CN-122021942-A - PINN-based temporomandibular joint acoustic pathology causal deducing method

CN122021942ACN 122021942 ACN122021942 ACN 122021942ACN-122021942-A

Abstract

The invention discloses a temporomandibular joint acoustic pathology causal deducing method based on PINN, which relates to the technical field of medical signal processing, computational biomechanics and artificial intelligence intersection and comprises the following steps of S1000, signal acquisition, S2000, input variable calculation, S3000, model construction, temporomandibular joint acoustic pathology causal deducing model construction based on a Physical Information Neural Network (PINN), S4000, model training, S5000 and temporomandibular joint pathology deducing. According to the invention, continuous quantitative evaluation of temporomandibular joint lesion degree is realized through joint modeling of mechanical parameters, acoustic response and time evolution relation thereof, and a network solution space is constrained through a partial differential equation, so that a displacement field, a sound pressure field and time-space evolution thereof output by the model have definite physical significance, and the credibility and traceability of a prediction result in clinical application are improved.

Inventors

TIAN YE
ZHANG XINTAO
DU LIMING
FAN HAO
WANG LIN
GUO JING

Assignees

宁波口腔医院集团有限公司
上海交通大学宁波人工智能研究院

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A method for causal inference of temporomandibular joint acoustics based on PINN, comprising the steps of: s1000, acquiring signals, namely acquiring original signals of the temporomandibular joint in the functional movement process; s2000, calculating an input variable, namely constructing a multi-source pathology related variable according to the original signal, and calculating the input variable; S3000, constructing a model, namely constructing a temporomandibular joint acoustic pathology causal inference model based on a physical information neural network, and establishing a mapping relation between the input variable and the temporomandibular joint key state quantity in a continuous function approximation mode; S4000, model training, namely calculating multi-physical constraint loss, and training the temporomandibular joint acoustic pathology causal inference model until the multi-physical constraint loss is minimized; S5000, temporomandibular joint pathology inference, wherein the temporomandibular joint pathology is analyzed by using the temporomandibular joint acoustic causal inference model.
2. The PINN-based temporomandibular joint acoustics causal inference method of claim 1, wherein the raw signals include raw acoustic signals, temporomandibular joint region spatial coordinates, functional movement cycle information, and individual movement strategy information.
3. The PINN-based temporomandibular joint acoustooptic causal inference method of claim 2, wherein said S2000 comprises: S2100, constructing an acoustic feature vector, denoising, segmenting and time-frequency analyzing the original acoustic signal, and extracting the acoustic feature vector reflecting pathological features of joints; S2200, constructing a space coordinate variable, and mapping space coordinates of the temporomandibular joint region to a three-dimensional space coordinate system; S2300, constructing a time variable, namely mapping a motion process into a continuous time variable according to functional motion period information, and describing the occurrence position of a pathological acoustic event on a time axis; S2400, constructing external load and control parameters, carrying out parameterization representation on stress states of different patients under different exercise conditions according to individual exercise strategy information, constructing the external load and the control parameters, and enhancing individuation analysis capability; S2500, construction of input vector, input vector The definition is as follows: , Wherein, the , Representing the real temporomandibular joint acoustic feature vector acquired and extracted by the patient, Is a real number domain, and has the dimension of ; In the form of a spatial coordinate system, Is a spatial coordinate variable, representing a spatial coordinate vector of the temporomandibular joint region in three-dimensional space, Three-dimensional coordinates, representing spatial components in three orthogonal directions, As a function of the time variable, Is an external load and control parameter.
4. The PINN-based temporomandibular joint acoustooptic causal inference method as recited in claim 1, wherein said temporomandibular joint critical state quantity The definition is as follows: , Wherein the displacement field Describing deformation and displacement behaviors of the joint disc, the joint head and the contact area in the motion process, and judging abnormal displacement, reset delay, abrupt change of motion path and local strain concentration of the joint disc; Sound pressure field Acoustic responses induced by friction, impact and material nonlinearities are described to convert subjectively perceptible joint acoustic anomalies into resolvable physical quantities.
5. A temporomandibular joint acoustics pathology causal inference method based on PINN according to claim 3, wherein S4000 comprises: s4100, constructing multiple physical constraints, wherein the multiple physical constraints comprise motion control physical constraint loss, patient real data driving physical constraint loss and elastic mechanical physical constraint loss based on motion control characteristics, acoustic response characteristics and elastic mechanical characteristics of the temporomandibular joint in the functional motion process; S4200, calculating multi-physical constraint loss, by which the multi-physical constraint and the real acoustic data of the patient are calculated And through a weight adjustment strategy, the self-adaptive attention aiming at different pathological mechanisms is realized; S4300, training a model, namely performing minimization optimization on the multi-physical constraint loss, training the temporomandibular joint acoustics pathological cause and effect inference model, and completing training when the multi-physical constraint loss is converged to be within a convergence threshold value and a solution space of the multi-physical constraint is met.
6. The PINN-based temporomandibular joint acoustooptic causal inference method of claim 5, wherein S4100 comprises: s4110, constructing motion control physical constraint loss, and constructing motion control physical constraint loss The formula is as follows: , Wherein, the For displacement fields predicted by the temporomandibular joint acoustics causal inference model, For a velocity field predicted by the temporomandibular joint acoustics causal inference model, For the automatic derivation of the velocity derived from the temporal derivative of the displacement field predicted by the temporomandibular joint acoustics causal inference model, For the acceleration derived from the temporal derivatives of the velocity field predicted by the temporomandibular joint acoustic pathology causal inference model, To at the same time Lower quilt 、 And time of A co-determined function of the integrated control force, As a function of the time variable, For the external load and the control parameters, Representing squaring the two norms of the physical residual terms in brackets; s4120, constructing a physical constraint loss driven by the real data of the patient, and constructing the physical constraint loss driven by the real data of the patient The formula is as follows: , Wherein, the And Respectively represents a sound pressure field and a displacement field which are predicted by a temporomandibular joint acoustic pathology causal inference model, For the predicted time derivative of the sound pressure field, For inferring model parameters from temporomandibular joint acoustics pathology cause and effect The resulting acousto-structural coupling term is predicted down, As a function of the acoustic-structural coupling coefficient, Representing the divergence of the velocity field predicted by the temporomandibular joint acoustics causality inference model, Is a data driven item; s4130, constructing an elastomechanical physical constraint loss, and constructing an elastomechanical physical constraint loss The formula is as follows: , Wherein, the And (3) with Representing the predicted stress tensor and the predicted displacement field respectively, Is a physical item, and represents gravity or equivalent distributed load; representing the equivalent density.
7. The PINN-based temporomandibular joint acoustooptic causal inference method of claim 6, wherein S4200 comprises: S4210, calculating multi-physical constraint loss, wherein the multi-physical constraint loss is calculated through the multi-physical constraint and the real acoustic data of the patient The formula is as follows: , Wherein, the Controlling the weight lost to physical constraints for the motion; The weight lost to physical constraints is driven for patient real data, The weight lost by the physical constraint of the elastic mechanics; S4220, weight adjustment, according to weight adjustment strategy adjustment Adapt to different pathological focus.
8. The PINN-based temporomandibular joint acoustooptic causal inference method of claim 5, wherein S5000 comprises: S5100, carrying out statistical analysis on the loss component, wherein the loss component comprises the motion control physical constraint loss, the patient real data driving physical constraint loss and the elastomechanical physical constraint loss, and calculating the mean value of the loss component, the relative duty ratio of the loss component in the multi-physical constraint loss and the change trend slope of the loss component in the later training period; s5200, determining a dominant physical mechanism, and determining the dominant physical mechanism of the abnormal acoustic event according to a dominant physical mechanism judging rule; s5300, quantitatively evaluating the structural degeneration degree of the temporomandibular joint structure when the dominant physical mechanism is an elastic mechanical mechanism; S5400, temporomandibular joint pathology inference, which infers temporomandibular joint pathology by a temporomandibular joint acoustic causal inference model.
9. The PINN-based temporomandibular joint acoustics causal inference method of claim 8, wherein S5100 comprises: s5110, calculating the mean value of the loss component, and training the mean value of the loss component of the last N iterations by using the temporomandibular joint acoustic pathology causal inference model The formula is as follows: , Wherein, the In order to lose the component(s), In order to lose the component number, Is the first Loss component (b) Training times, wherein N is an integer and ranges from 5% to 10% of training rounds ; S5120, calculating the relative duty ratio of the loss component in the multi-physical constraint loss The formula is as follows: , Wherein, the A mean value of the loss of the multiple physical constraints; s5130, calculating the change trend slope of the loss component in the later training period I.e. loss component variation per training round, the formula is as follows: , Wherein, the Is the first The number of loss components is calculated, Representing training rounds, each Representing the state of the temporomandibular joint acoustic pathology causal inference model after parameter update once, if Represents the first The individual loss components have converged and, Converging the threshold for the loss rate of change if Indicating that there is still a trend in increasing the degree of constraint violation.
10. The PINN-based temporomandibular joint acoustooptic causal inference method of claim 9, wherein S5400 comprises: S5410, calculating sound pressure mutation frequency, namely the number of sound pressure mutation times in unit time; s5420 defining sound pressure abrupt event, predicting sound pressure field at time t Is above a time derivative threshold; s5430 determining the sound pressure mutation frequency When the sound pressure mutation frequency in unit time is increased, the sound pressure mutation frequency is obviously increased; s5440, defining the acoustic energy release intensity, wherein the formula is as follows: , Wherein, the Is the joint acoustic action space area; When the acoustic energy release in unit time is enhanced, the acoustic energy release in unit time is shown to be in an increasing trend, and the corresponding displacement field gradient increase is judged to reflect the intensity level change of the abnormal acoustic event; S5450, judging multiple physical coupling trends, and when the increase of sound pressure mutation frequency in unit time, the enhancement of acoustic energy release in unit time and the continuous increase of loss components of dominant physical mechanisms are simultaneously met, indicating that positive feedback is formed between the abnormal acoustic event and structural degeneration, and the pathological state of the temporomandibular joint is evolving to a more unstable state, otherwise, the pathological state of the temporomandibular joint is stable or improved, and temporomandibular joint pathological inference is completed.

Description

PINN-based temporomandibular joint acoustic pathology causal deducing method Technical Field The invention relates to the technical field of medical signal processing, computational biomechanics and artificial intelligence intersection, in particular to a temporomandibular joint acoustooptic causal deducing method based on PINN. Background Temporomandibular joint disorder (Temporomandibular Disorders, TMD) is a common oromandibular joint disorder mainly represented by joint ringing, fricatives, pain and limited functions, and has a complex occurrence mechanism, involving various factors such as abnormal positions of the articular discs, degeneration of soft tissues, unbalanced muscle functions, and changes in the mechanical environment of the articular surfaces. During clinical diagnosis, abnormal acoustic signals generated during joint movement are considered to reflect important external representations of intra-articular pathological states, and therefore temporomandibular joint pathology analysis based on acoustic signals is increasingly attracting attention. In the prior art, the research method for the temporomandibular joint acoustic pathology mainly comprises the following categories: one type of method relies on clinical experience or simple signal analysis means, and acquires joint motion acoustic signals through equipment such as an electronic stethoscope, an acceleration sensor and the like, and extracts time domain or frequency domain characteristics such as amplitude, frequency spectrum distribution, energy indexes and the like, so as to assist doctors in judging whether abnormal bouncing or friction sounds exist. The method is simple to realize, but highly depends on manual experience, has poor adaptability to different patients and different movement modes, and is difficult to reveal the specific pathological mechanism behind acoustic anomalies. Another type of method introduces a traditional machine learning or deep learning technology, takes acoustic features as input, directly establishes a mapping relation between acoustic signals and disease labels, and classifies or grades temporomandibular joint diseases by using models such as a support vector machine, a convolutional neural network and the like. Although the method improves the automatic recognition precision to a certain extent, the method still essentially belongs to a data-driven 'black box model', the model output lacks physical and biomechanical explanation, the cause of acoustic abnormality is difficult to answer, and the stability performance is difficult to maintain when the sample distribution changes. In addition, there have been studies attempting to model the mechanical behavior of the temporomandibular joint by medical imaging or finite element methods to analyze the disc displacement, stress distribution and tissue deformation. The method can theoretically reflect the mechanical state in the joint, but usually depends on an idealized geometric model and material parameters, so that the acoustic information in the real motion process of a patient is difficult to introduce, the modeling and calculation cost is high, and the method is not suitable for rapid and continuous clinical evaluation. In summary, the prior art has the following general defects that firstly, most methods use acoustic signals only as classification features, causal connection between acoustic anomalies and joint motion and tissue mechanics is not established, secondly, a data driving model lacks physical constraint, the interpretation of a prediction result is insufficient, and the mechanistic analysis of the disease degree is difficult to support, thirdly, the true acoustic data of a patient are difficult to fuse by a pure mechanics or image modeling method, the individual difference depicting capability is limited, fourthly, the traditional method only can judge whether the acoustic anomalies exist, and the severity of the disease and the development of the disease are difficult to analyze. Accordingly, those skilled in the art have focused their efforts on developing a causal inference method based on PINN temporomandibular joint acoustics. Disclosure of Invention In view of the above-mentioned drawbacks of the prior art, the present invention is directed to resolving the acoustic abnormalities of the temporomandibular joint. The invention adopts a continuous function approximation mode based on a Physical Information Neural Network (PINN) to carry out joint modeling on a mechanical state variable and an acoustic response variable of the temporomandibular joint in the functional motion process. The pathological behavior reasoning of the temporomandibular joint is realized by uniformly mapping the acoustic feature vector, the space coordinate parameter, the time variable, the external load and the control parameter to the displacement field and the sound pressure field of the temporomandibular joint. In one embodiment of the invention, a method fo