CN111508606B - Glaucoma diagnostic system - Google Patents

Abstract

A glaucoma diagnosis system, comprising: an OCT scanner, a subject data input module, a data storage module, a data processing module and a display module, the OCT scanner measures the subject RNFL thickness, AL, DFA and DFD, and the subject data The input module inputs the subjects AGE and GENDER, and the data storage module is used to store the RNFL thickness, AL, DFA, DFD, AGE and GENDER data of the subjects and the RNFL thickness compensation results; the data processing module includes a first compensation unit and a second compensation unit unit, the first compensation unit has a built-in model for vertical compensation of RNFL thickness using AGE; the second compensation unit has a built-in model for vertical and horizontal compensation of RNFL thickness using AL, DFA, DFD and GENDER; the display module is used to view the input output data. Compared with the prior art, the ophthalmic diagnostic device of the present invention improves the accuracy of glaucoma diagnosis.

Description

A glaucoma diagnosis system

technical field

The invention belongs to the field of ophthalmic medical diagnosis equipment, and more particularly, relates to a glaucoma diagnosis system.

Background technique

Retinal Nerve Fiber Layer (RNFL) is used in the prior art to diagnose glaucoma, but certain factors may cause changes in the retinal nerve fiber layer, such as age (AGE), gender (GENDER), axial length of the eye (Axial Length, referred to as AL), optic disc-foveal angle (Disc-Foveal Angle, referred to as DFA) and optic disc-foveal distance (Disc-Foveal Distance, referred to as DFD), these factors have a significant relationship with the change of RNFL thickness, such as : AL increase leads to a decrease in RNFL thickness, DFD increase leads to a decrease in RNFL thickness in the Superior Quadrant and an increase in RNFL thickness in the Inferior Quadrant. Since these factors can vary RNFL thickness, ignoring the effect of these factors on RNFL thickness will reduce the accuracy of glaucoma diagnostic systems.

The existing glaucoma diagnosis system directly diagnoses glaucoma based on the measured RNFL thickness, and takes the average RNFL thickness measured by optical coherence tomography (OCT) of normal people in each age group, 1.96 standard deviations, as the lower limit of the normal value. , to analyze the sensitivity and specificity of OCT in measuring RNFL thickness. This method does not consider the above-mentioned factors that lead to the deformation of RNFL, which will affect the diagnostic accuracy of glaucoma.

There has never been a glaucoma diagnostic system using AGE, GENDER, AL, DFA and DFD to compensate the thickness of RNFL for assisting glaucoma diagnosis in the prior art.

SUMMARY OF THE INVENTION

In order to compensate the deformation effects of AGE, GENDER, AL, DFA and DFD and other factors on the RNFL, and to improve the accuracy of glaucoma diagnosis, a glaucoma diagnosis system is provided, which uses AGE, GENDER, AL, DFA and DFD through built-in modules. Compensating the RNFL, and the compensated RNFL curve can improve the accuracy of the glaucoma diagnosis system in diagnosing glaucoma.

The invention provides a glaucoma diagnosis system, comprising: an OCT scanner, a subject data input module, a data storage module, a data processing module and a display module, the OCT scanner measures the subject's RNFL thickness, AL, DFA and DFD, The subject data input module inputs subjects AGE and GENDER, and the data storage module is used to store subject RNFL thickness, AL, DFA, DFD, AGE and GENDER data and RNFL thickness compensation results; the data processing module includes a first compensation unit and the second compensation unit, the first compensation unit has a built-in model for vertical compensation of RNFL thickness using AGE; the second compensation unit has a built-in model for vertical and horizontal compensation of RNFL thickness using AL, DFA, DFD and GENDER; display module Used to view input and output data.

Preferably, the first compensation module has the following built-in model to compensate the thickness of the RNFL in the vertical direction:

y′ _AGE = [x1]×W _AGE ×AGE_RBF+y _train

where:

y′ _AGE represents the RNFL thickness after AGE compensation;

[x1] is an n×2 matrix composed of normalized AGE and bias;

W _AGE represents the weight matrix;

AGE_RBF represents radial basis matrix;

y _train represents the RNFL thickness of n subjects.

Preferably, the optimal solution of the weight matrix W _AGE is:

Randomly initialize the weight matrix W _AGE ;

Define the error function as the mean square error function MSE(y′ _AGE ,y _mean ) of y′ _AGE and y _mean ;

Minimize the error function MSE(y′ _AGE , y _mean ) to obtain the optimal solution of the weight matrix W _AGE .

Preferably, the second compensation unit has the following built-in model to compensate the thickness of the RNFL in the horizontal direction and the vertical direction:

y′=s×g(y,p)

where:

s represents the compensation amount of RNFL thickness in the vertical direction;

y represents the RNFL thickness before compensation using GENDER, AL, DFA and DFD factors;

p represents the compensation amount of RNFL thickness in the horizontal direction;

g(y,p) is the transformation formula, which transforms y from discrete to continuous.

Preferably, the transformation formula g(y,p) for discrete-to-continuous transformation of y is:

where:

p represents the compensation amount of RNFL thickness in the horizontal direction;

表示向下取整运算。p is the element in the vector p,

Represents a round-down operation.

Preferably, the compensation amount p of the RNFL thickness in the horizontal direction is obtained as follows:

p=i+Δp

Δp=L×tanh(Δp′)

where:

p is an N-dimensional column vector, representing the compensation amount of the RNFL thickness in the horizontal direction;

i is an N-dimensional column vector i=(1,2,...,N) ^T , indicating the initial position for obtaining the thickness of the RNFL;

Δp is an N-dimensional column vector Δp=(Δp ₁ ,Δp ₂ ,...,Δp _N ), which represents the compensation variation of the RNFL thickness in the horizontal direction;

L is a constant;

tanh() is the hyperbolic tangent function

由如下的公式计算获得，

It is calculated by the following formula,

Among them, i is the element in the N-dimensional column vector i=(1,2,...,N) ^T ;

m维列向量c表示把RNFL等分了m份，j为1至m的正整数；c _j is the element in the _m -dimensional column vector c=(c ₁ ,c ₂ ,...,cm ),

The m-dimensional column vector c represents that the RNFL is divided into m equal parts, and j is a positive integer from 1 to m;

The parameter σ controls the width of the Gaussian kernel;

w _j is the element in the m-dimensional column vector w_RBF

_{w_RBF} =z×WFC

z is a 4×n matrix, which consists of gender data and normalized AL, DFA, DFD data, and is represented by the following expression

AL _{standard_k} , DFA _{standard_k} , DFD _{standard_k} , GENDER _k represent the gender data of the k-th subject and the standardized AL, DFA, and DFD data in n normal subjects, respectively.

Preferably, the way to obtain the optimal solution of _WFC is:

1) Randomly initialize the weight matrix W _FC ;

2) Define the error function as the mean square error function MSE(y', y _mean ) of y' and y _mean . In order to prevent the parameter W _FC from overfitting during training, define the loss function Loss=MSE(y', y _mean )+ λ|W _FC | ² ;

3) Minimize the loss function Loss to obtain the optimal solution of the weight matrix W _FC and w_RBF.

Preferably, the compensation amount s of the RNFL thickness in the vertical direction is obtained as follows:

s=L×tanh(s′)

w _j are elements in the m-dimensional column vector s_w_RBF

s_w_RBF=z×s_W _FC

where:

L is a constant;

s_w_RBF represents the weight matrix of the RBF neural network in the process of calculating s;

s_W _FC represents the weight matrix of the FC neural network in the process of calculating s;

z is a 4xn matrix consisting of gender data and normalized AL, DFA, DFD data.

Preferably, the training method for obtaining s_W _FC and s_w_RBF is:

1) Randomly initialize the weight matrix s_W _FC ;

2) Define the error function as the mean square error function MSE(y', y _mean ) of y' and y _mean , define the loss function Loss=MSE(y', y _mean )+λ|s_W _FC | ² , the loss function has Regular term, the coefficient of the regular term is λ;

3) Minimize the loss function MSE by training the neural network, find the optimal solution of the weight matrix s_W _FC , and find s_w_RBF.

The beneficial effect of the present invention is that, compared with the prior art, in the validation set, compared with the results of the original RNFL for the glaucoma diagnostic system, the AUROC (Receiver Operating The area under the characteristic curve, Area Under the Receiver Operating Characteristic curve, referred to as AUROC) increased from 0.779 to 0.861. For the 10% of the samples with the longest AL in the validation set, the AUROC increased from 0.703 to 0.842; for the 20% of the samples with the longest AL in the validation set, the AUROC increased from 0.748 to 0.889; for the 30 with the longest AL in the validation set % of the samples, AUROC increased from 0.767 to 0.891.

Description of drawings

Fig. 1 is the structural block diagram of a kind of glaucoma diagnosis system of the present invention;

Fig. 2 is the built-in model schematic diagram of the data processing module of the glaucoma diagnosis system of the present invention;

3 is a schematic diagram of the RNFL thickness curve before and after compensation by the glaucoma diagnostic system of the present invention.

Detailed ways

The present application will be further described below with reference to the accompanying drawings. The following examples are only used to more clearly illustrate the technical solutions of the present invention, and cannot be used to limit the protection scope of the present application.

Bold and lowercase English letters represent vectors or matrices. A 1×n matrix is called an n-dimensional row vector, and an n×1 matrix is called an n-dimensional column vector.

As shown in FIG. 1, the present invention provides a glaucoma diagnosis system, including: an OCT (Optical Coherence Tomography, OCT for short) scanner, a subject data input module, a data storage module, and a data processing module and display module. The OCT scanner measured the RNFL thickness, AL (Axial Length, AL), DFA (Disc-Foveal Angle, DFA) and DFD (Disc-Foveal Distance, Disc-Foveal) Distance, referred to as DFD); subject data input module input subjects AGE (age), GENDER (sex); data storage module is used to store subjects RNFL thickness, AL, DFA, DFD, AGE and GENDER data, RNFL Thickness compensation result.

The data processing module includes a first compensation unit and a second compensation unit. The first compensation unit has a built-in compensation model for vertical compensation of RNFL thickness using AGE; and compensation model for horizontal compensation.

Use the AGE, GENDER, AL, DFA and DFD of n normal subjects to train a neural network, obtain a compensation model, and substitute AGE, GENDER, AL, DFA and DFD of glaucoma patients into the compensation model, the glaucoma diagnostic system will obtain glaucoma patients RNFL thickness compensation results to aid in the diagnosis of glaucoma.

As shown in FIG. 2 , the first compensation unit is used to compensate the thickness of the RNFL in the vertical direction by using the AGE. Its built-in model is:

y′ _AGE = [x1]×W _AGE ×AGE_RBF+y _train (1)

where:

y′ _AGE is an n×N matrix representing the age-compensated RNFL thickness;

[x1] is an n×2 matrix composed of normalized AGE and bias;

W _AGE is a 2×m matrix, representing the weight matrix;

AGE_RBF is an m×N matrix, representing a radial basis matrix;

y _train is an n × N matrix representing the RNFL thickness of n subjects.

The training process of this model is:

To standardize AGE, those skilled in the art can use various methods to standardize. As an example, this embodiment uses the following formula to standardize AGE:

where:

AGE _{standard_k} represents the normalized age of the kth normal subject among n normal subjects;

AGE ₁ AGE ₂ ... AGE _n respectively represent the age of each normal subject, AGE _k represents the age of the kth normal subject among n normal subjects;

AGE _mean represents the average age of n normal subjects;

AGE _std represents the standard deviation of the age of n normal subjects.

The normalized ages of n normal subjects form an n-dimensional column vector x, that is

x=(AGE _{standard_1} ,AGE _{standard_2} ,…,AGE _{standard_n} ) ^T .

Add a column offset of 1 to the n-dimensional column vector x to obtain an n × 2 matrix, that is, [x1].

N的优选值为768。使用y_train计算正常受试者RNFL厚度均值y_mean，即计算y_train每一列元素的均值，将其作为列元素形成y_mean，同一列中的元素均是该位置RNFL厚度均值，同一行中的元素是不同位置的RNFL厚度均值，即y_mean是如下的n×N矩阵，Define an N-dimensional column vector i=(1,2,...,N) ^T represents the initial position to obtain the RNFL thickness, and define the RNFL thickness of subject k as an annular N-dimensional row vector y _k =(t _k1 ,t _k2 ,... ,t _kN ),k=1,2,...,n, the RNFL thicknesses of n subjects form an n×N matrix

A preferred value for N is 768. Use y _train to calculate the mean RNFL thickness y _mean of normal subjects, that is, calculate the mean value of each column element of y _train , and use it as a column element to form y _mean . The elements are the mean values of RNFL thickness at different locations, i.e. y _mean is an n×N matrix as follows,

As shown in FIG. 3 , a schematic diagram of the RNFL thickness curve before and after compensation is shown, and the ring refers to the fact that the RNFL curve is circular. However, when doing subsequent calculations and drawing the RNFL thickness curve, the RNFL thickness curve is expanded, and after expansion is a wavy curve, then the wavy curve will be vertical and horizontal during compensation, that is There is movement in up and down and left and right directions.

Because age only changes the RNFL curve vertically, not horizontally, when compensating for RNFL with age, there is only a vertical change and no horizontal change. When compensating the RNFL thickness in the vertical direction with the following formula:

y′ _AGE = [x1]×W _AGE ×AGE_RBF+y _train

The weight matrix W _AGE is obtained by training the neural network represented by the above formula, and the preferred value of m is 10, that is, the weight matrix W _AGE is a 2×10 matrix, and the optimal solution of the weight matrix W _AGE is obtained by the following method:

1) Randomly initialize the weight matrix W _AGE ;

2) Define the error function as the mean square error function MSE(y′ _AGE , y _mean ) of y′ _AGE and y _mean ;

3) Use the Lagrange multiplier method with KKT conditions to minimize the error function MSE(y′ _AGE , y _mean ), and obtain the optimal solution of the weight matrix W _AGE .

The values of the elements of the radial basis matrix AGE_RBF are calculated by the following formula (4), which is a two-dimensional Gaussian kernel function.

where:

i is the element in the N-dimensional column vector i=(1,2,...,N) ^T ;

m维列向量c表示把RNFL等分了m份，j为1至m的正整数；m是一个超参数，在计算时可以由所属领域的技术人员进行设定，例如但不限于，设m为10、20、30等；分成m份，是为了利用第i个位置与第c_j个位置的距离信息，加高斯核就是为了利用距离信息，即高斯核里的欧式距离；参数σ控制高斯核的宽度；c _j is the element in the m-dimensional column vector c=(c ₁ ,c ₂ ,...,c _m ) ^T ,

The m-dimensional column vector c represents that the RNFL is divided into m equal parts, and j is a positive integer from 1 to m; m is a hyperparameter, which can be set by those skilled in the art during calculation, for example, but not limited to, let m It is 10, 20, 30, etc.; it is divided into m parts, in order to use the distance information between the ith position and the c _jth position, and the Gaussian kernel is added to use the distance information, that is, the Euclidean distance in the Gaussian kernel; the parameter σ controls the Gaussian the width of the nucleus;

As mentioned above, when the preferred value of m is 10 and the preferred value of N is 768, the radial basis matrix AGE_RBF is preferably a 10×768 matrix.

The optimal solution of the weight matrix W _AGE is put into formula (1), and then the compensated RNFL thickness y′ _AGE of normal subjects is obtained.

The second compensation unit is used to compensate the RNFL thickness obtained by the first compensation unit in vertical and horizontal directions by simultaneously using GENDER, AL, DFA and DFD factors on the basis of using age to compensate the RNFL thickness.

That is, on the basis of using age to compensate the RNFL thickness, continue to use the following built-in model to compensate the RNFL thickness in both the horizontal and vertical directions:

y′=s×g(y,p) (5)

where:

s represents the compensation amount of RNFL thickness in the vertical direction;

y represents the RNFL thickness before compensation using GENDER, AL, DFA and DFD factors;

p is an N-dimensional column vector, representing the compensation amount of the RNFL thickness in the horizontal direction;

g(y,p) is the transformation formula, which transforms y from discrete to continuous.

Since the actual RNFL thickness is continuous and the matrix y is discrete, an approximate discrete-to-continuous transformation of y is performed using the following formula:

where:

p is an N-dimensional column vector, representing the compensation amount of the RNFL thickness in the horizontal direction;

表示向下取整运算，结果是不大于p的最大整数，

mod N表示求余运算，例如p＝1000.3，N＝768，

p is an element in the vector p ie p=(p ₁ ,p ₂ ,...,p _N ) ^T ,

Represents a round-down operation, and the result is the largest integer not greater than p,

mod N represents the remainder operation, for example, p=1000.3, N=768,

y represents the RNFL thickness before compensation using GENDER, AL, DFA and DFD factors;

p is calculated by the following formula (7),

p=i+Δp (7)

where:

p is an N-dimensional column vector, representing the compensation amount of the RNFL thickness in the horizontal direction;

i is an N-dimensional column vector i=(1,2,...,N) ^T , indicating the initial position for obtaining the thickness of the RNFL;

Δp is an _N -dimensional column vector Δp=(Δp ₁ , Δp ₂ , .

Δp is calculated by the following formula (8),

Δp=L×tanh(Δp′) (8)

where:

L is a constant, and the value of L is preferably but not limited to L=200;

tanh() is the hyperbolic tangent function, that is, the following formula (9),

Δp' is an N-dimensional column vector Δp' = (Δp' ₁ , Δp' ₂ , . . . , Δp' _N ) ^T .

The element i in the initial position vector i is used as the input of the RBF neural network, and its output is the horizontal variable of the ith position of the vector y, that is, Δp' _i is calculated by the following formula (10),

由如下的公式(11)计算获得，

It is calculated by the following formula (11),

exp() represents the exponential function with the natural constant e as the base;

i is an N-dimensional column vector i=(1,2,...,N) ^T , indicating the initial position for obtaining the thickness of the RNFL;

m维列向量c表示把RNFL等分了m份，j为1至m的正整数，m的取值优选但不限于6；c _j is the element in the _m -dimensional column vector c=(c ₁ ,c ₂ ,...,cm ),

The m-dimensional column vector c represents that the RNFL is divided into m equal parts, j is a positive integer from 1 to m, and the value of m is preferably but not limited to 6;

The parameter σ controls the width of the Gaussian kernel;

w _j is the element in the m-dimensional column vector w_RBF

w_RBF=z×W _FC (12)

where:

w_RBF represents the weight matrix of the RBF neural network;

W _FC represents the weight matrix of the FC neural network;

z is a 4xn matrix consisting of gender data and normalized AL, DFA, DFD data.

The training method to obtain W _FC and w_RBF is:

1) Randomly initialize the weight matrix W _FC ;

2) Define the error function as the mean square error function MSE(y', y _mean ) of y' and y _mean . In order to prevent the parameter W _FC from overfitting during training, define the loss function Loss=MSE(y', y _mean )+ λ|W _FC | ² , the loss function has a regular term, the regular term coefficient is λ, and the value of λ is preferably but not limited to 0.003;

3) Train the neural network to minimize the loss function, find the optimal solution of the weight matrix W _FC , and find w_RBF.

Among them, z is a 4×n matrix, which consists of gender data and normalized AL, DFA, and DFD data, and is represented by the following expression (13)

where:

AL _{standard_k} , DFA _{standard_k} , DFD _{standard_k} , GENDER _k represent the gender data of the kth subject and the standardized AL, DFA, and DFD data in n normal subjects, respectively;

AL _{standard_k} , DFA _{standard_k} , DFD _{standard_k} are calculated by the following formula (13),

AL represents the subject's axial length.

AL ₁ AL ₂ ... AL _n represent the axial length of each normal subject, respectively.

AL _mean represents the mean axial length of n normal subjects.

AL _std represents the standard deviation of the axial length of n normal subjects.

DFA represents the subject's optic disc-foveal angle.

DFA ₁ DFA ₂ ... DFA _n represent the optic disc-foveal angle of each normal subject, respectively.

DFA _mean represents the mean optic disc-foveal angle of n normal subjects.

DFA _std represents the standard deviation of the optic disc-foveal angle of n normal subjects.

DFD represents the subject optic disc-foveal distance.

DFD ₁ DFD ₂ ... DFD _n represents the disc-foveal distance of each normal subject, respectively.

DFD _mean represents the mean disc-foveal distance of n normal subjects.

DFD _std represents the standard deviation of the optic disc-foveal distance in n normal subjects.

Represent the gender GENDER of n normal subjects as 0 or 1 without normalization.

The calculation process of the compensation amount s representing the thickness of the RNFL in the vertical direction is as follows:

s=L×tanh(s′)

w _j are elements in the m-dimensional column vector s_w_RBF

s_w_RBF=z×s_W _FC

where:

L is a constant, preferably, but not limited to, L=0.9.

s_w_RBF represents the weight matrix of the RBF neural network in the process of calculating s;

s_W _FC represents the weight matrix of the FC neural network in the process of calculating s;

z is a 4xn matrix consisting of gender data and normalized AL, DFA, DFD data.

The training method to obtain s_W _FC and s_w_RBF is:

1) Randomly initialize the weight matrix s_W _FC ;

2) Define the error function as the mean square error function MSE(y', y _mean ) of y' and y _mean . In order to prevent the parameter s_W _FC from overfitting during training, define the loss function Loss=MSE(y', y _mean )+ λ|s_W _FC | ² , the loss function has a regular term, the regular term coefficient is λ, and the value of λ is preferably but not limited to 0.003;

3) Minimize the loss function by training the neural network, find the optimal solution of the weight matrix s_W _FC , and find s_w_RBF.

The display module is used to display the subject's original RNFL thickness, AGE, GENDER, AL, DFA and DFD, as well as the compensated RNFL thickness.

The beneficial effect of the present invention is that, compared with the prior art, in the validation set, compared with the results of the original RNFL for the glaucoma diagnostic system, the AUROC (Receiver Operating The area under the characteristic curve, Area Under the Receiver Operating Characteristic curve, referred to as AUROC) increased from 0.779 to 0.861. For the 10% of the samples with the longest AL in the validation set, the AUROC increased from 0.703 to 0.842; for the 20% of the samples with the longest AL in the validation set, the AUROC increased from 0.748 to 0.889; for the 30 with the longest AL in the validation set % of the samples, AUROC increased from 0.767 to 0.891.

The applicant of the present invention has described and described the embodiments of the present invention in detail with reference to the accompanying drawings, but those skilled in the art should understand that the above embodiments are only preferred embodiments of the present invention, and the detailed description is only to help readers better Rather, any improvement or modification based on the spirit of the present invention should fall within the protection scope of the present invention.

Claims (6)

1. A glaucoma diagnostic system comprising: an OCT scanner, a data input module of a subject, a data storage module, a data processing module and a display module, which is characterized in that,

the OCT scanner measures the RNFL thickness, AL, DFA and DFD of the testees, the data input module of the testees inputs AGE and GENDER, the data storage module is used for storing the data of the RNFL thickness, AL, DFA, DFD, AGE and GENDER of the testees and the RNFL thickness compensation result;

training a neural network by using AGE, GENDER, AL, DFA and DFD of n normal subjects to obtain a compensation model, substituting AGE, GENDER, AL, DFA and DFD of glaucoma patients into the compensation model, and obtaining an RNFL thickness compensation result of the glaucoma patients by the glaucoma diagnosis system for assisting in diagnosing glaucoma; the compensation model comprises: a model for vertical direction compensation of RNFL thickness using AGE and a model for vertical and horizontal direction compensation of RNFL thickness using AL, DFA, DFD and GENDER;

the data processing module comprises a first compensation unit and a second compensation unit, wherein a model for compensating the thickness of the RNFL in the vertical direction by using AGE is arranged in the first compensation unit; a model for compensating the thickness of the RNFL in the vertical direction and the horizontal direction by using AL, DFA, DFD and GENDER is built in the second compensation unit; the first compensation unit is internally provided with the following model for compensating the thickness of the RNFL in the vertical direction:

y′ _AGE ＝[x1]×W _AGE ×AGE_RBF+y _train

in the formula:

y′ _AGE represents the RNFL thickness after AGE compensation;

[ x1] is an n × 2 matrix of normalized AGEs and offsets;

W _AGE representing a weight matrix;

AGE _ RBF denotes a radial basis matrix;

y _train represents RNFL thickness for n subjects;

weight matrix W _AGE The optimal solution of (c) is:

random initialization weight matrix W _AGE ；

Defining an error function as y' _AGE And y _mean Mean square error function MSE (y' _AGE ,y _mean )；

Minimizing an error function MSE (y' _AGE ,y _mean ) Obtaining a weight matrix W _AGE The optimal solution of (2);

the second compensation unit is internally provided with the following model for compensating the thickness of the RNFL in the horizontal direction and the vertical direction:

y′＝s×g(y,p)

in the formula:

s represents the compensation amount of the RNFL thickness in the vertical direction;

y represents the RNFL thickness before compensation using GENDER, AL, DFA and DFD factors;

p represents the compensation amount of the RNFL thickness in the horizontal direction;

g (y, p) is a transformation formula, and discrete-to-continuous transformation is performed on y;

the display module is used for displaying the original RNFL thickness, AGE, GENDER, AL, DFA and DFD of the subject and the compensated RNFL thickness;

OCT refers to optical coherence tomography, RNFL refers to retinal nerve fiber layer, AL axial length, DFA refers to disc-foveal angle, DFD refers to disc-foveal distance, AGE refers to AGE, GENDER refers to GENDER.

2. The glaucoma diagnostic system according to claim 1, wherein:

the transformation formula g (y, p) for discrete to continuous transformation of y is:

in the formula:

p represents the compensation amount of the RNFL thickness in the horizontal direction;

p is an element in the vector p,

indicating a rounding down operation.

3. The glaucoma diagnostic system according to claim 2, wherein:

the compensation amount p of the RNFL thickness in the horizontal direction is obtained as follows:

p＝i+Δp

Δp＝L×tanh(Δp′)

in the formula:

p is an N-dimensional column vector representing the amount of compensation of the RNFL thickness in the horizontal direction;

i is an N-dimensional column vector i ═ (1,2, …, N) ^T Indicating an initial position for obtaining RNFL thickness;

Δ p is an N-dimensional column vector Δ p ═ Δ p (Δ p) ₁ ,Δp ₂ ,…,Δp _N ) Indicating the amount of variation in the RNFL thickness compensation in the horizontal direction;

l is a constant;

tanh () is a hyperbolic tangent function;

is calculated by the following formula,

where i is an N-dimensional column vector i ═ (1,2, …, N) ^T The element (1) in (1);

c _j is m-dimensional column vector c ═ c ₁ ,c ₂ ,…,c _m ) The elements (A) and (B) in (B),

an m-dimensional column vector c represents that RNFL is equally divided into m parts, and j is a positive integer from 1 to m;

the parameter σ controls the width of the gaussian kernel;

w _j are elements in an m-dimensional column vector w _ RBF,

w_RBF＝z×W _FC

w _ RBF denotes the weight matrix of the RBF neural network, W _FC A weight matrix representing the FC neural network;

z is a 4 × n matrix composed of gender data and normalized AL, DFA, DFD data, and expressed by the following expression

AL _{standard_k} ,DFA _{standard_k} ,DFD _{standard_k} ,GENDER _k Represents the kth subject gender data and the normalized AL, DFA, DFD data, respectively, of n normal subjects.

4. The glaucoma diagnostic system according to claim 3, wherein:

obtaining W _FC The optimal solution mode is as follows:

1) random initialization weight matrix W _FC ；

2) Defining error functions as y' and y _mean Mean square error function MSE (y', y) _mean ) To prevent the weight matrix W from being used in training _FC Overfitting, defining Loss function Loss ═ MSE (y', y) _mean )+λ|W _FC | ² ；

3) Minimizing Loss function Loss to obtain weight matrix W _FC And an m-dimensional column vector w _ RBF.

5. The glaucoma diagnostic system according to claim 4, wherein:

the amount s of RNFL thickness compensation in the vertical direction is obtained as follows:

s＝L×tanh(s′)

w _j are elements in the m-dimensional column vector s _ w _ RBF,

s_w_RBF＝z×s_W _FC

in the formula:

l is a constant;

s _ w _ RBF represents a weight matrix of an RBF neural network in the s calculation process;

s_W _FC representing a weight matrix of the FC neural network in the process of calculating s;

z is a 4 × n matrix, composed of gender data and normalized AL, DFA, DFD data.

6. The glaucoma diagnostic system according to claim 5, wherein:

obtain s _ W _FC The training method of s _ w _ RBF is as follows:

1) random initialization weight matrix s _ W _FC ；

2) Defining error functions as y' and y _mean Mean square error function MSE (y', y) _mean ) Defining a Loss function Loss as MSE (y', y) _mean )+λ|s_W _FC | ² The loss function is provided with a regular term, and the coefficient of the regular term is lambda;

3) the weight matrix s _ W is obtained by minimizing the loss function MSE through a method for training a neural network _FC Is most preferredThe solution is obtained as s _ w _ RBF.

Images