiates the timing sequence for acquiring the first and second response signals. At each main trigger reference... At the appointed time, the synchronous acquisition unit simultaneously sends acquisition start commands to both the mechanical response sensors (such as laser Doppler vibration meters) and the thermal response sensors (such as infrared thermal imagers). The two sensors then begin acquisition within a preset time window. The data is recorded internally, with the acquisition time window typically set to 1.2 to 1.5 times the excitation period to ensure complete capture of the response and decay process of one cycle. This step produces a raw first response signal data block that is strictly aligned to the time start point. and the original second response signal data block .

[0042] Finally, based on the morphology of the acquired first response signal, the system adaptively fine-tunes the start time of subsequent acquisition sequences. Its engineering goal is to dynamically compensate for the minute phase drift between the response and excitation signals caused by tissue viscoelasticity or patient micro-movements through a feedback loop. After each acquisition cycle is completed, the fast signal processing algorithm is immediately invoked to analyze the raw first response signal data block. The algorithm can quickly locate a stable feature, such as the time when the first major peak of the signal occurs. The system will use this actual peak time. With a preset ideal peak reference time By comparison, the timing deviation for this period can be calculated. .

[0043] ; Here That is, the first The system calculates the response delay or lead for each cycle. This deviation value is used to calculate a timing correction to be applied to the next cycle. A first-order low-pass filter is typically used to smooth the correction process and avoid system oscillations.

[0044] ; Here, It is a smoothing coefficient between 0 and 1, for example, 0.2. In the... Each acquisition cycle, although the system is still composed of Triggered, but during post-processing, the collected data blocks will be shifted on the timeline. This allows for the precise alignment of the first and second response signals. Through this adaptive fine-tuning, the system ensures that key response feature points across different cycles can be accurately aligned even in the presence of dynamic disturbances, providing a foundation for the subsequent construction of a high-quality spatiotemporal correlated tensor field.

[0045] S3. Based on the dual-mode response signal set, construct a spatiotemporal correlation tensor field containing the first response signal, the second response signal, and intermodal correlation information representing the correlation information between the two.

[0046] In one specific embodiment of the present invention, before constructing the dual-mode response signal set, the first response signal and the second response signal need to be spatially registered. Specific methods include: using a calibration plate that is simultaneously sensitive to visible and infrared light for system calibration, obtaining the transformation matrix (homography matrix) between the mechanical sensor coordinate system and the thermal sensor coordinate system; based on this transformation matrix, using a bilinear interpolation algorithm to map the low-resolution thermal response image or point-scanned mechanical response data to a unified high-resolution grid coordinate system, ensuring that data at the same location in the tensor field corresponds to the same physical point on the tooth.

[0047] In a specific embodiment of the present invention, constructing a spatiotemporal correlation tensor field that includes a first response signal, a second response signal, and intermodal correlation information representing the correlation between the two includes: extracting the structural mechanical response data sequence corresponding to the first response signal and the heat transfer response data sequence corresponding to the second response signal from the dual-mode response signal set.

[0048] In one specific embodiment of the present invention, the structural mechanical response data is preferably a vibration velocity signal, as it is more sensitive to changes in structural damping; the heat transfer response data is preferably a temperature rise relative to the initial reference temperature (…). The signal is used to eliminate interference from ambient temperature fluctuations. Before integrating them into a tensor field, the two data sequences are normalized, for example using... Standardization or maximum / minimum normalization is used to eliminate dimensional differences and ensure the stability of neural network model training.

[0049] Calculate the real-time correlation strength between the structural mechanical response data sequence and the thermal transfer response data sequence at various points in time and space, and generate a modal correlation strength data sequence.

[0050] In one specific embodiment of the present invention, the physical basis of the intermodal correlation information lies in the fact that when microleakage channels exist at the edge of the repair, the presence of fluid will simultaneously cause changes in local structural damping (affecting mechanical vibration attenuation) and abrupt changes in heat capacity / thermal conductivity (affecting heat wave transmission). Therefore, in the leakage region, the amplitude change of the mechanical response and the temperature decay of the thermal response exhibit a specific strong correlation on the time axis (e.g., synchronous phase lag); while in the structurally intact region, due to the lack of coupling effect of the fluid medium, the synchronicity of this dynamic change is weaker.

[0051] In a specific embodiment of the present invention, the real-time correlation strength between the structural mechanics response data sequence and the heat transfer response data sequence at various points in time and space is calculated to generate a modal correlation strength data sequence, including: at the corresponding points in time and space of the structural mechanics response data sequence and the heat transfer response data sequence, short time segments are extracted using a sliding time window.

[0052] Calculate the mutual information or covariance coefficient between the structural mechanical response data and the thermal transfer response data within each short time segment to obtain the quantized correlation value.

[0053] The quantized correlation values ​​are used as the real-time correlation strength at spatiotemporal points to form a modal correlation strength data sequence.

[0054] Specifically, the engineering objective is to accurately calculate the modal correlation strength, that is, to transform the abstract concept of physical correlation into quantifiable data. This process is a key step in constructing the spatiotemporal correlation tensor field, revealing the dynamic coupling relationship between mechanical and thermal responses point by point and time by time through a local, sliding calculation mechanism.

[0055] The calculation process begins with localizing the input structural mechanics response data sequence and the thermal transfer response data sequence. The system at each spatial coordinate point... Up, along the time axis Define and extract a short time sequence segment. This is a sliding time window operation, where the length of the window is... A key preset engineering parameter is typically set between 5 and 15 time sampling points. The choice of window length needs to strike a balance between ensuring statistical stability and maintaining high temporal resolution. This step is performed for each time point. Extract two data sequences of length from the original data sequence. The vector, that is, with Short time segment of structural mechanical response centered on and short time segments of heat transfer response .

[0056] Next, the system performs correlation calculations on each extracted short time-series segment. This embodiment uses the covariance coefficient as a quantitative indicator because it effectively characterizes the linear correlation between two variables. For spatial points... Time point The covariance of a pair of short time segments is... Calculate using the following formula.

[0057] ; In this formula, the summation symbol This indicates that all data points within the window are summed, and their index is used for accumulation. Iterate from 1 to . That is, the length of the sliding time window, which is a dimensionless integer. and They are the first The structural mechanical response and thermal transfer response values ​​at each sampling time are obtained directly from the corresponding data sequence. and These are the two short time segments respectively. and The arithmetic mean of these values ​​is calculated instantly within the current window. The result of this formula... It has a clear physical meaning; its positive or negative sign indicates a positive or negative correlation, and its absolute value indicates the strength of the correlation.

[0058] Finally, the system will calculate the covariance coefficients. This serves as the modal correlation strength data for that spatiotemporal point. This means that the calculation results are directly assigned to the corresponding modal correlation strength data sequence. The elements of location. By repeatedly performing the above windowing and covariance calculation throughout the entire spatiotemporal domain, the system generates a complete sequence of modal correlation strength data point by point. This provides a third key data channel for the subsequent construction of the spatiotemporal correlation tensor field.

[0059] The structural mechanics response data sequence, thermal transfer response data sequence, and modal correlation strength data sequence are integrated and their dimensions are reconstructed to generate a spatiotemporal correlation tensor field.

[0060] Specifically, the engineering objective is to transform the temporally aligned bimodal raw signal streams acquired in the preceding steps into a structured, high-dimensional information carrier—a spatiotemporal correlated tensor field. This data structure not only preserves the original information but also explicitly encodes the physical correlation between the two physical modes in local spatiotemporal space, providing crucial input for subsequent deep fusion and inversion. This process is executed by the system's data processing module and consists of the following engineering steps.

[0061] First, the system extracts the structural mechanical response data sequence corresponding to the first response signal and the heat transfer response data sequence corresponding to the second response signal from the dual-mode response signal set. In engineering terms, this means preprocessing and formatting the raw data output by the sensors. For example, for data acquired by a two-dimensional infrared thermal imager and a two-dimensional ultrasonic vibration imager, the system registers and normalizes consecutive image frames within each acquisition cycle, generating two three-dimensional matrices. (Structural mechanical response data sequence) Represents spatial coordinates Discrete time points The vibration amplitude or velocity. Heat transfer response data sequence. This represents the temperature change at the same spatiotemporal coordinate point. Both data sequences have the same spatial resolution, e.g., 50 to 100 micrometers per pixel, and the same temporal sampling rate, e.g., 100 to 500 frames per second.

[0062] Next, the system calculates the real-time correlation strength between the structural mechanical response data sequence and the thermal transfer response data sequence at various points in space and time, generating a modal correlation strength data sequence. Its engineering goal is to quantify whether mechanical and thermal changes occur synchronously or independently in every tiny region and at every instant. To this end, the system at each spatial point... Up, along the time axis Apply a short-lived sliding window. Let the window length be... Each time sampling point, for example, 5 to 10 points. Within the central window, the system calculates and The covariance of two signal segments.

[0063] ; Here, and Represents respectively in Point, Time Window Vector segments of the internal structural mechanical response and thermal transfer response. Calculated... This is the real-time correlation strength value at that spatiotemporal point, which constitutes a correlation with... and A new three-dimensional matrix with the same dimensions is formed, which is the modal correlation strength data sequence. The numerical values ​​of this sequence directly reflect the degree of local coupling between the two physical effects.

[0064] Finally, the system integrates and reconstructs the dimensions of the structural mechanics response data sequence, thermal transfer response data sequence, and modal correlation strength data sequence to generate a spatiotemporal correlation tensor field. This step is a data stacking operation designed to create a unified multi-channel data structure suitable for use as input to subsequent neural network models. In engineering terms, the system creates a four-dimensional tensor. Their dimensions are space and feature channels .

[0065] ; Among them, when the feature channel index hour, The value is assigned ;when hour, The value is assigned ;when hour, The value is assigned Through this dimensional reconstruction, the two original independent data sequences and their derived correlation information are integrated into a single, structured spatiotemporal correlation tensor field. Each element of this tensor field simultaneously contains composite information about mechanics, thermodynamics, and their interactions.

[0066] S4. Input the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization for fusion inversion calculation, and decouple the mechanical state parameters that characterize the structural integrity of the edge interface and the heat transfer state parameters that characterize the material transport activity.

[0067] In one specific embodiment of the present invention, in order to train the neural network model, a training dataset containing multimodal inputs and corresponding gold standard labels is constructed. The training dataset is obtained by: constructing a finite element simulation model of the dental restoration, setting different interface strength values ​​and microleakage channel parameters as labels, and simulating and generating corresponding vibration and thermal response signals as input data; or conducting experiments using extracted teeth, and after collecting bimodal response signals, obtaining real interface structure integrity data and dye penetration depth data as labels through slice microscopy or micro-stretching tests.

[0068] In a specific embodiment of the present invention, the spatiotemporal correlation tensor field is input into a neural network model for collaborative optimization and fusion inversion calculation is performed to decouple the mechanical state parameters characterizing the structural integrity of the edge interface and the heat transfer state parameters characterizing the material transport activity. This includes: using the mechanical parameter prediction channel of the neural network model, and based on the spatiotemporal correlation tensor field, to preliminarily predict the mechanical state parameter map.

[0069] By using the heat transfer parameter prediction channel of the neural network model, and based on the spatiotemporal correlation tensor field, a preliminary prediction of the heat transfer state parameter map is made.

[0070] In one specific embodiment of the present invention, the mechanical state parameter is defined as the equivalent interfacial bonding strength (unit: ). ) or normalized structural integrity index (dimensionless value between 0 and 1); the heat transfer state parameter is defined as the equivalent thermal diffusivity (unit: ) or the normalized liquid permeability activity index (a dimensionless value between 0 and 1).

[0071] Specifically, the project aims to perform deep analysis of the input spatiotemporal correlated tensor field using a specially designed, pre-trained parameter co-optimization model, thereby accurately decoupling and reversing the mechanical and heat transfer states hidden behind complex signals. The core of this process lies in utilizing the correlation information between modes as strong constraints to achieve accurate inference beyond single-modal analysis.

[0072] The process begins with preliminary predictions through the mechanical parameter prediction channel of the parameter collaborative optimization model. In engineering terms, the mechanical parameter prediction channel is a branch of a deep convolutional neural network. It receives the complete spatiotemporal correlated tensor field. As input, but its network weights, after training, mainly target... Structural mechanical response data sequence Modal correlation strength data sequence More sensitive. This channel extracts deep features related to structural integrity from the input spatiotemporal data through a series of convolutional, pooling, and nonlinear activation layers, ultimately outputting a two-dimensional image with the same spatial dimension as the input data, i.e., a preliminary prediction of mechanical state parameters. Each pixel value in the figure represents a quantitative indicator of structural integrity at the corresponding location, such as equivalent bond strength.

[0073] In parallel, the system performs another preliminary prediction through the heat transfer parameter prediction channel of the parameter collaborative optimization model. This channel is also a branch of a deep convolutional neural network, with a structure similar to the mechanical parameter prediction channel but with independent weights. It also receives the complete spatiotemporal correlation tensor field. As input, but its weights are trained as the main focus Heat transfer response data sequence Modal correlation strength data sequence The task of this branch is to extract features related to mass transport activity and output a preliminary predicted graph of heat transfer state parameters. The pixel values ​​in the image represent the heat transfer efficiency at the corresponding location, used to characterize the presence of microleakage channels.

[0074] The final and crucial step involves the system jointly optimizing the two preliminary prediction results using the built-in correlation consistency constraint function of the parameter collaborative optimization model. This function is central to the entire model training and inference process. It defines an objective function. The model minimizes the value of this function during runtime.

[0075] ; In this formula, and These are traditional loss functions, used to measure... and The difference from the true labels (provided during training) can be addressed by using cross-entropy loss, for example. This is the core of the invention, namely the association consistency loss function. It is a hyperparameter used to balance the weights of various losses, and is usually set between 0.1 and 1.0. The design goal is to be in The region with high values ​​indicates a strong correlation between mechanical and thermal responses. and The predicted anomaly regions must highly overlap spatially. In engineering, This can be achieved by calculating a weighted Dice coefficient loss or a similar spatial overlap index.

[0076] ; Here, the summation symbol This indicates that for all spatial points Accumulate. It is a two-dimensional weighted graph obtained by integrating or taking the maximum value of the modal correlation strength data sequence over the time dimension. The function is used to calculate the two prediction plots in Local spatial overlap at a point. By minimizing The model is forced to learn an internal representation that can not only explain the mechanical and thermal signals separately, but also satisfy the inherent consistency between the two in their physical correlation. After this joint optimization process, the model finally outputs optimized mechanical state parameters. and heat transfer state parameters Its accuracy and reliability are significantly better than the results of independent predictions from the two channels.

[0077] In one specific embodiment of the present invention, when calculating the association consistency constraint function, in order to ensure the differentiability of the loss function to achieve backpropagation, binarization processing is adopted. function or The function performs a soft threshold approximation, making the predicted abnormal region distribution map a continuous probability distribution map rather than a discrete 0 / 1 mask, thus allowing the gradient to be backpropagated to the neural network model through the overlap calculation formula.

[0078] By utilizing the correlation consistency constraint function based on the intermodal correlation information in the spatiotemporal correlation tensor field, the mechanical state parameter diagram and the heat transfer state parameter diagram are jointly optimized, and the optimized mechanical state parameters and heat transfer state parameters are output.

[0079] In a specific embodiment of the present invention, the process of constructing the correlation consistency constraint function based on the correlation information between modes in the spatiotemporal correlation tensor field includes: calculating the first spatial distribution map of the abnormal region in the mechanical state parameter map.

[0080] The second spatial distribution diagram of the abnormal region in the calculation of heat transfer state parameters.

[0081] The overlap between the first spatial distribution map and the second spatial distribution map is calculated to obtain the overlap map.

[0082] Based on the intermodal correlation information in the spatiotemporal correlation tensor field, differentiated weights are assigned to different spatial locations of the overlap graph to construct an objective function for joint optimization.

[0083] Specifically, the project aims to elaborate on the technical implementation details of the correlation consistency constraint function, which is the core mechanism for the parameter collaborative optimization model to achieve a "1+1>2" synergistic effect. By applying differentiated constraint strengths to the prediction results at different spatial locations, the model is guided to focus on the regions that are most likely to experience collaborative failure, thereby improving the accuracy of decoupling.

[0084] The execution of this function begins with the binarization of two preliminary prediction results. The system first calculates the first spatial distribution map of the anomalous regions in the preliminary predicted mechanical state parameter map. In engineering terms, this means that the preliminary predicted mechanical state parameter map... Compared with the preset mechanical anomaly threshold Compare. All pixel values ​​higher than The location is marked as 1 (abnormal), and the rest are marked as 0, thus generating a binary mechanical state anomaly map. Similarly, the system calculates the second spatial distribution map of the anomalous regions in the preliminary predicted heat transfer state parameter diagram. This is achieved by mapping the preliminary predicted heat transfer state parameters. With thermal anomaly threshold It was obtained through comparison.

[0085] In a specific embodiment of the present invention, a preset mechanical anomaly threshold is defined. An empirical value, such as 0.7, can be chosen. The threshold is not arbitrarily set, but rather determined through statistical analysis of a validation dataset containing a large number of healthy samples and samples with known defect levels. This involves plotting receiver operating characteristic (ROC) curves and calculating the Youden index to select the parameter value that maximizes the ability to distinguish between healthy and abnormal states. A value of 0.7 means that when the integrity of the local interface structure predicted by the neural network model decreases to below 70% of the ideal state, the area is initially marked as a risk area with mechanical abnormalities. This setting aims to ensure high sensitivity for detecting early, minor structural defects while maintaining specificity to avoid excessive false positives.

[0086] Next, the system calculates the spatial overlap between the first and second spatial distribution maps. This step aims to quantify the consistency between the two modalities in predicting anomalous regions. In engineering practice, the most direct calculation method is to perform a logical AND operation pixel by pixel, i.e. However, in order to obtain a differentiable index that reflects local similarity, this embodiment calculates a local... Coefficient. The system at each pixel. Define a small neighborhood window around the perimeter, such as 3x3 or 5x5 pixels, and then calculate the area within that window. and of coefficient.

[0087] ; Here, It is at point The degree of local overlap. and It is the set of pixels within the neighborhood window of that point. This indicates the number of pixels in the set. The intersection of sets. The calculated... This forms an overlap map with values ​​ranging from 0 to 1. Higher values ​​indicate better local overlap.

[0088] Finally, and most crucially, the system assigns differentiated weights to the overlap calculation at different spatial locations based on modal correlation strength data in the spatiotemporal correlation tensor field. Its engineering goal is to enable the model to impose a greater penalty on inconsistencies in prediction results in regions of strong physical correlation. First, the system needs to process the three-dimensional modal correlation strength data sequence... Generate a two-dimensional weight graph. This is typically done through the time dimension. This is achieved by taking the maximum or average value. Then, the system applies this weighted graph to the association consistency loss. In the calculation.

[0089] ; In this formula, the summation symbol Represents all spatial pixels Accumulate. This is the final association consistency loss value. It can be seen that, in terms of modal association strength... The higher the area, This term, namely the penalty for non-overlap, is also weighted and amplified. Therefore, during model training, in order to minimize... The model must prioritize ensuring a high degree of consistency between the anomalous regions predicted by mechanical and thermal forces in these highly correlated areas. This weighting mechanism effectively encodes prior physical knowledge (i.e., leakage causes simultaneous changes in both mechanical and thermal properties) into the model's optimization objective.

[0090] S5. Based on the combined relationship between mechanical state parameters and heat transfer state parameters, generate detection results characterizing the leakage state of the edge interface.

[0091] In a specific embodiment of the present invention, a detection result characterizing the leakage state of the edge interface is generated based on the combination relationship between mechanical state parameters and heat transfer state parameters, including: comparing the mechanical state parameters with a threshold characterizing structural integrity to generate a mechanical state anomaly map.

[0092] In one specific embodiment of the present invention, the structural integrity threshold can be set to 0.7. This value has clear physical meaning and statistical basis: it means that when the structural integrity index of a local area decreases by more than 30% from the ideal state (value 1), the system determines that the area has a risk of structural defects. Through threshold judgment, the system can transform a continuous parameter distribution map into an intuitive mechanical state anomaly map containing only "normal" and "abnormal" states. All areas with parameter values ​​below 0.7 are highlighted as abnormal, providing a clear spatial location basis for subsequent fusion analysis with the heat transfer state.

[0093] The heat transfer state parameters are compared with thresholds characterizing mass transport activity to generate a heat transfer state anomaly map.

[0094] Spatial logic analysis is performed on the mechanical state anomaly map and the heat transfer state anomaly map to identify regions with different risk levels, and graded detection results are generated based on these regions.

[0095] Specifically, the project aims to transform the continuous mechanical and thermal transfer parameters output by the parameter co-optimization model into intuitive test results with clear clinical guidance. This process, through setting thresholds, performing logical judgments and spatial analysis, ultimately achieves accurate classification and location of microleakage risks.

[0096] The process first binarizes the two decoupled parametric maps to identify potential anomaly regions. The system then optimizes the mechanical state parameters. Compared with a pre-defined structural integrity threshold derived from clinical database statistics A comparison is made. This threshold represents the minimum acceptable level of structural integrity. Through pixel-by-pixel comparison, the system generates a binary map of mechanical state anomalies. A pixel value of 1 indicates that the structural integrity at that location is below the threshold, and there may be weak adhesion or microcracks.

[0097] ; In parallel, the system will optimize the heat transfer state parameters. With preset material transport activity threshold A comparison is then made. This threshold represents the critical point that distinguishes normal heat conduction from anomalous heat convection (usually caused by liquid leakage). After the comparison, the system generates a binary anomaly map of heat transfer states. A pixel value of 1 indicates abnormal material transport activity at that location, suggesting a possible leakage channel.

[0098] ; Next, the system overlays and performs spatial logical analysis on the two binary anomaly maps to identify regions with different risk levels. This step is crucial for achieving refined diagnosis, leveraging the advantages of combining bimodal information. The system then... and Perform pixel-by-pixel logical operations to generate multiple risk area maps.

[0099] ; ; ; Here, This represents a highly active leakage zone, which is a region that simultaneously exhibits structural defects and abnormal material transport, and is the highest priority for clinical intervention. This represents a structural risk zone, which is an area where there are only structural integrity issues but no active leakage channels have yet formed, indicating that close monitoring is needed. This may represent an abnormality caused solely by thermal artifacts or unrelated heat sources, and its clinical significance is low due to the lack of evidence of mechanical structural abnormalities.

[0100] In one specific embodiment of the present invention, for a preset structural integrity threshold =0.6: It acts on the normalized (0 to 1) mechanical state parameter map, where 1 represents a perfect structure. This threshold means that when the structural integrity index of a region drops below 60% of the ideal state, the system will ultimately identify it as a clinically significant structural defect.

[0101] For the preset mass transport activity threshold =0.65: This value applies to the normalized (0 to 1) heat transfer state parameter map, where 1 represents the maximum detectable leakage activity. This threshold means that when the mass transfer (i.e., liquid flow) activity in a region exceeds 65% of the maximum value, the system will ultimately identify it as having a functional leak.

[0102] Finally, the system generates tiered detection results based on the above identification results. This means that... , The logical partitioning diagram is transformed into a comprehensive risk map. For example, the system can generate a pseudo-color image, marking highly active leakage areas in red and structural risk areas in yellow, and overlaying it onto the original restoration image. Simultaneously, the system calculates quantitative indicators such as the area and location of each risk area, and automatically generates a text report containing conclusions such as "immediate intervention recommended," "regular follow-up recommended," or "good condition" based on preset clinical guidelines, such as if the area of ​​a highly active leakage area exceeds a certain threshold (e.g., 0.5 square millimeters). This visually appealing test result provides a comprehensive and intuitive basis for clinical decision-making.

[0103] Acquire static image data of the marginal interface of the target tooth restoration.

[0104] The detection results are spatially registered with the static image data to generate a registered fusion dataset.

[0105] Based on the fused dataset, mechanical state parameters, heat transfer state parameters, and leakage risk levels determined by the combination of the two are superimposed and labeled on static image data using different visual elements.

[0106] Specifically, the engineering objective is to precisely integrate the abstract functional test results generated in the preceding steps with specific anatomical structures, thereby providing clinicians with an intuitive, interactive, augmented reality view overlaid with functional information. This process is executed by the system's visualization and interaction module.

[0107] The process begins with acquiring static image data of the marginal interface of the target dental restoration. In engineering, this is typically done before or after dynamic inspection using a separate, high-precision 3D imaging device. For example, the system activates and controls an intraoral 3D scanner to capture the surface geometry of the dental restoration and its adjacent areas, generating a 3D point cloud or triangular mesh model with high spatial resolution, typically between 10 and 50 micrometers; this model is the static image data.

[0108] Next, the system spatially registers and overlays the detection results with the static image data. The core of this step is resolving the alignment issue between data in two independent coordinate systems. The system first automatically extracts several stable anatomical feature points from the static image data, such as cusp tips and developmental groove intersections. Simultaneously, during dynamic detection, the system's probe positioning module records its spatial trajectory in its own coordinate system, thereby constructing the spatial framework of the detection result atlas. The system then applies an iterative nearest-point algorithm, i.e., the ICP algorithm, to solve for a rigid transformation matrix that can convert the detection result coordinate system to the static image data coordinate system. .

[0109] ; In this formula, It is the three-dimensional coordinate vector of any point in the detection result. It is the coordinate vector corresponding to its position in the static image data coordinate system. It is a 4x4 rigid transformation matrix, determined by minimizing the sum of squared Euclidean distances between corresponding feature points in the two sets of data. Once solved... The system can then accurately "attach" the entire detection result map onto the three-dimensional tooth model.

[0110] Finally, in the overlay display, the system annotates the mechanical state parameters, thermal transfer state parameters, and the leakage risk level determined by their combination using different visual elements. In engineering implementation, this is a multi-layer rendering process. The system first renders the underlying static image data using a semi-transparent material that simulates tooth enamel. Then, on the registered restoration edge interface, the system continuously renders the mechanical state parameters... and heat transfer state parameters The system renders heatmaps using two independent color chromatograms, one from blue to red and the other from green to yellow, creating two independently switchable semi-transparent layers. At the top layer, the system overlays discrete, graded leakage risk levels, such as highly active leakage areas, as opaque, flashing red highlights, while highlighting structural risk areas with outlining. Through this multi-layered, multi-channel visual encoding, clinicians can simultaneously and clearly observe anatomical structures, the continuous distribution of two physical parameters, and the final risk assessment conclusion.

[0111] Reference Figure 2 The second aspect of the present invention provides a multimodal imaging-based system for detecting microleakage at the edge of dental restorations, comprising: a composite dynamic excitation signal acquisition module, a dual-mode response signal set generation module, a spatiotemporal correlation tensor field construction module, a neural network model decoupling module, and an edge leakage detection result generation module.

[0112] The composite dynamic excitation signal acquisition module is connected to the dual-mode response signal set generation module, the dual-mode response signal set generation module is connected to the spatiotemporal correlation tensor field construction module, the spatiotemporal correlation tensor field construction module is connected to the neural network model decoupling module, and the neural network model decoupling module is connected to the edge leakage detection result generation module.

[0113] The composite dynamic excitation signal acquisition module acquires composite dynamic excitation signals used to excite the target tooth restoration edge interface to generate a coupling response.

[0114] The dual-mode response signal set generation module, based on the composite dynamic excitation signal, synchronously acquires the first response signal and the second response signal generated by the coupled response to generate a dual-mode response signal set.

[0115] The spatiotemporal correlation tensor field construction module constructs a spatiotemporal correlation tensor field based on a dual-mode response signal set. This field includes a first response signal, a second response signal, and intermodal correlation information representing the correlation between the two.

[0116] The neural network model decoupling module inputs the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization for fusion and inversion calculation, decoupling the mechanical state parameters characterizing the structural integrity of the edge interface and the heat transfer state parameters characterizing the material transport activity.

[0117] The edge leakage detection result generation module generates detection results characterizing the leakage state of the edge interface based on the combination relationship between mechanical state parameters and heat transfer state parameters.

[0118] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for detecting marginal microleakage of dental restorations under multimodal imaging, characterized in that, include: S1. Acquire a composite dynamic excitation signal to stimulate a coupling response at the edge interface of the target tooth restoration; S2. Based on the composite dynamic excitation signal, the first response signal and the second response signal generated by the coupling response are acquired simultaneously to generate a dual-mode response signal set; S3. Based on the dual-mode response signal set, construct a spatiotemporal correlation tensor field containing the first response signal, the second response signal, and intermodal correlation information representing the correlation information between the two; S4. Input the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization to perform fusion inversion calculation, and decouple the mechanical state parameters that characterize the structural integrity of the edge interface and the heat transfer state parameters that characterize the material transport activity. S5. Based on the combined relationship between mechanical state parameters and heat transfer state parameters, generate detection results characterizing the leakage state of the edge interface.

2. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 1, characterized in that, The acquisition of the composite dynamic excitation signal used to induce a coupling response at the marginal interface of the target tooth restoration includes: Generate dynamic mechanical excitation signals to simulate alternating loads; Generate a transient thermal excitation signal that is synchronously triggered at a specific phase point of the dynamic mechanical excitation signal; The dynamic mechanical excitation signal and the transient thermal excitation signal are coupled in a programmed manner to generate a composite dynamic excitation signal.

3. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 1, characterized in that, The synchronous acquisition of the first and second response signals generated by the coupling response generates a dual-mode response signal set, including: The periodic events of the dynamic mechanical excitation component in the composite dynamic excitation signal are used as the main triggering reference; Based on the main trigger reference, the timing sequence for acquiring the first response signal and the second response signal is started synchronously; Based on the shape of the first acquired response signal, the timing of subsequent acquisitions is adaptively fine-tuned to generate an aligned dual-mode response signal set.

4. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 1, characterized in that, The construction of the spatiotemporal correlation tensor field, which includes a first response signal, a second response signal, and intermodal correlation information representing the correlation between the two, includes: Extract the structural mechanical response data sequence corresponding to the first response signal and the heat transfer response data sequence corresponding to the second response signal from the dual-mode response signal set. Calculate the real-time correlation strength between the structural mechanical response data sequence and the thermal transfer response data sequence at various points in time and space, and generate a modal correlation strength data sequence. The structural mechanics response data sequence, thermal transfer response data sequence, and modal correlation strength data sequence are integrated and their dimensions are reconstructed to generate a spatiotemporal correlation tensor field.

5. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 4, characterized in that, The calculation of the real-time correlation strength between the structural mechanical response data sequence and the thermal transfer response data sequence at various points in time and space generates a modal correlation strength data sequence, including: At the corresponding spatiotemporal points of the structural mechanics response data sequence and the heat transfer response data sequence, short time segments are extracted using a sliding time window; Calculate the mutual information or covariance coefficient between the structural mechanical response data and the thermal transfer response data within each short time segment to obtain the quantized correlation value; The quantized correlation values ​​are used as the real-time correlation strength at spatiotemporal points to form a modal correlation strength data sequence.

6. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 5, characterized in that, The process of inputting the spatiotemporal correlated tensor field into a neural network model for collaborative optimization and performing fusion inversion calculations decouples the mechanical state parameters characterizing the structural integrity of the edge interface and the heat transfer state parameters characterizing the material transport activity, including: Through the mechanical parameter prediction channel of the neural network model, based on the spatiotemporal correlation tensor field, a preliminary prediction of the mechanical state parameter map is made. Through the heat transfer parameter prediction channel of the neural network model, and based on the spatiotemporal correlation tensor field, a preliminary prediction of the heat transfer state parameter map is made. By utilizing the correlation consistency constraint function based on the intermodal correlation information in the spatiotemporal correlation tensor field, the mechanical state parameter diagram and the heat transfer state parameter diagram are jointly optimized, and the optimized mechanical state parameters and heat transfer state parameters are output.

7. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 6, characterized in that, The process of constructing the correlation consistency constraint function based on intermodal correlation information in the spatiotemporal correlation tensor field includes: First spatial distribution diagram of abnormal regions in the computational mechanical state parameter diagram; The second spatial distribution diagram of the abnormal region in the calculation of heat transfer state parameters; The overlap between the first spatial distribution map and the second spatial distribution map is calculated to obtain an overlap map. Based on the intermodal correlation information in the spatiotemporal correlation tensor field, differentiated weights are assigned to different spatial locations in the overlap graph to construct an objective function for joint optimization.

8. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 1, characterized in that, The generation of detection results characterizing the leakage state at the edge interface based on the combined relationship between mechanical state parameters and heat transfer state parameters includes: The mechanical state parameters are compared with thresholds characterizing structural integrity to generate a mechanical state anomaly map; The heat transfer state parameters are compared with thresholds characterizing mass transport activity to generate a heat transfer state anomaly map. Spatial logic analysis is performed on the mechanical state anomaly map and the heat transfer state anomaly map to identify regions with different risk levels, and graded detection results are generated based on these regions.

9. The method for detecting marginal microleakage of dental restorations under multimodal imaging according to claim 1, characterized in that, The step of generating detection results characterizing the leakage state at the edge interface based on the combined relationship between mechanical state parameters and heat transfer state parameters also includes: Acquire static image data of the marginal interface of the target tooth restoration; The detection results are spatially registered with the static image data to generate a registered fusion dataset. Based on the fused dataset, mechanical state parameters, heat transfer state parameters, and leakage risk levels determined by the combination of the two are superimposed and labeled on static image data using different visual elements.

10. A multimodal imaging-based system for detecting microleakage at the margins of dental restorations, characterized in that, include: The composite dynamic excitation signal acquisition module acquires a composite dynamic excitation signal used to excite the coupling response at the edge interface of the target tooth restoration. The dual-mode response signal set generation module, based on the composite dynamic excitation signal, synchronously acquires the first response signal and the second response signal generated by the coupled response to generate a dual-mode response signal set; The spatiotemporal correlation tensor field construction module constructs a spatiotemporal correlation tensor field based on a dual-mode response signal set, which includes a first response signal, a second response signal, and intermodal correlation information representing the correlation between the two. The neural network model decoupling module inputs the spatiotemporal correlation tensor field into the neural network model used for collaborative optimization for fusion and inversion calculation, decoupling the mechanical state parameters characterizing the structural integrity of the edge interface and the heat transfer state parameters characterizing the material transport activity. The edge leakage detection result generation module generates detection results characterizing the leakage state of the edge interface based on the combination relationship between mechanical state parameters and heat transfer state parameters.

Images