A method for constructing a parameter prediction model

Abstract

The present invention discloses a method for constructing a parameter prediction model, comprising the following steps: Step 1: Select the input values and expected output values in the model training stage; Step 2: Construct a BP neural network model; Step 3: Train the model to obtain a BP neural network model applicable to the prediction of the cutting thickness of SMILE surgery. Advantages: Deep learning is performed on a large amount of existing data based on the BP neural network, and the prediction function can be realized independently of the SMILE surgery machine. Therefore, the present invention can fundamentally realize the scientific and rapid screening of the indications for femtosecond laser in situ keratomileusis while ensuring the prediction accuracy, and can save clinical time.

Description

Technical Field

[0001] The present invention relates to a method for constructing a parameter prediction model. Background Art

[0002] At present, the application of femtosecond laser in refractive surgery of the cornea at a near visual angle is becoming increasingly widespread. In particular, small incision lenticule extraction (SMILE) with femtosecond laser shows good safety, effectiveness, stability, and predictability. However, the postoperative clinical effects of different patients are still different. Therefore, how to achieve true individualization still needs to be continuously explored, and there are still many unknown factors in the surgical parameter design. In recent years, refractive surgeons have continuously explored and optimized the design of SMILE surgical parameters on the premise of following the basic theory of visual optics to ensure the accuracy of the surgery and improve the postoperative visual quality of patients.

[0003] As a relatively new corneal refractive surgery method, SMILE is currently in the development stage of continuous exploration. The most important problem is the lack of sufficient experience in setting surgical parameters. In order to reduce the damage to corneal tissue, achieve individualized surgical design, and achieve the best surgical effect, it is necessary to carefully analyze the setting of surgical parameters and related surgical designs, and give optimization and necessary adjustments.

[0004] The BP neural network (Back-Propagation Neural Network), also known as the backpropagation neural network, is a multi-layer feedforward neural network through the error backpropagation algorithm. This network has good non-linear function approximation ability and can achieve accurate prediction.

[0005] The key parameters of the SMILE surgery are mainly the cutting thickness and the remaining basal thickness. Among them, the cutting thickness is affected by data such as the spherical power (SPH), cylindrical power (CYL), corneal curvature radius (K), and lenticule diameter (Diameter) of the patient, while the value of the remaining basal thickness is affected by parameters such as corneal thickness, cutting thickness, basal thickness, and cap thickness.

[0006] Due to the key parameters of the SMILE surgery being affected by the above-mentioned multiple factors, it is difficult to directly obtain a judgment result clinically, including the cutting thickness, the remaining basal thickness, and whether the patient meets the surgical conditions.

[0007] Usually, to obtain the surgical parameter conclusion, it is necessary to rely on the SMILE surgery machine and input various parameters to obtain results such as the cutting thickness and the remaining basal thickness. However, this method will occupy the time of the SMILE surgery machine and cannot quickly perform calculations and judgments during the preoperative screening process.

[0008] Therefore, if a scientific SIMLE surgical parameter prediction method can be established based on some data provided by the manufacturer and long-term accumulated patient data as prior conditions, quickly predict the cutting thickness and remaining basal thickness, help clinicians judge whether the patient meets the surgical conditions, and at the same time avoid occupying the time of the SMILE surgery machine, better economic and social benefits can be obtained. Summary of the Invention

[0009] In view of the above deficiencies, the present invention provides a method for constructing a parameter prediction model, which is based on the machine learning mode of BP neural network, simulates and predicts the corneal cutting formula of SMILE surgery, calculates the corneal thickness to be cut in SMILE surgery by inputting the current surgical data of the patient, predicts the feasibility of the patient's surgery, assists in evaluating the surgical risk, and guides the surgeon to select a personalized surgical plan.

[0010] The technical solution adopted by the present invention is as follows:

[0011] A method for constructing a parameter prediction model includes the following steps:

[0012] Step 1: According to the official data provided by the manufacturer and the historical patient data, select four indicators of the patient's spherical lens degree (SPH), cylindrical lens degree (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) as the input values in the model training stage, and select the corresponding cutting thickness value Y of the input quantity as the expected output value in the model training stage;

[0013] Step 2: According to the input quantity and expected output quantity in Step 1, construct a BP neural network model. The BP neural network model includes a three-layer feedforward neural network structure, namely an input layer, a hidden layer, and an output layer. The input indicators of the input layer are the input quantities selected in Step 1, and the output indicators of the output layer are the expected output quantities; according to the actual prediction accuracy requirements, set the expected error W;

[0014] Among them, the activation functions of the hidden layer and the output layer are both selected as tansig, the network training function is trainglm, and the network performance function is mse;

[0015] The BP neural network model selects the S-shaped transfer function logsig, and its expression is Through the backpropagation error function Continuously adjust the network weights and thresholds to make the error function E less than the expected error W; where, T i is the expected output, Q i is the network calculated output;

[0016] The number of neurons in the BP neural network model, the number of neurons L in the hidden layer is determined by referring to the following formula:

[0017]

[0018] c is the number of nodes in the input layer, b is the number of nodes in the output layer, and a is a constant between [1, 10];

[0019] Step 3: Train the BP neural network model in Step 2; the specific training steps are as follows:

[0020] Step 31: Generate an input vector P1 from the data of the 4 selected indicators in Step 1; generate an output vector T1 from the data of the selected expected output values in Step 1, and this output vector T1 is used as the expected output vector;

[0021] P1 = [S1, S2, S3,..., S n ; C1, C2, C3,..., C n ; K1 ， K2, K3,..., K n ; D1, D2, D3,..., D n ; T1 = [Y1, Y2, Y3,..., Y n ;

[0022] Among them, n is the number of historical sample data. The input vector P consists of the spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient. The spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient are simply referred to as S, C, K, and D respectively;

[0023] Step 32: Input the input vector P1 into the BP neural network model to obtain the actual output vector, and this actual output vector is the predicted value of the cutting thickness value Y;

[0024] Step 33: Input the output vector T1 into the BP neural network model, and calculate the root mean square error between the predicted value of the cutting thickness value Y and the expected value;

[0025] Step 34: Use this root mean square error as the input data of the BP neural network error backpropagation algorithm, and train the BP neural network model repeatedly until the error between the predicted value of the cutting thickness value Y output and the expected value is less than the set expected error W. The model training is completed, and the BP neural network model is saved;

[0026] Step 35: Obtain a BP neural network model applicable to the prediction of the cutting thickness of SMILE surgery.

[0027] The present invention helps to screen the indications for SMILE surgery. Especially for doctors who are initially performing SMILE surgery, they can conveniently and effectively obtain the patient's surgical parameters without relying on the Zeiss VISUMAX machine for input verification, assisting doctors in selecting a safe and reasonable surgical method.

[0028] For the optimization of the technical solution of the present invention, in step 2, the hidden layer of the BP neural network model can have one or more layers.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The prior art can only verify and judge the patient's surgical parameters through the manufacturer's machine, which may occupy the surgical time and have low efficiency. In addition, the look-up table method also has a certain feasibility, but the look-up table method is only effective for discrete input data and has low execution efficiency in the case of multi-dimensional parameter input. However, the network after deep learning training reaches the expected error, and for any input parameter within a reasonable range, it can achieve fast and accurate data prediction, quickly judge whether the patient meets the SMILE surgery conditions, and save clinical time.

[0031] The present invention performs deep learning on a large amount of existing data based on the BP neural network and can realize the prediction function independently of the SMILE surgery machine. Therefore, on the basis of ensuring the prediction accuracy, the present invention can fundamentally realize the scientific and rapid screening of the indications for the femtosecond laser in situ keratomileusis surgery and save clinical time. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of the construction method of the parameter prediction model related to the present invention.

[0033] Figure 2 It is a flowchart of the SMILE surgery parameter prediction method based on the BP neural network model in Embodiment 1 of the present invention.

[0034] Figure 3 It is a schematic diagram of the composition of the BP neural network for predicting the cutting thickness of the SMILE surgery in Embodiment 1 of the present invention.

[0035] Figure 4 It is the training result of the fitting degree of the training samples.

[0036] Figure 5 It is the training result of the fitting degree of the verification samples.

[0037] Figure 6 It is the training result of the fitting degree of the test samples.

[0038] Figure 7 It is the training result of the fitting degree of all samples.

[0039] Figure 8 It is a schematic diagram showing the change of the training effect of the BP neural network model in training step 2 with the number of iterations. Specific implementation manners

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the following further elaborates on the present invention in conjunction with the appended Figure 1 - appended Figure 8 drawings and embodiments. It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] The following further elaborates on the present invention, where tansig, logsig, and traingdx are existing calculation formulas or functions.

[0042] As Figure 1 shown, a method for constructing a parameter prediction model specifically includes the following steps:

[0043] Step 1: According to the official data provided by the manufacturer and the historical patient data, select four indicators, namely the spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient, as the input values in the model training stage, and select the corresponding cutting thickness value Y of each input quantity as the expected output value in the model training stage;

[0044] Step 2: According to the input quantity and the expected output quantity in Step 1, construct a BP neural network model. The BP neural network model includes a three-layer feedforward neural network structure, namely an input layer, a hidden layer, and an output layer. The input indicators of the input layer are the input quantities selected in Step 1, and the output indicators of the output layer are the expected output quantities; according to the actual prediction accuracy requirements, set the expected error W;

[0045] Among them, the activation functions of the hidden layer and the output layer are both selected as tansig, the network training function is trainglm, and the network performance function is mse;

[0046] The BP neural network model selects the S-shaped transfer function logsig, and its expression is Through the backpropagation error function continuously adjust the network weights and thresholds to make the error function E less than the expected error W; where, T i is the expected output, and Q i is the network calculated output;

[0047] For the number of neurons in the BP neural network model, the number of neurons L in the hidden layer is determined by referring to the following formula:

[0048]

[0049] c is the number of nodes in the input layer, b is the number of nodes in the output layer, and a is a constant between [1, 10];

[0050] Step 3: Train the BP neural network model in Step 2; the specific training steps are as follows:

[0051] Step 31: Generate an input vector P1 from the data of the 4 indicators selected in Step 1; generate an output vector T1 from the data of the expected output values selected in Step 1, and this output vector T1 is used as the expected output vector;

[0052] P1 = [S1, S2, S3,..., S n ; C1, C2, C3,..., C n ; K1, K2, K3,..., K n ; D1, D2, D3,..., D n ; T1 = [Y1, Y2, Y3,..., Y n ;

[0053] Among them, n is the number of historical sample data, and n takes the value of 13188; the input vector P consists of the spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient. The spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient are abbreviated as S, C, K, and D respectively;

[0054] Step 32: Input the input vector P1 into the BP neural network model to obtain the actual output vector, and this actual output vector is the predicted value of the cutting thickness value Y;

[0055] Step 33: Input the output vector T1 into the BP neural network model, and calculate the root mean square error between the predicted value of the cutting thickness value Y and the expected value;

[0056] Step 34: Use this root mean square error as the input data of the error backpropagation algorithm of the BP neural network, and train the BP neural network model repeatedly until the error between the predicted value of the cutting thickness value Y and the expected value is less than the set expected error W, the model training is completed, and save this BP neural network model;

[0057] Step 35: Obtain a BP neural network model suitable for predicting the cutting thickness of SMILE surgery.

[0058] The application example of the parameter prediction model in this embodiment is as follows:

[0059] Example 1

[0060] As Figure 2 shown, this example proposes a method for predicting SMILE surgical parameters based on a BP neural network model, which specifically includes the following steps:

[0061] Step 1: According to the official data provided by the manufacturer and the historical patient data, select four indicators, namely the spherical lens power (SPH), cylindrical lens power (CYL), corneal curvature radius (K), and stromal lens diameter (Diameter) of the patient as the input values in the model training stage, and select the corresponding cutting thickness value Y of the input quantity as the expected output value in the model training stage;

[0062] Step 2: According to the input quantity and expected output quantity in Step 1, construct a BP neural network model. The BP neural network model includes a three-layer feedforward neural network structure, namely the input layer, hidden layer, and output layer. The input indicators of the input layer are the input quantities selected in Step 1, and the output indicators of the output layer are the expected output quantities; According to the actual prediction accuracy requirements, set the expected error W;

[0063] Among them, the activation functions of the hidden layer and the output layer are both selected as tansig, the network training function is trainglm, and the network performance function is mse;

[0064] The BP neural network model selects the S-shaped transfer function logsig, and its expression is Through the backpropagation error function Continuously adjust the network weights and thresholds to make the error function E less than the expected error W; Among them, T i is the expected output, and Q i is the network calculated output;

[0065] The number of neurons in the BP neural network model, and the number of neurons L in the hidden layer are determined by referring to the following formula:

[0066]

[0067] c is the number of nodes in the input layer, b is the number of nodes in the output layer, and a is a constant between [1, 10];

[0068] Step 3: Train the BP neural network model in Step 2; The specific training steps are as follows:

[0069] Step 31: Generate an input vector P1 with the data of the four indicators selected in Step 1; Generate an output vector T1 with the data of the expected output value selected in Step 1, and this output vector T1 is used as the expected output vector;

[0070] P1 = [S1, S2, S3,..., Sn ; C1, C2, C3, ..., C n ; K1, K2, K3, ..., K n ; D1, D2, D3, ..., D n ;

[0071] T1 = [Y1, Y2, Y3, ..., Y n ;

[0072] Among them, n is the number of historical sample data, and n takes the value of 13188; the input vector P is composed of the spherical lens power (SPH) of the patient, the cylindrical lens power (CYL), the corneal curvature radius (K), and the stromal lens diameter (Diameter). The spherical lens power (SPH), the cylindrical lens power (CYL), the corneal curvature radius (K), and the stromal lens diameter (Diameter) of the patient are simply referred to as S, C, K, and D respectively;

[0073] Step 32: Input the input vector P1 into the BP neural network model to obtain the actual output vector, and this actual output vector is the predicted value of the cutting thickness Y;

[0074] Step 33: Input the output vector T1 into the BP neural network model, and calculate the root mean square error between the predicted value of the cutting thickness Y and the expected value;

[0075] Step 34: Use this root mean square error as the input data of the error backpropagation algorithm of the BP neural network, and train the BP neural network model repeatedly until the error between the predicted value of the cutting thickness Y output and the expected value is less than the set expected error W, the model training is completed, and save this BP neural network model;

[0076] Step 35: Obtain the BP neural network model applicable to the prediction of the cutting thickness of SMILE surgery.

[0077] Step 4: According to the surgical data of the current patient to be predicted, select the four indicators of S, C, K, and D as the input values in the model application stage, and generate the input vector P2 from the data of the selected input values in the model application stage,

[0078] P2 = [S1; C1; K1; D1];

[0079] Step 5: Model application, input the input vector P2 in Step 4 into the BP neural network model in Step 35 to obtain the output vector T2, and this output vector T2 is the predicted value of the cutting thickness Y of this patient;

[0080] Step 6: Calculate the predicted value of the remaining basal thickness, and the formula is as follows:

[0081] Remaining basal thickness = corneal thickness - basal thickness - cap thickness - cutting thickness value Y

[0082] Among them, the cutting thickness value Y in the formula is the predicted value of the cutting thickness value Y obtained in step 5, and the corneal thickness, basal thickness, and cap thickness are the surgical data of the current patient;

[0083] Step 7: If the remaining basal thickness ≥ 300 μm, it indicates a low surgical risk and surgery can be considered; if the remaining stromal thickness < 280 μm, it indicates a relatively high surgical risk and surgery is not recommended; if the remaining stromal thickness is between 280 μm and 300 μm, surgery can be performed under certain circumstances.

[0084] In Example 1 of the present invention, when the remaining stromal thickness is between 280 μm and 300 μm, surgery can be performed under certain circumstances. At this time, the remaining corneal stromal thickness can be increased to the surgically acceptable range by appropriately adjusting the preset Diameter, basal thickness, cap diameter, and cap thickness. However, there is a certain range for adjusting the preset parameters. For example, if the Diameter is too small, it is necessary to consider that the introduced higher-order aberrations may be too large, thus affecting the visual quality. Therefore, it is necessary to reasonably select the preset parameters of the patient to select a safer, more effective, and reasonable surgical plan.

[0085] Such as Figure 3 As shown, in step 2 of the embodiment of the present invention, the hidden layer of the BP neural network model can have one or more layers; among them, the activation functions of the hidden layer and the output layer are both selected as tansig, the network training function is trainglm, the network performance function is mse, the number of training times is set to 5000, and the expected error W is set to 10 according to the actual prediction accuracy requirements -3 。

[0086] In this embodiment, the BP neural network model takes four groups of parameter indicators as inputs and the cutting thickness value as the output. The number of nodes c in the input layer is 4, the number of nodes b in the output layer is 1, and the number of neurons L in the hidden layer is determined by referring to the following formula:

[0087]

[0088] Among them, a is a constant between [1, 10]. Therefore, the number of neurons in the hidden layer is ultimately determined by the training effect of the network.

[0089] For example, in the embodiment, a is taken as 3.764, so the number of neurons L in the hidden layer is set to 6. Using the root mean square error between the predicted cutting thickness value and the expected cutting thickness value as the input data for the error backpropagation algorithm of the BP neural network, the BP neural network model is trained repeatedly until the error between the output predicted value and the actual value is less than the set threshold of 10 -3 is the end of training.

[0090] In this embodiment, in step 3, the BP neural network model in step 2 is trained. After the training is completed, the fitting degree between the neural network and the corresponding data is measured by drawing a regression line, as Figure 4 , 5 , 6 and 7 show.

[0091] Figure 4 is the training result of the fitting degree of the training samples, Figure 5 is the training result of the fitting degree of the validation samples, Figure 6 is the training result of the fitting degree of the test samples, Figure 7 is the training result of the fitting degree of all samples; where R refers to the regression coefficient, and the closer R is to 1, the better the fitting degree.

[0092] In this embodiment, for the BP neural network model in step 2, it is randomly grouped, and the input vector P1 and the output vector T are divided into three data sets, one is the training set, one is the validation set, and one is the test set.

[0093] In the process of step 3 training the BP neural network model in step 2, the prediction effectiveness of the BP neural network model is verified through the data of the validation set, and the test is carried out through the data of the test set; if the root mean square error of the verification and test results is large compared with the expected error, then it will return to the training set for continued training until the result less than the expected error is achieved, or the pre-set number of iterations in the early stage is completed. If the regression analysis of the results of the three data sets can all reach the expected results, then the BP neural network model will be saved in the.mat format.

[0094] As Figure 8 shows, in this embodiment, in step 3, the BP neural network model in step 2 is trained, and the best validation performance is achieved after the number of iterations reaches 2313.

[0095] In this embodiment, in step 3, after training, the trained BP neural network model is saved to obtain a BP neural network model suitable for predicting the cutting thickness of SMILE surgery; the trained BP neural network model is saved in the.mat file format. After providing the input vector to this file, the output vector, that is, the predicted value of the cutting thickness value Y, can be obtained.

[0096] After the total of 13,188 historical sample data are input into the saved network model in reverse, the reverse prediction result can be obtained. Since the predicted cutting thickness value should be an integer in the actual use process, the residual will also be rounded. After rounding, the distribution is shown in Table 1. On the premise that the machine provides a data total of 13,188, the final error accuracy is (-0.00083 ± 0.561255) μm.

[0097] Table 1 Distribution of the residuals after rounding the results of the reverse prediction machine-provided data

[0098]

[0099] In the prediction method of this embodiment, in step 3, after training, the trained BP neural network model is saved to obtain a BP neural network model applicable to the prediction of the cutting thickness of SMILE surgery. To verify the accuracy of the saved BP neural network model, 4840 pieces of historical sample data information of patients are input into the BP neural network model. By comparing the residuals between the predicted cutting thickness value and the actual cutting thickness value, its prediction accuracy is verified. The results are shown in Table 2. On the premise that the total amount of actual patient data is 4840, the final error accuracy is (-0.003791 ± 0.42211) μm.

[0100] Table 2 Distribution of the residuals after rounding of the results of reverse prediction of actual patient data

[0101]

[0102]

[0103] From the results, the vast majority of the errors are distributed within the range of ±1 μm, and with the increase in the amount of data and the continuous optimization and training of the BP neural network model, the accuracy of the method involved in the present invention will be higher and higher.

[0104] In the embodiment of the present invention, in step 4, for the BP neural network model specifically applicable to the prediction of the cutting thickness of SMILE surgery, according to the current surgical data of the patient to be predicted, the BP neural network model is used to calculate the predicted value of the cutting thickness value Y of this patient. Input a set of patient surgical data into the BP neural network model to obtain a predicted value of the cutting thickness value Y.

[0105] The above embodiments are only used to illustrate the technical idea of the present invention, and the protection scope of the present invention cannot be limited by this. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the present invention.

Claims

1. A method for constructing a parameter prediction model, characterized in that, It includes the following steps: Step 1: According to the official data provided by the manufacturer and the historical patient data, select four indicators, namely the spherical lens power (SPH), the cylindrical lens power (CYL), the corneal curvature radius (K), and the stromal lens diameter (Diameter), as the input values in the model training stage, and select the corresponding cutting thickness value Y for each input quantity as the expected output value in the model training stage; Step 2: Based on the input quantity and the expected output quantity in Step 1, construct a BP neural network model. The BP neural network model includes a three-layer feedforward neural network structure, namely the input layer, the hidden layer, and the output layer. The input indicators of the input layer are the input quantities selected in Step 1, and the output indicators of the output layer are the expected output quantities; Set the expected error W according to the actual prediction accuracy requirements; Among them, the activation functions of the hidden layer and the output layer are both selected as tansig, the network training function is trainglm, and the network performance function is mse; The BP neural network model selects the sigmoid transfer function logsig, and its expression is Through the backpropagation error function Continuously adjust the network weights and thresholds to make the error function E less than the desired error W; where T i is the desired output, and Q i is the network calculated output; The number of neurons in the BP neural network model, and the number of neurons L in the hidden layer are determined by referring to the following formula: c is the number of nodes in the input layer, b is the number of nodes in the output layer, and a is a constant between [1, 10]; Step 3: Train the BP neural network model in Step 2; The specific training steps are as follows: Step 31: Generate an input vector P1 from the data of the four indicators selected in Step 1; Generate an output vector T1 from the data of the expected output value selected in Step 1, and this output vector T1 is used as the expected output vector; P1 = [S1, S2, S3, ..., S n ; C1, C2, C3, ..., C n ; K1, K2, K3, ..., K n ; D1, D2, D3, ..., D n ; T1 = [Y1, Y2, Y3, ..., Y n ; Among them, n is the number of historical sample data, the input vector P is composed of the spherical lens power (SPH), the cylindrical lens power (CYL), the corneal curvature radius (K), and the stromal lens diameter (Diameter). The spherical lens power (SPH), the cylindrical lens power (CYL), the corneal curvature radius (K), and the stromal lens diameter (Diameter) are abbreviated as S, C, K, and D respectively; Step 32: Input the input vector P1 into the BP neural network model to obtain the actual output vector, and this actual output vector is the predicted value of the cutting thickness value Y; Step 33: Input the output vector T1 into the BP neural network model, and calculate the root mean square error between the predicted value and the expected value of the cutting thickness value Y; Step 34: Use this root mean square error as the input data of the BP neural network error backpropagation algorithm, and perform cyclic training on the BP neural network model until the error between the predicted value and the expected value of the output cutting thickness value Y is less than the set expected error W, the model training is completed, and save this BP neural network model; Step 35: Obtain a BP neural network model applicable to the prediction of the cutting thickness of SMILE surgery.

2. The method for constructing a parameter prediction model according to claim 1, wherein In Step 2, the hidden layer of the BP neural network model can have one or more layers.

Images