Method for establishing fetal fetal lung maturity recognition model based on ultrasonic imaging omics and fetal fetal lung maturity recognition model

Abstract

The invention provides a method for establishing a fetal lung maturity recognition model based on ultrasonic imaging omics. The method comprises the following steps: acquiring lung ultrasonic image data of a normal fetus with the gestational age of 20 weeks to 40 + 6 weeks; acquiring corresponding image data according to the lung ultrasonic image data, and sketching a region of interest in each lung ultrasonic image in the image data; carrying out primary screening of image omics characteristics on the delineated region of interest; carrying out secondary screening on the image omics features screened out for the first time by using ANOVA F inspection, and carrying out dimension reduction processing on the features obtained after the secondary screening; and taking the image omics characteristics used for establishing the fetal lung maturity identification model as input, taking the pregnant woman pregnancy age as a label as output, training an AdeptiveXGBoost classifier, and storing the trained AdeptiveXGBoost classifier so as to obtain the fetal lung maturity identification model. According to the method, the aim of identifying and predicting the maturity of the fetal lung of the fetus is fulfilled through a plurality of finally screened radiomics features.

Description

Technical Field

[0001] The present invention relates to the technical field of fetal lung maturity recognition, and in particular to a method for establishing a fetal lung maturity recognition model based on ultrasound imaging, and a fetal lung maturity recognition model. Background Art

[0002] In recent years, the incidence of pregnancy complications has increased year by year, especially hypertensive disorders of pregnancy (HDP), the pathogenesis of which is the disorder of trophoblast remodeling of the mother's uterine spiral artery, which leads to placental ischemia, hypoxia and dysfunction. The fetus cannot obtain enough oxygen and nutrients from the placenta, and is in a state of chronic hypoxia for a long time. This state makes the fetal lung development lag behind that of a normal fetus of the same gestational age, that is, the fetal lung is immature, which can cause many respiratory diseases in newborns and affect their long-term quality of life. Therefore, the assessment of fetal lung maturity is helpful for clinical decision-making and personalized care after the fetus is born. However, current prenatal ultrasound cannot accurately assess fetal lung maturity.

[0003] Radiomics can automatically analyze a large amount of image information that cannot be discerned by the naked eye from medical images at a high throughput, and extract features of medical images using machine learning methods for disease diagnosis and prediction. How to use ultrasound radiomics features to help solve the problem of fetal lung maturity assessment is currently the focus of research by scholars.

[0004] In the related technology, Du Yanran, Jiao Jing, Ren Yunyun, Zhou Jianqiao and others disclosed the "Application of Ultrasound Imaging Omics Technology in Assessing Fetal Lung Maturity". Each pregnant woman collected a standard fetal lung ultrasound image collected within 72 hours before delivery. Based on ultrasound imaging omics technology, fetal lung texture analysis was performed on the fetal lung ultrasound image. Combined with the presence or absence of pregnancy complications, high-throughput imaging omics features were extracted. Taking the neonatal prognosis results as the gold standard, fetal lung maturity assessment models suitable for different gestational weeks were established respectively. Subsequently, in the corresponding validation set, the fetal lung maturity model was verified to predict the risk of respiratory diseases in newborns. However, this study mainly focuses on the relationship between fetal lung maturity and the prediction of respiratory diseases in newborns. However, in many cases, due to pregnancy complications of pregnant women, the fetal lung maturity of the fetus is inconsistent with the normal fetal lung maturity corresponding to the gestational age of the pregnant woman, and this study cannot analyze and identify the fetal lung maturity corresponding to the gestational age of the fetus based on ultrasound imaging omics features. For example, the gestational age is 30 weeks, but the fetal lung maturity only corresponds to the normal fetal lung maturity of a 27-week fetus. If the fetal lung maturity is not monitored in time during pregnancy, respiratory diseases may occur in the fetus after birth, and personalized care will not be provided in a timely and effective manner. Summary of the invention

[0005] In view of the problems existing in the above-mentioned related technologies, the present invention aims to provide a method for establishing a fetal lung maturity recognition model based on ultrasound imaging omics. Specifically, the method for establishing a fetal lung maturity recognition model based on ultrasound imaging omics, by using several imaging omics features in ultrasound images as recognition features of fetal lung maturity, achieves the purpose of identifying the gestational age of the pregnant woman corresponding to the fetal lung maturity of the fetus based on ultrasound imaging omics features, so as to provide an effective basis for the formulation of personalized nursing methods for the fetus after birth.

[0006] To this end, the first purpose of the present invention is to propose a method for establishing a fetal lung maturity recognition model based on ultrasound imaging.

[0007] The second objective of the present invention is to provide a fetal lung maturity recognition model.

[0008] In order to achieve the first objective, the technical solution of the present invention provides a method for establishing a fetal lung maturity recognition model based on ultrasound imaging. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging comprises: step S1: collecting fetal lung maturity recognition models of gestational age from 20 weeks to 40 weeks; +6 Lung ultrasound image data of normal fetuses at 40 weeks of age; +6 Week means 40 weeks plus 6 days; step S2: acquiring corresponding image data according to the lung ultrasound image data, and outlining the region of interest in each lung ultrasound image in the image data; step S3: performing primary screening of imaging genomics features on the outlined region of interest; step S4: performing secondary screening of the imaging genomics features initially screened using the ANOVA F test, and then performing dimensionality reduction processing on the features obtained after the secondary screening to determine the imaging genomics features for establishing a fetal lung maturity recognition model; step S5: using the imaging genomics features for establishing a fetal lung maturity recognition model as input, and using the gestational age of the pregnant woman as a label as output, training an AdeptiveXGBoost classifier, and saving the trained AdeptiveXGBoost classifier to obtain a fetal lung maturity recognition model; wherein the fetal lung maturity recognition model realizes the recognition of the gestational age of the pregnant woman corresponding to the fetal lung maturity.

[0009] Preferably, the lung ultrasound image data described in step S1 includes one of the following or a combination thereof: fetal weight, fetal chest circumference, fetal abdominal circumference, fetal bilateral lung volume, fetal main pulmonary artery pulsatility index, fetal main pulmonary artery resistance index, fetal main pulmonary artery valve systolic peak flow velocity, fetal main pulmonary artery systolic blood flow acceleration time, and the ratio of fetal main pulmonary artery systolic blood flow acceleration time to ejection time.

[0010] Preferably, the specific operation of step S3 includes the following steps: step S301: performing norm normalization preprocessing on the outlined region of interest and performing preliminary extraction of imaging genomics features; step S302: constructing a FEL autoencoder, performing initial screening on the preliminary extracted imaging genomics features, and obtaining 128 imaging genomics features; wherein the FEL autoencoder includes an input layer, an encoder, a decoder and an output layer connected in sequence.

[0011] Preferably, the encoder is a feedforward neural network, and the encoder includes a first fully connected layer, a first Dropout layer, a second fully connected layer, a second Dropout layer, and a third fully connected layer connected in sequence; the mathematical expression corresponding to the calculation process of the encoder is:

[0012] h=f(W e x+b e ) (1)

[0013] In formula (1), x is the manually drawn region of interest, and the dimension of x is d. x is compressed into a latent space, which is represented by h, and the dimension of h is k. Then k <d;W e is the weight matrix of the encoder; b e is the bias vector of the encoder; f is the activation function.

[0014] Preferably, the decoder is also a feedforward neural network, and the decoder includes a fourth fully connected layer, a third Dropout layer, a fifth fully connected layer, and a fourth Dropout layer connected in sequence; the mathematical expression corresponding to the calculation process of the decoder is:

[0015]

[0016] In formula (2), W d is the weight matrix of the decoder; b d is the bias vector of the decoder; g is the activation function of the decoder; is the reconstructed input data.

[0017] Preferably, the loss function of the FEL autoencoder is a cross entropy loss function, which is used to measure the original input data x and the reconstructed input data The difference between.

[0018] Preferably, the specific operation of step S4 includes the following steps: step S401: using the ANOVA F test to perform a secondary screening of the imaging genomics features initially screened out in step S3 to determine the features of the fetal lung region of interest that are linearly correlated with the gestational age and select the top 5% features; step S402: based on the recursive feature elimination method of Dynamic RFECV, the top 5% features are subjected to dimensionality reduction processing to determine the imaging genomics features for establishing a fetal lung maturity recognition model.

[0019] In order to achieve the above-mentioned second purpose, the technical solution of the present invention also provides a fetal lung maturity recognition model, which is a fetal lung maturity recognition model established by using the method for establishing a fetal lung maturity recognition model based on ultrasound imaging genomics in any of the above-mentioned technical solutions.

[0020] Beneficial effects of the present invention:

[0021] (1) The method for establishing a fetal lung maturity recognition model based on ultrasound imaging provided by the present invention is based on ultrasound images and a number of imaging features that are finally screened out to achieve the purpose of identifying and predicting the fetal lung maturity of the fetus. When the gestational age corresponding to the identified fetal lung maturity is inconsistent with the actual gestational age of the pregnant woman, it is possible to predict the respiratory diseases that may occur in the fetus after birth based on the fetal lung maturity, which helps to formulate personalized treatment and reasonable plans in advance, thereby reducing various risks caused by respiratory diseases after the fetus is born.

[0022] (2) The method for establishing a fetal lung maturity recognition model based on ultrasound imaging provided by the present invention maps the features of the manually drawn region of interest (ROI) to a latent space of lower dimension by constructing a FEL autoencoder, thereby achieving the purpose of dimensionality reduction, helping to remove redundant features and reducing computational complexity. At the same time, it also has a certain anti-interference ability against noise and outliers, so that the selected feature representation is more reliable and more suitable for subsequent fetal lung maturity analysis tasks.

[0023] Additional aspects and advantages of the invention will become apparent from the following description, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic flow chart showing a method for establishing a fetal lung maturity recognition model based on ultrasound imaging omics according to an embodiment of the present invention;

[0025] Figure 2 A schematic flow chart showing a method for establishing a fetal lung maturity recognition model based on ultrasound imaging omics according to another embodiment of the present invention;

[0026] Figure 3 A schematic diagram showing the radiomics features of 51 fetal lung ROIs that are linearly correlated with gestational age determined using the ANOVA F test according to an embodiment of the present invention.

[0027] Figure 4 A schematic diagram showing 12 radiomics features for establishing a fetal lung maturity recognition model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0030] Figure 1 FIG. 1 is a schematic flow chart showing a method for establishing a fetal lung maturity recognition model based on ultrasound imaging genomics according to an embodiment of the present invention. Figure 1 As shown, the method for establishing the fetal lung maturity recognition model based on ultrasound imaging genomics includes:

[0031] Step S1: Collect data from gestational age between 20 weeks and 40 weeks +6 Lung ultrasound image data of normal fetuses at 40 weeks of age; +6 Week means 40 weeks plus 6 days;

[0032] Step S2: acquiring corresponding image data according to the lung ultrasound image data, and outlining a region of interest in each lung ultrasound image in the image data;

[0033] Step S3: initial screening of radiomics features for the delineated regions of interest;

[0034] Step S4: ANOVA F test is used to perform secondary screening on the imaging features obtained from the primary screening, and then the features obtained after the secondary screening are subjected to dimensionality reduction processing to determine the imaging features used for establishing the fetal lung maturity recognition model;

[0035] Step S5: Using the radiomics features used to establish the fetal lung maturity recognition model as input, and using the gestational age of the pregnant woman as a label as output, training the AdeptiveXGBoost classifier, and saving the trained AdeptiveXGBoost classifier to obtain the fetal lung maturity recognition model; wherein the fetal lung maturity recognition model can recognize the gestational age of the pregnant woman corresponding to the fetal lung maturity.

[0036] In this embodiment, the method for establishing a fetal lung maturity recognition model based on ultrasound imaging provided by the present invention is based on ultrasound images and, through the final screening of a number of imaging features, achieves the purpose of identifying and predicting the fetal lung maturity of the fetus. When the gestational age corresponding to the identified fetal lung maturity is inconsistent with the actual gestational age of the pregnant woman, it is possible to predict the respiratory diseases that may occur after the birth of the fetus based on the fetal lung maturity, which helps to formulate personalized treatments and reasonable plans in advance, thereby reducing various risks caused by respiratory diseases after the birth of the fetus.

[0037] The method for establishing a fetal lung maturity recognition model based on ultrasound imaging provided by the present invention maps the features of the manually drawn region of interest (ROI) to a latent space of lower dimension by constructing a FEL autoencoder, thereby achieving the purpose of dimensionality reduction, helping to remove redundant features and reducing computational complexity. At the same time, it also has a certain anti-interference ability to noise and outliers, so that the selected feature representation is more reliable and more suitable for subsequent fetal lung maturity analysis tasks.

[0038] In one embodiment of the present invention, the lung ultrasound image data described in step S1 includes one of the following or a combination thereof: fetal weight, fetal chest circumference, fetal abdominal circumference, fetal bilateral lung volume, fetal main pulmonary artery pulsatility index, fetal main pulmonary artery resistance index, fetal main pulmonary artery valve systolic peak flow velocity, fetal main pulmonary artery systolic blood flow acceleration time, and the ratio of fetal main pulmonary artery systolic blood flow acceleration time to ejection time.

[0039] In one embodiment of the present invention, the specific operation of step S3 includes the following steps: step S301: performing norm normalization preprocessing on the outlined region of interest and performing preliminary extraction of imaging genomics features; step S302: constructing a FEL autoencoder, performing initial screening on the preliminary extracted imaging genomics features, and obtaining 128 imaging genomics features; wherein the FEL autoencoder includes an input layer, an encoder, a decoder, and an output layer connected in sequence.

[0040] In one embodiment of the present invention, the encoder is a feedforward neural network, and the encoder includes a first fully connected layer, a first Dropout layer, a second fully connected layer, a second Dropout layer, and a third fully connected layer connected in sequence; the mathematical expression corresponding to the calculation process of the encoder is:

[0041] h=f(W e x+b e ) (1)

[0042] In formula (1), x is the manually drawn region of interest, and the dimension of x is d. x is compressed into a latent space, which is represented by h, and the dimension of h is k. Then k <d;W e is the weight matrix of the encoder; b e is the bias vector of the encoder; f is the activation function.

[0043] In one embodiment of the present invention, the decoder is also a feedforward neural network, and the decoder includes a fourth fully connected layer, a third Dropout layer, a fifth fully connected layer, and a fourth Dropout layer connected in sequence; the mathematical expression corresponding to the calculation process of the decoder is:

[0044]

[0045] In formula (2), W d is the weight matrix of the decoder; b d is the bias vector of the decoder; g is the activation function of the decoder; is the reconstructed input data.

[0046] In one embodiment of the present invention, the loss function of the FEL autoencoder is a cross entropy loss function, which is used to measure the original input data x and the reconstructed input data The difference between.

[0047] In one embodiment of the present invention, the specific operation of step S4 includes the following steps: step S401: using the ANOVA F test to perform a secondary screening of the imaging genomics features initially screened in step S3 to determine the features of the fetal lung region of interest that are linearly correlated with the gestational age and select the top 5% features; step S402: based on the recursive feature elimination method of DynamicRFECV, the top 5% features are subjected to dimensionality reduction processing to determine the imaging genomics features used for establishing a fetal lung maturity recognition model.

[0048] In one embodiment of the present invention, the technical solution of the present invention also provides a fetal lung maturity recognition model, which is a fetal lung maturity recognition model established by using the method for establishing a fetal lung maturity recognition model based on ultrasound imaging in any of the above embodiments.

[0049] The technical solution of the present invention will be demonstrated with a specific embodiment below. Figure 2 As shown, the implementation steps of the method for establishing a fetal lung maturity recognition model based on ultrasound imaging genomics in this specific embodiment of the present invention are as follows:

[0050] 1. Set up the inclusion and exclusion criteria

[0051] Instrument: GE Voluson E10 ultrasonic diagnostic instrument, probe model C6-1, frequency 1-6MHZ, acquisition 20 weeks to 40 +6 Two-dimensional ultrasound image of the fetal lungs in a normal pregnant woman with a fetal four-chamber heart. +6 A week means 40 weeks plus 6 days.

[0052] Inclusion criteria for pregnant women: ① No abnormalities in medical history, physical examination, routine electrocardiogram, blood sugar, and blood pressure; ② Single pregnancy, clear last menstrual period; ③ No known congenital malformation or chromosomal abnormality in the fetus; ④ Women did not take steroids before delivery. Exclusion criteria: ① Oligohydramnios; ② Respiratory diseases after delivery.

[0053] 2. Obtain clinical data

[0054] The 20 to 40 +6 The invention relates to a method for measuring the lungs of a normal fetus at 4 weeks of age using ultrasound data. The lung ultrasound image data obtained by these measurements include one of the following or a combination thereof: fetal body weight (EFW for short), fetal chest circumference (TC for short), fetal abdominal circumference (AC for short), fetal bilateral lung volume (FLV for short), fetal main pulmonary artery pulsatility index (PI for short), fetal main pulmonary artery resistance index (RI for short), fetal main pulmonary artery valve systolic peak flow velocity (PSV for short), fetal main pulmonary artery systolic blood flow acceleration time (AT for short) and the ratio of fetal main pulmonary artery systolic blood flow acceleration time to ejection time (AT / ET for short).

[0055] 3. Obtain image data (directly input the acquired image in practice), outline the region of interest and perform feature screening;

[0056] (1) Obtain imaging data: Select the level of the fetal four-chamber heart cross-section, observe the fetal chest cross-section, and adjust the probe to ensure that there is no obvious fetal rib acoustic shadow in at least one lung. To obtain the best image quality, adjust the settings according to the relevant characteristics of each pregnant woman and fetus, including depth, gain, frequency, time gain compensation, and harmonics for obese patients.

[0057] (2) Manual delineation of the region of interest (ROI) of the acquired image data: Two ultrasound physicians manually delineated the contours of the bilateral lung parenchyma (ROI), avoiding the ribs and large blood vessels. If the physicians disagreed, another senior ultrasound physician (with more than ten years of work experience) participated in the discussion until a group consensus was reached. After obtaining the region of interest, interpolation processing of the region of interest data was also performed. For missing data, the average value was used as the difference.

[0058] (3) For ROI, perform radiomics feature extraction and screening;

[0059] Specifically, ① the delineated ROI was preprocessed by norm normalization;

[0060] More specifically, the process of norm normalization is: divide each element of each dimension of all samples by the L2 norm of the feature to make it have a unit norm.

[0061] For a given feature X∈R N×d , where N is the total number of samples, d is the feature dimension, and the norm normalization is applied to each column feature x∈R N Do the following:

[0062]

[0063] Where 1≤i≤N and norm(x) is the norm of x.

[0064] The L2 norm of feature X is

[0065]

[0066] The delineated ROI was divided into a training set and a validation set (8:2). The Python open source data package was used to extract features of the region of interest, and 1130 radiomic features (5 clinical features + 1125 radiomic features) were obtained. These features were further screened to obtain the most effective features for classification modeling and to optimize the performance of the model.

[0067] ②Perform feature screening

[0068] In this specific embodiment, an FEL autoencoder (abbreviated as FELAutoEncode) was constructed to train the features, so as to extract effective features from 1130 radiomics features, perform feature screening, and finally extract 128 features.

[0069] The FEL autoencoder includes an input layer, an encoder, a decoder, and an output layer; assuming that the manually delineated ROI is x, the dimension of x is d, and it is desired to compress it into a latent space representation h, the dimension of h is k, then k < d; then an encoder Encoder needs to be constructed. The encoder Encoder is a feedforward neural network, and the mathematical expression corresponding to the calculation process of this encoder Encoder is:

[0070] h = f(W e x + b e ) (1)

[0071] In formula (1), W e is the weight matrix of the encoder; b e is the bias vector of the encoder; f is the activation function.

[0072] The encoder in this specific embodiment includes a first fully connected layer, a first Dropout layer, a second fully connected layer, a second Dropout layer, and a third fully connected layer connected in sequence; the network features of each layer are shown in Table 1 below.

[0073] Corresponding to the encoder, a decoder is needed. The decoder is also a feedforward neural network, and the mathematical expression corresponding to the calculation process of this decoder is:

[0074]

[0075] In formula (2), W d is the weight matrix of the decoder; b d is the bias vector of the decoder; g is the activation function of the decoder; is the reconstructed input data.

[0076] The decoder in this specific embodiment includes a fourth fully connected layer, a third Dropout layer, a fifth fully connected layer, and a fourth Dropout layer connected in sequence; a fully connected layer is connected to the last Dropout layer (i.e., the fourth Dropout layer) of the decoder as the output layer, and the network features of each layer are shown in Table 1 below.

[0077] Table 1 Network features of the autoencoder

[0078]

[0079] The goal of the autoencoder is to minimize the reconstruction error, which is the difference between the original input data x and the reconstructed data Therefore, the loss function of the automatic encoder in this specific embodiment selects the cross entropy loss function.

[0080] After determining the encoder and decoder, the entire autoencoder needs to be trained. The process of training the autoencoder is to minimize the above loss function. This is usually done through the gradient descent method. The specific steps are as follows:

[0081] a. Initialize the parameters of the encoder and decoder;

[0082] b. For each manually drawn ROI x, the encoder calculates the latent representation h;

[0083] c. Take h as the input of the decoder and calculate the reconstructed input data through the decoder

[0084] d. Calculate the loss function

[0085] e. Use the back-propagation algorithm to calculate the gradient of the loss with respect to all parameters;

[0086] f. Update the parameters to reduce the loss, which can be done through gradient descent or other optimization algorithms (such as Adam);

[0087] g. Repeat steps b to f until convergence or the predetermined number of iterations is reached.

[0088] The advantage of using the FEL autoencoder constructed in the present invention for feature screening is that dimensionality reduction is achieved by mapping the features of the manually drawn ROI to a latent space of lower dimension, which helps to remove redundant features and reduce computational complexity. At the same time, it has a certain anti-interference ability to noise and outliers, so that the feature representation obtained is more reliable and more suitable for subsequent analysis tasks. 128 features can be obtained through this step.

[0089] 4. ANOVA F test was used to further screen the 128 radiomic features to determine the features of the fetal lung ROI that were linearly correlated with gestational age and select the top 5% features;

[0090] Specifically, the ANOVA F test was used to find the features of the fetal lung ROI that were linearly correlated with the gestational age by calculating the variance F value, and the top 5% of the features were selected, totaling 51. Figure 3 As shown, Figure 3The image omics features of 51 fetal lung ROIs that were linearly correlated with gestational age determined using the ANOVA F test, and the names of the 51 screened image omics features are shown.

[0091] More specifically, the calculation method of the variance F value is:

[0092] Select pregnant women of different gestational age categories, take one gestational age category as a group, and the pregnant women in each group as individual samples. Let k be the number of groups, nj be the sample size of the jth group, is the sample mean of the jth population, is the grand mean of all data, SS B represents the sum of squares between groups, SS w represents the within-group sum of squares, then:

[0093]

[0094] Total sum of squares SS T It is expressed as:

[0095] SS T =SS B +SS W

[0096] The ANOVA F statistic calculation formula is:

[0097]

[0098] in, is the between-group mean square; is the within-group mean square, and N is the total number of all observations (all training samples).

[0099] 5. The top 5% of features are further reduced in dimension to determine the radiomics features for fetal lung maturity recognition and classification modeling;

[0100] Specifically, the top 5% of the features were reduced in dimension again, and the 51 features obtained in step 4 were selected using the recursive feature elimination method of Dynamic RFECV (Dynamic Recursive Feature Elimination with Cross-Validation): ① All features were evaluated and ranked using the gradient boosted tree (GBDT) model; ② Features were removed with a predetermined step size of 1; ③ After removing some features, the model was retrained using the remaining features, and the performance of the model was evaluated by cross-validation; ④ The learning rate was dynamically adjusted according to the model performance; ⑤ The above steps were iterated until the preset minimum number of features of 12 was reached or the performance improvement was no longer significant, so as to obtain the optimal radiomics features.

[0101] Among them, the basic update of the gradient boosted tree gbdt model is as follows:

[0102] F m (x) = F m-1 (x)+η1.h m (x)

[0103] In the formula, F m (x) represents the predicted value of the model for input x after the mth iteration; F m-1 (x) represents the predicted value at the m-1th iteration; η1 is the learning rate of the gradient boosted tree gbdt model, which is between 0 and 1 and is used to reduce the weight of each tree. m (x) is the output of the m-th tree (that is, the m-th weak learner) for the input x, which is constructed based on the residual (that is, negative gradient) of the current model.

[0104] The performance of the model is evaluated by cross-validation, and the learning rate is adjusted dynamically. If the fluctuation of the score is below 0.1 for three consecutive times, the learning rate is divided by 5 and the next round of iteration is performed to accelerate the convergence of the model. It should be noted that the recursive feature elimination method of Dynamic RFECV is a prior art, and the basic update formula of the gradient boosting tree gbdt model mentioned above only explains the principle of model update. The detailed process of selecting 51 features is not repeated here.

[0105] In this specific embodiment, the final model converges to 12 radiomics features for establishing a fetal lung maturity recognition model. Figure 4 As shown, specifically including:

[0106] wavelet-HL_firstorder_RobustMeanAbsoluteDeviation_us_paramsName1

[0107] gradient_firstorder_Energy_us_paramsName1

[0108] wavelet-HL_gldm_GrayLevelNonUniformity_us_paramsName1

[0109] wavelet-HH_firstorder_10Percentile_us_paramsNamel

[0110] lbp-2D_firstorder_Energy_us_paramsName1

[0111] wavelet-LL_glszm_LargeAreaHighGrayLevelEmphasis_us_paramsName1

[0112] original_glszm_LargeAreaHighGrayLevelEmphasis_us_paramsName1

[0113] wavelet-HL_gldm_DependenceNonUniformity_us_paramsNamel

[0114] wavelet-LL_glszm_GrayLevelNonUniformity_us_paramsNamel

[0115] squareroot_glszm_LargeAreaHighGrayLevelEmphasis_us_paramsName1

[0116] gradient_firstorder_TotalEnergy_us_paramsNamel

[0117] lbp-2D_firstorder_TotalEnergy_us_paramsNamel.

[0118] 6. Use the 12 radiomics features obtained in step 5 as input and the gestational age of the pregnant woman as the label as output to train the AdeptiveXGBoost classifier, and save the trained AdeptiveXGBoost classifier to obtain a recognition model of fetal lung maturity corresponding to gestational age.

[0119] Specifically, the basic update formula of the AdeptiveXGBoost classifier can be expressed as:

[0120]

[0121] in, Represents the predicted value for the i-th sample after the t-th iteration; represents the predicted value of the i-th sample after the t-1th iteration, η2 is the learning rate, and f t (x i ) is the decision tree (weak learner) added in the tth iteration for sample x i The predicted value of .

[0122] In addition, in the optimization objective function of AdaptiveXGBoost, the learning rate also affects the rate of descent of the objective function, that is,

[0123]

[0124] in, is the loss function, Ω(f t ) is the regularization term, and λ is the regularization strength.

[0125] It should be noted here that the updating principle of the AdeptiveXGBoost classifier is also a prior art, and the detailed classification process of the 12 radiomics features is not described in detail in this specific embodiment.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for establishing a fetal lung maturity recognition model based on ultrasound imaging, characterized in that: include: Step S1: Collect data from gestational age between 20 weeks and 40 weeks +6 Lung ultrasound image data of normal fetuses at 40 weeks of age; +6 Week means 40 weeks plus 6 days; Step S2: acquiring corresponding image data according to the lung ultrasound image data, and outlining a region of interest in each lung ultrasound image in the image data; Step S3: initial screening of radiomics features for the delineated regions of interest; Step S4: ANOVA F test is used to perform secondary screening on the imaging features obtained from the primary screening, and then the features obtained after the secondary screening are subjected to dimensionality reduction processing to determine the imaging features used for establishing the fetal lung maturity recognition model; Step S5: Using the radiomics features used to establish the fetal lung maturity recognition model as input, and using the gestational age of the pregnant woman as a label as output, training the AdeptiveXGBoost classifier, and saving the trained AdeptiveXGBoost classifier to obtain the fetal lung maturity recognition model; wherein the fetal lung maturity recognition model can recognize the gestational age of the pregnant woman corresponding to the fetal lung maturity.

2. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 1, characterized in that: The lung ultrasound image data described in step S1 includes one of the following or a combination thereof: fetal weight, fetal chest circumference, fetal abdominal circumference, fetal bilateral lung volume, fetal main pulmonary artery pulsatility index, fetal main pulmonary artery resistance index, fetal main pulmonary artery valve systolic peak flow velocity, fetal main pulmonary artery systolic blood flow acceleration time, and the ratio of fetal main pulmonary artery systolic blood flow acceleration time to ejection time.

3. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 1, characterized in that: The specific operation of step S3 includes the following steps: Step S301: performing norm normalization preprocessing on the delineated region of interest and performing preliminary extraction of radiomics features; Step S302: constructing a FEL autoencoder to perform a preliminary screening of the initially extracted radiomics features to obtain 128 radiomics features; The FEL autoencoder comprises an input layer, an encoder, a decoder and an output layer which are connected in sequence.

4. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 3, characterized in that: The encoder is a feedforward neural network, and includes a first fully connected layer, a first Dropout layer, a second fully connected layer, a second Dropout layer, and a third fully connected layer connected in sequence; the mathematical expression corresponding to the calculation process of the encoder is: h=f(W e x+b e ) (1) In formula (1), x is the manually drawn region of interest, and the dimension of x is d. x is compressed into a latent space, which is represented by h, and the dimension of h is k. Then k <d;W e is the weight matrix of the encoder; b e is the bias vector of the encoder; f is the activation function.

5. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 4, characterized in that: The decoder is also a feedforward neural network, and includes a fourth fully connected layer, a third Dropout layer, a fifth fully connected layer, and a fourth Dropout layer connected in sequence; the mathematical expression corresponding to the calculation process of the decoder is: In formula (2), W d is the weight matrix of the decoder; b d is the bias vector of the decoder; g is the activation function of the decoder; is the reconstructed input data.

6. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 5, characterized in that: The loss function of the FEL autoencoder is a cross entropy loss function, where the cross entropy loss ^ The function is used to measure the difference between the original input data x and the reconstructed input data X.

7. The method for establishing a fetal lung maturity recognition model based on ultrasound imaging according to claim 6, characterized in that: The specific operation of step S4 includes the following steps: Step S401: Perform secondary screening of the radiomics features initially screened in step S3 using the ANOVA F test to determine the features of the fetal lung region of interest that are linearly correlated with the gestational age and select the top 5% features; Step S402: Based on the recursive feature elimination method of Dynamic RFECV, the top 5% of the features are subjected to dimensionality reduction processing to determine the radiomics features for establishing the fetal lung maturity recognition model.

8. A fetal lung maturity recognition model, characterized in that: The recognition model is a fetal lung maturity recognition model established by using the method for establishing a fetal lung maturity recognition model based on ultrasound imaging genomics as described in any one of claims 1 to 7.

Images