A method for predicting the efficacy of OSA based on physical neural network processing image data

Abstract

The present invention discloses a method for predicting the efficacy of OSA based on physical neural network processing of image data, relates to the technical field of OSA efficacy prediction, constructs a PINN model and the input layer receives key feature points. The present invention has the advantages of establishing a physical information neural network model PINN, and the method can comprehensively consider factors such as individual differences of patients, preoperative upper airway morphology and size, and accurately predict the changes in upper airway morphology and size after surgery, which greatly improves the accuracy and reliability of the prediction, and the PINN model can more accurately reflect the actual situation of the patient after surgery by introducing physical constraints and neural network calculations, which reduces the uncertainty of the prediction results and provides strong support for doctors to formulate personalized treatment plans. Based on accurate prediction results, doctors can provide patients with personalized treatment plans, which can not only improve the treatment effect, but also reduce the pain and surgical risks of patients.

Description

Technical Field

[0001] The present invention relates to the technical field of OSA therapeutic effect prediction, and in particular to an OSA therapeutic effect prediction method based on physical neural network processing of image data. Background Art

[0002] Obstructive sleep apnea (OSA) is a sleep-related breathing disorder characterized by repeated partial or complete obstruction of the upper airway during sleep, resulting in decreased blood oxygen saturation and fragmented sleep. Patients often experience symptoms such as snoring, daytime sleepiness, fatigue, and persistent morning headaches, which increase the incidence of motor vehicle accidents. OSA can also lead to a series of health problems. Untreated OSA increases the risk of hypertension, coronary heart disease, heart failure, myocardial infarction, stroke, arrhythmias, and insulin resistance. OSA seriously threatens overall health and is an important public health issue facing the world. A large number of OSA patients are in urgent need of effective diagnosis and treatment.

[0003] Mandibular advancement surgery can increase the upper airway volume and the size of various oropharyngeal cross-sections by moving the mandibular block forward. It can significantly improve OSA symptoms in patients with skeletal Class II mandibular retrusion accompanied by OSA. However, the cause of OSA is complex and its diagnosis and treatment involves multiple disciplines. For some patients, they often seek medical treatment for aesthetic appeals on the basis of solving sleep and breathing problems. This requires that the evaluation of surgical treatment effects needs to take into account both morphology and function. Predicting the changes in the patient's airway morphology and size due to the movement of the mandibular block can help doctors provide patients with the best treatment options that are beneficial to their health throughout their life cycle when formulating a treatment plan.

[0004] However, the current prediction methods may not fully consider the key factors affecting the upper airway morphology and size when evaluating the postoperative upper airway characteristics of OSA with bony class II. For example, individual differences among patients may lead to deviations in the evaluation results. In addition, the existing prediction methods may lack personalized prediction models, and thus cannot predict the efficacy of OSA based on the actual situation of the patient, resulting in large errors in the prediction results. In addition, the upper airway structure of OSA patients is complex and changeable. Traditional prediction methods may not be able to accurately reflect the actual postoperative situation of the patient, which leads to greater uncertainty for doctors when formulating treatment plans, increasing the risk of surgery and the pain of patients. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the efficacy of OSA based on physical neural network processing of image data, which solves the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for predicting the efficacy of OSA based on physical neural network processing of image data, comprising the following prediction method:

[0007] Step Ⅰ: collect image data of OSA patients and extract key feature points including velopharyngeal length (L-RPA) and mandibular body length (Go-Me);

[0008] Step II, construct a PINN model, which includes an input layer, a hidden layer and an output layer. The input layer receives the key feature points of velopharyngeal length (L-RPA) and mandibular body length (Go-Me), and the output layer outputs the transverse diameter LAT of the minimum cross-sectional area of ​​the upper airway MAA, the anteroposterior diameter AP of the minimum cross-sectional area of ​​the upper airway MAA and the relative changes of the joint angle (∠S-Ar-Go) and intermediate variables;

[0009] Step III, training the PINN model, optimizing the PINN model through data loss terms and physical loss terms, the data loss term enables the PINN model to learn the changing rules of the upper airway parameters according to historical data, and the physical loss term enables the prediction results of the PINN model to conform to the physical principles of the upper airway;

[0010] Step IV: Based on the prediction results of the PINN model, evaluate the efficacy of OSA with skeletal Class II surgery and output a prediction report. Optionally, the construction of the PINN model includes establishing a strain-stress relationship, calculating the first principal stress of stress, and changing the plane area; the process of establishing the strain-stress relationship is as follows: applying the stress-strain relationship and introducing the intermediate variable f j (j=1…6) respectively correspond to The quantities related to normal strain and shear strain are calculated by using the changing quantities to regard the intermediate variables as partial differentials along the x-axis, y-axis, z-axis and different planes. The expressions are as follows;

[0011] ; ; ; ;

[0012] ; ; Where: When substituting into the formula for calculation, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is ;f j (j=1…6) is the expression of the three-dimensional stress-strain relationship, which converts the stress components that are difficult to process directly into a form that can be learned by the neural network, representing the adjustment of the relationship between strain and stress in different directions and planes; q i(i=1,2,3,4,5,6) represent plane stress respectively and shear stress E is the elastic modulus, v is Poisson's ratio. Optionally, the process of calculating the first principal stress from stress is as follows: based on the stress invariance criterion, the first principal stress ζ1 is calculated from stress; the stress state can be represented by the stress tensor σ, and the expression is as follows: ; Based on the stress invariance criterion, the principal stress ζ1 is the eigenvalue of the stress tensor, and the characteristic equation of the stress tensor is det(σ-ζI)=0, where I is the unit matrix, ζ is the principal stress, and the characteristic equation of the stress tensor is about the plane stress and shear stress The function of these six variables results in three principal stresses ζ1, ζ2 and ζ3, which represent the first principal stress, the second principal stress and the third principal stress respectively;

[0013] The establishment of the strain-stress relationship in the construction of the PINN model is interrelated with the calculation of the first principal stress in the stress calculation. The changes in the normal strain and the shear strain will affect the distribution of the stress tensor, and thus affect the calculation of the principal stress. The stress value learned by the PINN model through the neural network will affect the calculation result of the first principal stress. The relationship between the PINN model and the first principal stress is as follows: ; At the same time, combined with the strain-stress relationship established in the PINN model, the introduced intermediate variable has a functional relationship with the stress, so:

[0014] Translates to:

[0015] .

[0016] Optionally, the process of changing the plane area is as follows:

[0017] Based on the relationship between principal stress and air pressure, the plane area changes. In a three-dimensional coordinate system, the change in the plane area is considered as the area change ΔS of a certain surface;

[0018] Based on the relationship between principal stress and air pressure, there is a relationship ΔS=I(ζ1). The change of surface area is affected by the principal stress, and the principal stress is related to the air pressure. The change of principal stress will cause the deformation of the surface, thereby causing the change of area. Because there is a certain relationship between the first principal stress and the intermediate variable, the expression is as follows:

[0019] ; The functional relationship between principal stress and air pressure is obtained:

[0020] ;in;

[0021] When substituting into the formula, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is ; This expression relates the change in surface area to and In connection, the change in the plane area of ​​the upper airway reflects the change in its geometric shape. In OSA patients, when the plane area decreases, the cross-sectional area through which airflow passes becomes smaller. According to the principles of fluid mechanics, the airflow resistance will increase. This increase in airflow resistance causes more dramatic changes in pressure in the airway during breathing, making the upper airway more prone to collapse.

[0022] Optionally, the PINN model specifically includes:

[0023] Part I: Establishing the strain-stress relationship, introducing intermediate variables corresponding to the quantities related to normal strain and shear strain, and learning the relationship between these intermediate variables and stress through neural networks;

[0024] Part II: Combined with physical constraints, the corresponding relationship between strain and stress in three dimensions, the stress invariance criterion, and the relationship between the change of plane area and principal stress are introduced into the neural network as constraints;

[0025] Part III: By minimizing the data loss term and the physical constraint loss term, the PINN model is trained to predict the morphology and size changes of the upper airway after surgery.

[0026] Optionally, the training PINN model includes construction of a loss function and a total loss function;

[0027] The construction of the loss function includes a data loss term and a physical constraint loss term;

[0028] The data loss term is expressed as: ;in: For neural networks, is the known historical data, y i =(a i , b i , c i );

[0029] When substituting into the formula, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is , the transverse diameter LAT is a, the anteroposterior diameter AP is b, and the joint angle (∠S-Ar-Go) is c;

[0030] The physical constraint loss term includes a first part constraint, a second part constraint and a third part constraint;

[0031] The first part of the constraints is:

[0032] ; The second part constraint is:

[0033] ; The third part constraint is:

[0034] ; Where: ΔS represents the change in surface area; H* represents a function, and H* is the result of receiving six f functions; I Represents the identity matrix.

[0035] Optionally, the total loss function is expressed as: ; Among them, the total loss function includes two parts: data fitting and physical constraints;

[0036] L data The difference between the output of the neural network and the actual observed data is the output loss term; L 1. L 2. L 3The physical constraints that reflect different factors are collectively referred to as physical constraint loss terms.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention proposes key points for judging the physical characteristics of the upper airway morphology and size after OSA and bony class II surgery, and the upper airway minimum cross-sectional area MAA is the physical characteristic that best reflects the upper airway after OSA surgery. The Cr point at the spine closest to MAA is used instead of the joint angle Ar of the skull base attachment, and the bony landmarks of the cervical spine are incorporated into the JARABAK pentagon, which can not only better reflect the influence of rigid structures such as the cervical spine on the deformation of the soft structure of the airway, but also uses the angle between the mandible and the neck as a new parameter to reflect the relative position of the skull and the cervical spine.

[0039] 2. The present invention proposes to use the PINN network to predict the upper airway morphology and size after OSA with bony Class II surgery. To this end, the mutual relationship between human body structures is analyzed, and then the parameters of each level of the PINN network are defined, and the loss function L is designed. The PINN network is successfully applied to the prediction of upper airway characteristics after OSA surgery. Through the data loss term, the PINN can learn the change law of the upper airway parameters in the historical data, so as to predict the new patient data. According to the known velopharyngeal length, the distance from the mandibular angle point to the submental point, and the historical data of the key parameters of the upper airway, the neural network can predict these parameter values ​​of the new patient. At the same time, the physical constraint loss term introduces the physical laws such as the relationship between strain and stress, the calculation of the first principal stress, and the relationship between the change of plane area and the respiratory effect into the learning of the neural network. This enables the PINN to not only fit the data, but also follow the physical characteristics of the upper airway, thereby improving the accuracy and reliability of the prediction. The upper airway structure and mechanical properties of different patients may be different, so personalized treatment is the key to the treatment of OSA with bony Class II patients. PINN The loss function can be adjusted according to the specific data and physical characteristics of each patient to improve the accuracy of the PINN model in predicting changes in upper airway characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a method for predicting the efficacy of OSA based on physical neural network processing of image data;

[0041] Figure 2 It is a projection measurement index Jarabak diagram of the prior art;

[0042] Figure 3 The Jarabak chart is the cephalometric index of this OSA efficacy prediction method;

[0043] Figure 4 This is the upper airway morphology variable diagram for this OSA efficacy prediction method;

[0044] Figure 5 This is a flow chart of the steps of the OSA efficacy prediction method. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] About this OSA efficacy prediction method based on physical neural network processing image data: PINN is a physical knowledge embedded neural network, and its principle is to approximate the partial differential equation solution by training to minimize the residual, where the residual integrates the partial differential equation of discrete points in the calculation domain, boundary conditions and initial conditions;

[0047] Most of the existing literature is a case-control study, comparing the differences in craniofacial and upper airway morphology between normal healthy adults and OSA patients. The differences are mainly reflected in six aspects: skull base, face height, maxilla and mandible, soft palate and tongue, hyoid bone, and upper airway. The commonly used cephalometric indicators are Jarabak measurements (SN, S-Ar, Ar-Go, Go-Me, ∠NS-Ar, S-Ar-Go, ∠Ar-Go-Me, ∠Sum), such as Figure 2 The five points are: nasion point (N), sella turcica point (S), joint point (Ar), mandibular angle point (Go) and submental point (Me);

[0048] The eight measurement indicators mean:

[0049] Anterior skull base length (SN): the distance from the sella turcica point (S) to the nasion point (N);

[0050] Posterior skull base length (S-Ar): distance from sella turcica point (S) to articulation point (Ar);

[0051] Mandibular ramus length (Ar-Go): distance from the articulation point (Ar) to the mandibular angle point (Go);

[0052] Mandibular body length (Go-Me): the distance from the mandibular angle (Go) to the submental point (Me);

[0053] Saddle angle (∠NS-Ar): the angle between the nasion point (N) and the articulation point (Ar) at the sella turcica point (S);

[0054] Joint angle (∠S-Ar-Go): the angle between the sella turcica point (S) and the mandibular angle point (Go) at the joint point (Ar);

[0055] Mandibular angle (∠Ar-Go-Me): The angle between the articulation point (Ar) at the mandibular angle point (Go) and the submental point (Me);

[0056] Total angle (∠Sum): the sum of (∠NS-Ar) + (∠S-Ar-Go) + (∠Ar-Go-Me);

[0057] The present invention is directed to Figure 3Patient case analysis, using the second cervical vertebra anterior inferior point (Cr) instead of the joint point (Ar), because the original five points are all located on the skull, regardless of the head posture, their relative positions change very little, mandibular advancement is one of the effective methods for treating OSA, the upper airway is a flexible structure when the human body is in a supine position, the cervical spine pressure on the airway is much greater than standing, the rigid structure of the bone will inevitably exert force on the airway and cause deformation, so the second cervical vertebra anterior inferior point, the spinal structure closest to the upper airway, is used as the physical feature, taking the human body posture into consideration;

[0058] The present invention proposes a scheme for predicting the upper airway morphology and size changes after surgery in patients with OSA and bony class II using the PINN model, wherein the position of the upper airway minimum cross-sectional area MAA is as follows: Figure 4 As shown, the upper airway volume can be expressed by its length in three-dimensional coordinates, which are:

[0059] Transverse diameter LAT;

[0060] anteroposterior diameter AP;

[0061] Upper airway length L;

[0062] Considering the relationship between the minimum cross-sectional area MAA of the upper airway and the plane angle, in addition to the transverse diameter LAT and the anteroposterior diameter AP, the joint angle (∠S-Ar-Go) was added, and then the minimum cross-sectional area MAA of the upper airway can be calculated through three parameters;

[0063] However, in actual operation, it is difficult to measure the transverse diameter LAT, anteroposterior diameter AP and joint angle (∠S-Ar-Go). As key factors affecting the upper airway, they are also manifested in the differences in human body structural parameters. According to Poiseuille's law, airflow resistance is directly proportional to the length of the upper airway and inversely proportional to the radius of the upper airway. When the airflow resistance increases, the pressure in the upper airway decreases. From the perspective of the upper airway collapse mechanism, within the range where the upper airway can collapse, when the pressure in the upper airway is less than the external pressure, collapse occurs. Therefore, when all other variables remain unchanged, a longer upper airway will increase the tendency of upper airway collapse. Further segmenting the upper airway revealed that the length of the upper airway in the velopharyngeal part is L has a great influence on the change of the minimum cross-sectional area of ​​the upper airway MAA. In other words, the smaller the area of ​​the mandibular closed area, the more severe the OSA symptoms, and the difference in bone structure may be the main factor leading to OSA in non-obese patients. Therefore, it is concluded that the mandibular body length (Go-Me) is one of the determinants of AHI in non-obese OSA patients. After controlling the demographic variables, that is, age, weight, height and body mass index, the present invention analyzes that the mandibular body length is a significant predictor of OSA severity; in summary, the relationship between the transverse diameter LAT, the anteroposterior diameter AP, the joint angle (∠S-Ar-Go) and other physiological parameters is not clear, but the main influencing factors are:

[0064] Velopharyngeal length (L-RPA), which is the vertical distance between the upper and lower borders of the velopharynx;

[0065] Mandibular body length (Go-Me), which is the distance from the mandibular angle (Go) to the submental point (Me);

[0066] Therefore, the velopharyngeal length (L-RPA) and the mandibular body length (Go-Me) are the key parameters affecting the transverse diameter LAT, the anteroposterior diameter AP and the joint angle (∠S-Ar-Go).

[0067] The present invention uses a neural network to analyze historical data and mine its related connections. The transverse diameter LAT, the anteroposterior diameter AP, and the joint angle (∠S-Ar-Go) are three-dimensional parameters for calculating the minimum cross-sectional area MAA of the upper airway, and the relationship between the area and the force is calculable.

[0068] Velopharyngeal length (L-RPA) and mandibular body length (Go-Me) are key parameters that affect the transverse diameter LAT, anteroposterior diameter AP and joint angle (∠S-Ar-Go). The neural network is used to analyze the historical data and design the overall architecture of PINN. Figure 1 shown.

[0069] Example 1: Please refer to Figures 1 to 5 This embodiment provides a method for predicting the efficacy of OSA based on physical neural network processing image data, including the following prediction methods:

[0070] Step Ⅰ: collect image data of OSA patients and extract key feature points including velopharyngeal length (L-RPA) and mandibular body length (Go-Me);

[0071] Step II, construct a PINN model, which includes an input layer, a hidden layer and an output layer. The input layer receives the key feature points of velopharyngeal length (L-RPA) and mandibular body length (Go-Me), and the output layer outputs the transverse diameter LAT of the minimum cross-sectional area of ​​the upper airway MAA, the anteroposterior diameter AP of the minimum cross-sectional area of ​​the upper airway MAA and the relative changes of the joint angle (∠S-Ar-Go) and intermediate variables;

[0072] Step III, training the PINN model, optimizing the PINN model through data loss terms and physical loss terms, the data loss term enables the PINN model to learn the changing rules of the upper airway parameters according to historical data, and the physical loss term enables the prediction results of the PINN model to conform to the physical principles of the upper airway;

[0073] Step IV: Based on the prediction results of the PINN model, evaluate the efficacy of surgery for OSA with skeletal Class II and output a prediction report.

[0074] The construction of PINN model includes the establishment of strain-stress relationship, stress calculation of first principal stress, and plane area change;

[0075] Specifically:

[0076] Building a Neural Network , where the input is the change value of mandibular body length (Go-Me) and velopharyngeal length (L-RPA), and the output is the relative change of transverse diameter LAT, anteroposterior diameter AP, joint angle (∠S-Ar-Go) and intermediate variables.

[0077] refer to Figure 1 PINN is an inner loop, and its network itself is equivalent to being nested in NN. Every time NN learning and training is performed, the physical constraint PINN will be called. It has no output itself, and the training results after adding physical constraints will be fed back to NN to improve the learning and prediction accuracy of NN.

[0078] The process of establishing the strain-stress relationship is as follows: Apply the stress-strain relationship and introduce the intermediate variable f j (j=1…6) respectively correspond to The quantities related to normal strain and shear strain are calculated by using the changing quantities to regard the intermediate variables as partial differentials along the x-axis, y-axis, z-axis and different planes;

[0079] ; ; ; ;

[0080] ; ; Where: When substituting the above formula into the calculation, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is .

[0081] f j (j=1…6) is the expression of the three-dimensional stress-strain relationship, which transforms the stress components that are difficult to process directly into a form that can be learned by the neural network. It represents an adjustment of the relationship between strain and stress in different directions and planes; q i (i=1,2,3,4,5,6) represent plane stress respectively and shear stress E is the elastic modulus, v Poisson's ratio, the initial value can be set to the average value of human muscle, or it can be designed as a learnable variable based on parameters such as age and BMI;

[0082] In this embodiment, the introduction of functions and parameters provides the model with greater flexibility and adaptability. The consideration of the variation reflects the dynamic changes of the upper airway under different conditions. In actual situations, the mechanical properties of the upper airway will change with changes in factors such as breathing and body position. The use of variation can better capture these dynamic changes. By analyzing the variation, we can better understand the changes of the upper airway in different treatment stages or under different physiological conditions. The size and mechanical properties of the upper airway may change before and after treatment. By monitoring these variations, the effect of treatment can be evaluated. Partial differential calculations can help determine the relationship between key parameters of the upper airway and the treatment effect. By analyzing partial differentials along different directions and planes, the degree of influence of parameters such as velopharyngeal length (L-RPA) and mandibular body length (Go-Me) on the changes in upper airway characteristics of OSA patients can be determined, providing a basis for formulating personalized treatment plans. During the training process of the neural network, the input velopharyngeal length (L-RPA) and mandibular body length (Go-Me) are linked to intermediate variables.

[0083] The neurons in the hidden layer transform the input geometric structure parameters into feature representations related to the intermediate variables by learning weights and biases. This connection can be seen as a way for the neural network to encode the input information, so that subsequent calculations can be performed based on intermediate variables with clear physical meanings.

[0084] By introducing intermediate variables, the relationship between normal strain, shear strain and plane stress and shear stress is introduced into the OSA problem, which provides more comprehensive physical constraints for PINN, helps to improve the reliability and interpretability of the model when dealing with three-dimensional upper airway problems. Considering the change amount and treating it as a partial differential calculation can better capture the dynamic changes of variables in the three-dimensional system. For the analysis of the changes of the upper airway in three-dimensional space, this processing method is more in line with the actual situation and helps to deeply understand the mechanical properties of the upper airway. The definition of intermediate variables and the relationship with stress will play an important role in the construction and optimization of the loss function, helping to determine the relationship between the key parameters of the upper airway and factors such as velopharyngeal length (L-RPA) and mandibular body length (Go-Me) in three-dimensional space. In the construction of the loss function of PINN, these formulas can be used as part of the physical constraint loss term to prompt the output of the neural network to conform to the correspondence between stress and strain, so as to better explore the relationship between velopharyngeal length (L-RPA), mandibular body length (Go-Me) and key parameters of the upper airway, and provide a more accurate basis for evaluating the prediction of changes in upper airway characteristics after surgery for OSA with skeletal class II.

[0085] The process of stress calculation of the first principal stress is as follows: Based on the stress invariance criterion, stress calculates the first principal stress ζ1; the stress state can be expressed by the stress tensor σ: ; Based on the stress invariance criterion, the principal stress ζ1 is the eigenvalue of the stress tensor, and the characteristic equation of the stress tensor is det(σ-ζI)=0, where I is the unit matrix, ζ is the principal stress, and the characteristic equation of the stress tensor is about the plane stress:

[0086] and shear stress The function of these six variables results in three principal stresses ζ1, ζ2 and ζ3, which represent the first principal stress, the second principal stress and the third principal stress respectively;

[0087] The establishment of the strain-stress relationship in the PINN model is interrelated with the calculation of the first principal stress in the stress calculation. The changes in normal strain and shear strain will affect the distribution of the stress tensor, and then affect the calculation of the principal stress. The stress value learned by the PINN model through the neural network will affect the calculation result of the first principal stress. The relationship between the PINN model and the first principal stress is as follows: ; At the same time, combined with the strain-stress relationship established in the PINN model, the introduced intermediate variable has a functional relationship with the stress, so:

[0088] Translates to:

[0089] ;in:

[0090] When substituting the above formula into the calculation, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is .

[0091] In this embodiment: The idea of ​​this transformation is that in the OSA problem, the present invention hopes to calculate the first principal stress through the known function related to the velopharyngeal length (L-RPA) and the mandibular body length (Go-Me). and shear stress It is difficult to directly measure or determine the relationship with the key parameters in the OSA problem, so the relationship is established through the introduction of intermediate variables, namely, functions related to velopharyngeal length (L-RPA) and mandibular body length (Go-Me);

[0092] By calculating the first principal stress, we can have a deeper understanding of the stress of the upper airway under different conditions. The principal stress reflects the direction and magnitude of the maximum stress at a certain point inside an object. For a complex structure like the upper airway, the calculation of the principal stress helps to reveal the stress distribution it bears during breathing. In OSA patients, the upper airway may be subjected to greater stress in some parts, resulting in airway stenosis or collapse. By analyzing the principal stress, these potential problem areas can be identified to provide a basis for further treatment. The increase in principal stress will cause deformation and collapse of the upper airway soft tissue, thereby causing apnea. By studying the relationship between principal stress and OSA symptoms, we can better understand the occurrence and development of the disease. There are differences in the upper airway structure and mechanical properties of different patients. The calculation of principal stress can provide a basis for personalized treatment. The calculation of the first principal stress can provide a reference for the formulation of personalized treatment plans based on the specific conditions of the patient. Therefore, the calculation of the first principal stress can be used as an important characteristic variable, together with other upper airway parameters, to predict the changes in upper airway morphology and size after surgery in OSA patients with bony Class II. By incorporating principal stress into the prediction model, the accuracy and reliability of the model can be improved.

[0093] The process of changing the plane area is as follows:

[0094] Based on the relationship between principal stress and air pressure, the change in plane area under atmospheric pressure is calculated. In a three-dimensional coordinate system, the change in plane area is considered as the area change ΔS of a certain surface.

[0095] Based on the relationship between principal stress and air pressure, there is a relationship ΔS=I(ζ1). The change of surface area is affected by the principal stress, and the principal stress is related to factors such as air pressure. The change of principal stress may cause deformation of the surface, thereby causing a change in area. Because there is a certain relationship between the first principal stress and the intermediate variable:

[0096] ; The functional relationship between principal stress and air pressure is obtained:

[0097] ;in:

[0098] When substituting the above formula into the calculation, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is .

[0099] This expression links the change in surface area to the velopharyngeal length (L-RPA) and the mandibular body length (Go-Me). Changes in the plane area of ​​the upper airway reflect changes in its geometry. Upper airway stenosis is a key issue in OSA patients. When the plane area decreases, the cross-sectional area through which airflow passes becomes smaller. According to the principles of fluid mechanics, the airflow resistance increases. This increase in airflow resistance can cause more dramatic changes in pressure in the airway during breathing, making the upper airway more susceptible to collapse.

[0100] See also Figures 1 to 5 ,In step 2, an improved definition of key points of the human structure is ,adopted when constructing the PINN model, and the Y-axis-S-anterior ,inferior point of the second cervical vertebra (Cr-Me) is used instead of the original ,(S-Ar-Go).

[0101] See also Figures 1 to 5 , building the PINN model specifically includes:

[0102] Part I: Establishing the strain-stress relationship, introducing intermediate variables corresponding to the quantities related to normal strain and shear strain, and learning the relationship between these intermediate variables and stress through neural networks;

[0103] Part II: Combined with physical constraints, the corresponding relationship between strain and stress in three dimensions, the stress invariance criterion, and the relationship between the change of plane area and principal stress are introduced into the neural network as constraints;

[0104] Part 3: Through data loss terms and physical constraint loss terms, the PINN model is trained to predict the morphology and size changes of the upper airway after surgery.

[0105] Example 2: Please refer to Figures 1 to 5 ,Training PINN model includes: construction of loss function and total loss function;

[0106] The construction of the loss function includes: data loss term and physical constraint loss term;

[0107] The data loss term is expressed as: ;in: For neural networks, is the known historical data, y i =(a i , b i , c i );

[0108] When substituting the above formula into the calculation, the velopharyngeal length (L-RPA) is set to , the mandibular body length (Go-Me) is , the transverse diameter LAT is a, the anterior-posterior diameter AP is b, and the joint angle (∠S-Ar-Go) is c.

[0109] This data loss term is calculated based on the difference between the output of the neural network and the actual data, with the goal of making the neural network fit the historical data as closely as possible;

[0110] The physical constraint loss term includes: first part constraint; second part constraint; third part constraint;

[0111] The first part of the constraints is:

[0112] ; The second part of the constraint is:

[0113] ; The third part of the constraints is:

[0114] ; Where: ΔS represents the change in surface area; H* represents a function, and H* is the result of receiving six f functions; I represents the identity matrix;

[0115] The total loss function is expressed as: ;

[0116] Among them: the total loss function includes two parts: data fitting and physical constraints; L data The difference between the output of the neural network and the actual observed data is the output loss term; L 1. L 2. L 3The physical constraints that reflect different factors are collectively referred to as physical constraint loss terms.

[0117] In this embodiment: L 1 This part of the constraint aims to establish the corresponding relationship between strain and stress in three dimensions, and to realize the constraint of physical laws in neural networks by introducing intermediate variables;

[0118] L 2 This part of the constraint is based on the stress invariance criterion, and the first principal stress is introduced into the neural network as a physical constraint;

[0119] L 3 The third part of the constraint considers the change of plane area as the change of curved surface area in three dimensions, and establishes its relationship with principal stress and breathing effect;

[0120] In summary, the total loss function includes two parts: data fitting and physical constraints. L data The difference between the output of the neural network and the actual observed data is the output loss term. L 1. L 2. L3 Physical constraints that reflect different factors are collectively referred to as physical constraint loss terms. The data loss term enables the model to learn based on a large amount of data, thereby adapting to the individual differences of different patients. The physical loss term ensures that the physical principles of the upper airway are followed in the evaluation of the prediction of the upper airway morphology and size after surgery, and conforms to the physical laws of stress-strain relationship, stress tensor change, and stress-induced area change. The upper airway structure of OSA patients in different patients is unique. The total loss function can adjust the model's prediction based on the specific situation of each patient, such as the unique upper airway geometry, different tissue mechanical properties, etc. It also helps the model accurately predict the effect of surgery on specific patients. By minimizing this loss function, the model can learn the relationship between the upper airway-related parameters of different patients and the upper airway morphology and size after surgery.

[0121] In summary, by constructing a neural network and loss function and using the idea of ​​the PINN algorithm, the relationship between the velopharyngeal length (L-RPA), mandibular body length (Go-Me), transverse diameter LAT, anteroposterior diameter AP and the joint point Ar at the mandibular angle can be established, and the accuracy and reliability of the model can be improved through intermediate variables and physical constraints of each part.

[0122] The outstanding essential features and remarkable progress in the present invention are:

[0123] 1. The present invention analyzes the key factors affecting OSA, improves the key points of physical characteristics proposed in the prior art for judging the upper airway characteristics of OSA with skeletal class II surgery, and uses the Y-axis-S-anteroinferior point of the second cervical vertebra (Cr-Me) instead of the original (S-Ar-Go).

[0124] 2. The present invention designs a framework of a physical information neural network model PINN for predicting the upper airway morphology and size after surgery for OSA with skeletal class II. The model's input, output, and intermediate variables are associated with human body structure parameters. Structurally, characteristic quantities reflecting the efficacy of OSA are used as model inputs, and PINN is used to predict changes in upper airway characteristics after surgery.

[0125] 3. In the first part of the PINN framework, the present invention adds a problem with physical constraints and designs an intermediate variable as a variable for partial differential calculation, which provides a new method for understanding the mechanical behavior of the upper airway in three-dimensional space. It uses a fusion of physical knowledge and data-driven methods to learn the relationship between intermediate variables and stress using neural networks, follows physical laws, and avoids overfitting or unreasonable prediction results that may occur in pure data-driven methods.

[0126] 4. In the second part of the PINN framework, the present invention solves the principal stress by establishing the characteristic equation of the stress tensor, which provides a tool for analyzing the stress condition and potential collapse mechanism of the upper airway in three-dimensional space, and applies the three-dimensional principal stress calculation to the prediction of upper airway morphology and size after surgery for OSA with skeletal class II. The influence of these learnable variables on the three-dimensional stress-strain relationship and principal stress calculation is further considered. According to the individual differences of different patients, such as age, BMI and other factors, the neural network can learn the elastic modulus and Poisson's ratio suitable for the three-dimensional upper airway problem, so that the model can be more adapted to the three-dimensional upper airway mechanical characteristics of different patients, thereby improving the accuracy and personalization of the prediction.

[0127] 5. In the third part of the PINN framework, the present invention associates the change in surface area with the three-dimensional stress state, providing a new indicator for evaluating the upper airway morphology and size after surgery for OSA with bony class II. This method of associating the change in three-dimensional surface area with the respiratory effect uses the change in surface area as a variable in the verification equation, providing an effective verification method for the accuracy and reliability of the model. By monitoring the change in surface area, problems and deficiencies in the model can be discovered in a timely manner, and the structure and parameters of the model can be further optimized. At the same time, this also provides a new model evaluation and improvement idea for other similar three-dimensional medical problems.

[0128] 6. The present invention proposes a processing method for applying relevant physical quantities to PINN, and introduces the relationship between normal strain and plane stress and between shear strain and shear stress into the processing of LAT, AP and LR. Through the strain and stress of the upper airway under different conditions, these physical relationships can be used to infer the size change of the upper airway, and then the principal stress can be linked to the size parameters of the upper airway. For the velopharyngeal length (L-RPA) and the mandibular body length (Go-Me), they can be used as one of the factors affecting the change of the surface area. The relationship between them and the change of the surface area can be learned through neural networks. At the same time, combined with the laws of physics and the calculation of principal stress, the changes in respiratory effects can be predicted more accurately, and the mechanical behavior of the upper airway and the changes in the characteristics of the upper airway after surgery can be better understood.

[0129] 7. This paper designs the loss function of the PINN model. The loss function is an important component of the PINN model and is responsible for integrating physical information into the neural network. The loss function of the present invention takes into account the relationship between stress-strain, first principal stress, principal stress and air pressure, and calculates the change of plane volume by force change, which enables the PINN model to use physical information analysis to predict the changes in upper airway morphology and size after surgery for OSA with skeletal class II.

[0130] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the efficacy of OSA based on physical neural network processing of image data, characterized in that: The following forecasting methods are included: Step Ⅰ: collect image data of OSA patients and extract key feature points including velopharyngeal length and mandibular body length; Step II, construct a PINN model, which includes an input layer, a hidden layer and an output layer. The input layer receives the key feature points of the velopharyngeal length and the mandibular body length, and the output layer outputs the transverse diameter LAT of the minimum cross-sectional area of ​​the upper airway MAA, the anteroposterior diameter AP of the minimum cross-sectional area of ​​the upper airway MAA, the relative change of the joint angle and the intermediate variables; Step III, training the PINN model, optimizing the PINN model through data loss terms and physical loss terms, the data loss term enables the PINN model to learn the changing rules of the upper airway parameters according to historical data, and the physical loss term enables the prediction results of the PINN model to conform to the physical principles of the upper airway; Step IV: Based on the prediction results of the PINN model, evaluate the efficacy of OSA with skeletal Class II surgery and output a prediction report; The construction of the PINN model includes establishing strain-stress relationship, calculating the first principal stress of stress, and changing the plane area; The process of establishing the strain-stress relationship is as follows: Apply the stress-strain relationship and introduce the intermediate variable f j (j=1…6) respectively correspond to ε x , ε y , ε z , γ xy , γ yz , γ xz The quantities related to normal strain and shear strain are calculated by using the changing quantities to regard the intermediate variables as partial differentials along the x-axis, y-axis, z-axis and different planes. The expressions are as follows; in: When substituting into the formula for calculation, the velopharyngeal length is set to l and the mandibular body length is set to m; f j (j=1…6) is the expression of the three-dimensional stress-strain relationship, which transforms the stress components that are difficult to process directly into a form that can be learned by the neural network, representing the adjustment of the relationship between strain and stress in different directions and planes; q i (i=1,2,3,4,5,6) represent the plane stress σ x , σ y , σ z and shear stress τ xy , τ xz , τ yz , E is the elastic modulus, v is the Poisson’s ratio; The process of calculating the first principal stress is as follows: Based on the stress invariance criterion, the first principal stress ζ1 is calculated; The stress state can be expressed by the stress tensor σ, which is expressed as follows: Based on the stress invariance criterion, the principal stress ζ1 is the eigenvalue of the stress tensor, and the characteristic equation of the stress tensor is det(σ-ζI)=0, where I is the unit matrix, ζ is the principal stress, and the characteristic equation of the stress tensor is about the plane stress σ x , σ y , σ z and shear stress τ xy , τ xz , τ yz The function of these six variables results in three principal stresses ζ1, ζ2 and ζ3, which represent the first principal stress, the second principal stress and the third principal stress respectively; The establishment of the strain-stress relationship in the construction of the PINN model is interrelated with the calculation of the first principal stress in the stress calculation. The changes in the normal strain and the shear strain will affect the distribution of the stress tensor, and thus affect the calculation of the principal stress. The stress value learned by the PINN model through the neural network will affect the calculation result of the first principal stress. The relationship between the PINN model and the first principal stress is as follows: ζ1=H(σ x ,s x ,s x ,t xy ,t yz ,t xz ); At the same time, combined with the strain-stress relationship established in the PINN model, the introduced intermediate variable has a functional relationship with the stress, so ζ1=H(σ x ,σ x ,σ x ,τ xy ,τ yz ,τ xz ) is transformed into: ζ1=H * (f1(l,m),f2(l,m),f3(l,m),f4(l,m),f5(l,m),f6(l,m)); The process of changing the plane area is as follows: Based on the relationship between principal stress and air pressure, the plane area changes. In a three-dimensional coordinate system, the change in the plane area is considered as the area change ΔS of a certain surface; Based on the relationship between principal stress and air pressure, there is a relationship ΔS=I(ζ1). The change of surface area is affected by the principal stress, and the principal stress is related to the air pressure. The change of principal stress will cause the deformation of the surface, thereby causing the change of area. Because there is a certain relationship between the first principal stress and the intermediate variable, the expression is as follows: ζ1=H * (f1 l,m,f2 l,m,f3 l,m),f4(l,m),f5(l,m),f6(l,m); The functional relationship between principal stress and air pressure is obtained: ΔS=I * (H(f1 l,m,f2 l,m,f3 l,m,f4 l,m),f5(l,m),f6(l,m))); in; When substituting into the formula for calculation, the velopharyngeal length is set to l and the mandibular body length is set to m; The PINN model specifically includes: Part I: Establishing the strain-stress relationship, introducing intermediate variables corresponding to the quantities related to normal strain and shear strain, and learning the relationship between these intermediate variables and stress through neural networks; Part II: Combined with physical constraints, the corresponding relationship between strain and stress in three dimensions, the stress invariance criterion, and the relationship between the change of plane area and principal stress are introduced into the neural network as constraints; Part III: By minimizing the data loss term and the physical constraint loss term, the PINN model is trained to predict the morphology and size changes of the upper airway after surgery; The training PINN model includes the construction of a loss function and a total loss function; The construction of the loss function includes a data loss term and a physical constraint loss term; The data loss term is expressed as: Where: k(l i , m i ) is a neural network, (l i , m i ,y i ) is the known historical data; y i =(a i ,b i ,c i ); When substituting into the formula for calculation, the velopharyngeal length is set as l, the mandibular body length is set as m, the transverse diameter LAT is set as a, the anterior-posterior diameter AP is set as b, and the joint angle is set as c; The physical constraint loss term includes a first part constraint, a second part constraint and a third part constraint; The first part of the constraints is: The second part of the constraints is: The third part of the constraints is: in: ΔS represents the change in area of ​​the surface; H * Represents the function, H * It is the result of receiving six f functions; I represents the identity matrix.

2. The OSA therapeutic effect prediction method based on physical neural network processing image data according to claim 1, characterized in that: The total loss function is expressed as: <h2 style=";text-align:left;direction:ltr">L=L<h2 style=";text-align:left;direction:ltr"> data <h2 style=";text-align:left;direction:ltr"> +L1+L2+L3; Among them, the total loss function includes two parts: data fitting and physical constraints; L data The difference between the output of the neural network and the actual observed data is the output loss term; L1, L2, and L3 reflect different physical constraints and are collectively referred to as physical constraint loss terms.

Images