A system for predicting fetal growth restriction

Abstract

The present application relates to the technical field of machine learning, in particular to a prediction system for fetal growth restriction, comprising: a collection module, which collects magnetic resonance images and ultrasound data for a patient; a feature extraction module, which extracts an image feature group from the magnetic resonance images, and extracts an ultrasound feature group from the ultrasound data; a joint prediction module, which generates a prediction result according to the image feature group and the ultrasound feature group. The beneficial effect lies in: in view of the problem that the measurement effect of the related indicators for fetal growth restriction in the prior art is poor, by extracting specific image features from the magnetic resonance images as the image feature group, a better representation effect can be achieved for whether the placenta has placental dysfunction, avoiding the influence of subjective factors caused by relying on subjective judgment and evaluation of doctors in the prior art, and by processing the ultrasound data to form the ultrasound feature group, the image feature group and the ultrasound feature group are combined, achieving higher prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a prediction system for fetal growth restriction. BACKGROUND

[0002] Fetal growth restriction (FGR) is a common pregnancy complication worldwide, which refers to the pathological reason that the fetus cannot achieve its genetic growth potential during intrauterine development. The symptom is associated with intrauterine death, premature birth, neonatal death and postpartum complications. Therefore, accurate early diagnosis of fetal growth restriction is an important problem to be solved in clinical practice.

[0003] In the prior art, there are corresponding diagnostic indicators for fetal growth restriction. For example, when the estimated fetal weight obtained by an ultrasound device is lower than the tenth percentile of the corresponding gestational age, and the fetal Doppler parameter is abnormal, it indicates that there may be fetal growth restriction. Or, when there is placental dysfunction, it may also cause fetal growth restriction.

[0004] However, in actual implementation, the inventors found that the above scheme has the problem of poor accuracy of the measured indicators in the implementation process. For example, relevant research shows that there may be more than 15% deviation between the estimated fetal weight obtained by ultrasound and the actual birth weight of the fetus. It is relatively difficult for doctors to distinguish the morphological and functional changes of the placenta through magnetic resonance imaging (MRI), and there are subjective factors. These reasons lead to the deviation of the related indicators measured based on the prior art, and further lead to the decrease of the accuracy of the diagnosis. SUMMARY

[0005] In view of the above problems in the prior art, the present application provides a prediction system for fetal growth restriction.

[0006] The specific technical solutions are as follows:

[0007] A prediction system for fetal growth restriction comprises:

[0008] An acquisition module, which acquires magnetic resonance imaging and ultrasound data for a patient;

[0009] A feature extraction module, which is connected to the acquisition module, extracts an image feature group from the magnetic resonance imaging, and extracts an ultrasound feature group from the ultrasound data;

[0010] a joint prediction module connected to the feature extraction module, the joint prediction module generating a prediction result according to the image feature set and the ultrasound feature set.

[0011] In another aspect, the image feature set comprises placenta flatness, placenta area volume ratio, and placenta region gray scale matrix feature.

[0012] In another aspect, the ultrasound feature set comprises fetal weight, abdominal circumference, and umbilical artery pulsatility index.

[0013] In another aspect, the feature extraction module comprises an image feature extraction module, the image feature extraction module comprising:

[0014] a wavelet filtering module receiving the magnetic resonance image, the wavelet filtering module generating a wavelet filtered image according to the magnetic resonance image;

[0015] a laplacian of gaussian filtering module connected to the wavelet filtering module, the laplacian of gaussian filtering module processing the wavelet filtered image to obtain a derived image;

[0016] a feature generation module connected to the laplacian of gaussian filtering module, the feature generation module extracting image features from the derived image and the magnetic resonance image respectively;

[0017] a feature calculation module generating the image feature set according to the image features.

[0018] In another aspect, the feature generation module comprises:

[0019] a shape extraction module extracting placenta shape from the derived image and the magnetic resonance image and adding to the image features;

[0020] a first order statistics extraction module extracting first order statistics features from the derived image and the magnetic resonance image and adding to the image features;

[0021] a texture feature extraction module extracting placenta texture from the derived image and the magnetic resonance image and adding to the image features.

[0022] In another aspect, the feature extraction module comprises a pulsatility index calculation module, the pulsatility index calculation module comprising:

[0023] a first flow rate generation module generating systolic flow rate and diastolic flow rate according to the ultrasound data;

[0024] a second flow rate generation module configured to generate an average flow rate based on the ultrasound data;

[0025] a pulsatility index generation module connected to the first flow rate generation module and the second flow rate generation module, respectively, the pulsatility index generation module configured to generate the pulsatility index based on the systolic flow rate, the diastolic flow rate and the average flow rate.

[0026] In another aspect, the joint prediction module is configured to generate the prediction result based on the image feature set and the ultrasound feature set using a random forest algorithm;

[0027] The prediction system further comprises a model training module, the model training module comprising:

[0028] a sample feature generation module configured to receive sample data and extract image sample features corresponding to the image feature set and ultrasound sample features corresponding to the ultrasound feature set from the sample data, and combine the image sample features and the ultrasound sample features to form a sample feature set;

[0029] a sampling module connected to the sample feature generation module, the sampling module configured to sample from a plurality of the sample feature sets, and generate a decision tree based on the sampled sample feature set;

[0030] a node generation module connected to the sampling module, the node generation module configured to generate a tree node based on the image sample features and / or the ultrasound sample features of the sample feature set for each of the decision trees;

[0031] a model combination module connected to the node generation module, the model combination module configured to generate the random forest model based on the decision trees;

[0032] a model iteration module connected to the sample feature generation module, the sampling module, the node generation module and the model combination module, respectively, the model iteration module configured to validate the random forest model based on a validation feature set corresponding to the sample feature set, and determine whether to output the current random forest model.

[0033] In another aspect, the prediction system further comprises a feature set screening module configured to select usable features from a plurality of candidate features in the magnetic resonance image;

[0034] the feature extraction module is configured to extract the image feature set from the magnetic resonance image according to the usable features;

[0035] the feature set screening module comprises:

[0036] a univariate analysis module, which receives the magnetic resonance image, extracts a plurality of the candidate features from the magnetic resonance image, and respectively generates a predictive evaluation result for each of the candidate features;

[0037] The univariate analysis module further screens the candidate features according to the predictive evaluation results to obtain a plurality of first screening features;

[0038] a variance inflation measurement module, which is connected to the univariate analysis module, respectively generates a variance inflation factor of each of the first screening features;

[0039] The variance inflation measurement module further screens the first screening features according to the variance inflation factors to obtain a plurality of second screening features;

[0040] a multivariate analysis module, which is connected to the variance inflation measurement module, generates the available features and a feature threshold value of each of the available features according to the second screening features.

[0041] The above technical solution has the following advantages or beneficial effects:

[0042] To solve the problem of poor measurement effect of related indicators of fetal growth restriction in the prior art, the prediction process is improved in the present solution, specific image features are extracted from the magnetic resonance image as an image feature group, which can achieve better representation effect of whether the placenta has placental dysfunction, avoids the influence of subjective factors caused by subjective judgment and evaluation of doctors in the prior art, and forms an ultrasound feature group by processing ultrasound data, combines the image feature group and the ultrasound feature group, and achieves better prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0043] Reference will now be made to the drawings to describe the embodiments of the application in more detail. However, the accompanying drawings are only used for illustration and explanation, and do not constitute a limitation on the scope of the application.

[0044] Figure 1 is a schematic diagram of the whole embodiment of the present application;

[0045] Figure 2 is a schematic diagram of the image feature extraction module in the embodiment of the present application;

[0046] Figure 3 is a schematic diagram of the feature generation module in the embodiment of the present application;

[0047] Figure 4 is a schematic diagram of the beat index calculation module in the embodiment of the present application;

[0048] Figure 5 Fig. 1 is a schematic diagram of a model training module in an embodiment of the present application;

[0049] Figure 6 Fig. 2 is a schematic diagram of a feature group screening module in an embodiment of the present application;

[0050] Figure 7 Fig. 3 is a schematic diagram of ROC curves of various models in a training set in an embodiment of the present application;

[0051] Figure 8 Fig. 4 is a schematic diagram of ROC curves of various models in a test set in an embodiment of the present application;

[0052] Figure 9 Fig. 5 is a schematic diagram of ROC curves of various models in a validation set in an embodiment of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0054] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0055] Some block diagrams shown in the drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0056] The present application will be further described below with reference to the drawings and specific embodiments, but is not limited by the embodiments.

[0057] The present application includes:

[0058] A prediction system for fetal growth restriction, as shown in Fig. 1, comprising: Figure 1 A collection module 1 for collecting magnetic resonance images and ultrasound data for a patient;

[0059] A feature extraction module 2 connected to the collection module 1, the feature extraction module 2 extracting an image feature group from the magnetic resonance images and an ultrasound feature group from the ultrasound data;

[0060]

[0061] ​A joint prediction module 3 is connected to the feature extraction module 2, and generates a prediction result according to the image feature group and the ultrasound feature group.

[0062] The image feature group includes placenta flatness, placenta area-volume ratio, and placenta region gray matrix features.

[0063] The ultrasound feature group includes fetal weight, abdominal circumference, and pulsatility index.

[0064] Specifically, to solve the problem that the accuracy of the doctor in diagnosing the fetal growth restriction disease is reduced in the subsequent diagnosis process due to the poor accuracy of the indicators measured from two dimensions of ultrasound and magnetic resonance images in the prior art, in the embodiment, the feature extraction module 2 is arranged to extract the magnetic resonance images and ultrasound data obtained by the acquisition module 1, so as to obtain an image feature group composed of placenta flatness, placenta area-volume ratio, and placenta region gray matrix features, and an ultrasound feature group composed of fetal weight, abdominal circumference, and pulsatility index. Since the extraction process of the image feature group is realized by the image processing algorithm in the feature extraction module 2, the evaluation effect is more accurate than that realized by the doctor in the prior art. Moreover, by selecting the image features as the indicators for prediction and combining them with the ultrasound feature group, the joint prediction module 3 can realize more accurate prediction of the fetal state.

[0065] In the implementation process, the above prediction system is arranged in a corresponding computer device, such as a hospital information system, as a software embodiment, for processing the input magnetic resonance images and ultrasound data and outputting the corresponding prediction results to assist the doctor in diagnosis. The acquisition module 1 is connected to external magnetic resonance imaging devices or equivalent devices and ultrasound scanning devices or equivalent devices in the form of a corresponding software interface, a data bus, etc. In an embodiment, the magnetic resonance device uses a conventional 1.5T Siemens magnetic resonance system (MAGNETOM Aera, Germany Siemens), which has 18 channels and performs placenta magnetic resonance imaging according to a standard scheme, and the imaging result is a T2WI image; the ultrasound scanning device is selected from GE Voluson E8, GE Voluson E6, and Philips HD11DX color Doppler ultrasound examination instruments, the frequency of the transducer is set to 3.5-5.0MHz, and is used to perform fetal biometry and Doppler blood flow velocity measurement. The feature extraction module 2 is provided with a corresponding processing algorithm, which can perform image processing, feature recognition, and extraction on the input magnetic resonance images according to a specific execution process, and process the signals in the ultrasound data in time-frequency or recognize and extract the ultrasound images reconstructed according to the ultrasound data to obtain the corresponding features.

[0066] In one embodiment, the feature extraction module 2 comprises an image feature extraction module 21, as shown in Figure 2 The image feature extraction module 21 comprises:

[0067] a wavelet filtering module 211, the wavelet filtering module 211 receives the magnetic resonance image, and the wavelet filtering module 211 generates a wavelet filtered image according to the magnetic resonance image;

[0068] a Gaussian Laplace filtering module 212, the Gaussian Laplace filtering module 212 is connected to the wavelet filtering module 211, and the Gaussian Laplace filtering module 212 processes the wavelet filtered image to obtain a Gaussian Laplace filtered image;

[0069] a feature generation module 213, the feature generation module 213 is connected to the wavelet filtering module 211 and the Gaussian Laplace filtering module 212, and the feature generation module 213 extracts image features from the wavelet filtered image, the Gaussian Laplace filtered image and the magnetic resonance image respectively;

[0070] a feature calculation module 214, the feature calculation module 214 generates an image feature group according to the image features.

[0071] Specifically, in order to achieve better image feature extraction effect, in the embodiment, the image feature extraction module 21 is arranged in the feature extraction module 2 to achieve better processing effect of the magnetic resonance image. Among them, the image feature extraction module 21 comprises a wavelet filtering module 211 and a Gaussian Laplace filtering module 212, which processes the original magnetic resonance image through a wavelet filter and a Gaussian Laplace filter to obtain a derived image based on the magnetic resonance image. Specifically, the wavelet filtering module 211 decomposes the magnetic resonance image through the wavelet filter, and calculates eight different wavelets to obtain wavelet features, which enriches the image features of the placenta at multiple scales; the Gaussian Laplace filtering module 212 processes the magnetic resonance image through the Gaussian-Laplace filtering method, which enhances the edge information, so that the subsequent imaging model can achieve better edge segmentation effect. Then, the feature generation module 213 extracts image features from the derived image and the magnetic resonance image respectively, including placenta shape, first-order statistics and placenta texture, so that the feature calculation module 214 can accurately calculate the image feature group.

[0072] Further, in order to achieve better feature extraction effect, in the embodiment, the feature generation module 213 further introduces the derived image, including the wavelet filtered image and the Gaussian Laplace filtered image, on the basis of the magnetic resonance image in the process of calculating the placenta texture and the first-order statistics, which achieves better extraction effect of these two image features.

[0073] In one embodiment, as shown in Figure 3As shown, the feature generation module 213 includes:

[0074] a shape extraction module 2131 that extracts a placenta shape from the derived image and the magnetic resonance image and adds the placenta shape to the image features;

[0075] a first-order statistics extraction module 2132 that extracts first-order statistics features from the derived image and the magnetic resonance image and adds the first-order statistics features to the image features;

[0076] a texture feature extraction module 2133 that extracts placenta texture from the derived image and the magnetic resonance image and adds the placenta texture to the image features.

[0077] In particular, to achieve better evaluation of each placenta index in the magnetic resonance image, in this embodiment, the shape extraction module 2131 is arranged in the feature generation module 213 to extract the placenta shape, the first-order statistics extraction module 2132 is arranged to extract the gray histogram features of the derived image and the magnetic resonance image, and the texture feature extraction module 2133 is arranged to extract the placenta texture of the derived image and the magnetic resonance image, so that each index of the placenta in the image is better selected, facilitating the subsequent index evaluation process.

[0078] In the implementation process, the above modules can be implemented by using image processing algorithms. For example, in the shape extraction module 2131, the edge part of the placenta region and the background image can be recognized by using an image segmentation model, so as to extract the placenta region as the shape feature. In the first-order statistics extraction module 2132, the first-order statistics features can be easily extracted by converting the image into a gray histogram.

[0079] In one embodiment, the feature extraction module 2 includes a pulsatility index calculation module 22, as shown in Figure 4 As shown, the pulsatility index calculation module 22 includes:

[0080] a first flow rate generation module 221 that generates a systolic flow rate and a diastolic flow rate according to the ultrasound data;

[0081] a second flow rate generation module 222 that generates an average flow rate according to the ultrasound data;

[0082] a pulsatility index generation module 223 that is connected to the first flow rate generation module 221 and the second flow rate generation module 222, respectively, and generates a pulsatility index according to the systolic flow rate, the diastolic flow rate, and the average flow rate.

[0083] Specifically, to achieve a better calculation result of pulsatility index, in the embodiment, a pulsatility index calculation module 22 is further arranged in the feature extraction module 2, which comprises a first flow velocity generation module 221, a second flow velocity generation module 222 and a pulsatility index generation module 223. The first flow velocity generation module 221 calculates the systolic flow velocity and diastolic flow velocity by performing spectral Doppler calculation on the echo signals in the ultrasound data. The second flow velocity generation module 222 obtains the curve envelope of a cardiac cycle by intercepting a predetermined length of echo signal sequence in the ultrasound data, and then integrates the curve envelope to obtain the average flow velocity. Based on the flow velocity data calculated above, the pulsatility index generation module 223 easily calculates the ratio of the difference between the systolic flow velocity and the diastolic flow velocity to the average flow velocity as the pulsatility index.

[0084] In one embodiment, the joint prediction module 3 generates the prediction result according to the image feature group and the ultrasound feature group by using a random forest algorithm.

[0085] The prediction system further comprises a model training module 4, as shown in Figure 5 The model training module 4 comprises:

[0086] a sample feature generation module 41, which receives sample data and extracts image sample features corresponding to the image feature group and ultrasound sample features corresponding to the ultrasound feature group from the sample data.

[0087] and combines to form a sample feature group;

[0088] a sampling module 42, which is connected to the sample feature generation module 41, samples from the multiple sample feature groups, and generates a decision tree according to the sampled sample feature group;

[0089] a node generation module 43, which is connected to the sampling module 42, and generates a tree node according to the image sample features and / or the ultrasound sample features of the sample feature group for each decision tree;

[0090] a model combination module 44, which is connected to the node generation module 43, and generates a random forest model according to the decision tree;

[0091] a model iteration module 45, which is connected to the sample feature generation module 41, the sampling module 42, the node generation module 43 and the model combination module 44 respectively, verifies the random forest model according to a verification feature group corresponding to the sample feature group, and determines whether to output the current random forest model.

[0092] Specifically, in order to achieve better prediction effect on the extracted features, in the embodiment, a random forest model is selected to predict the image feature group extracted based on the magnetic resonance image and the ultrasound feature group generated based on the ultrasound data, so as to achieve better prediction effect. In order to achieve better prediction effect, a model training module 4 is constructed in the prediction system for training. Specifically, for the random forest model, sample data corresponding to the actual patient is collected in advance. The sample data includes magnetic resonance images, ultrasound data and the diagnosis result of the patient. Subsequently, the sample feature generation module 41 extracts corresponding features from the magnetic resonance images and the ultrasound data as image sample features and ultrasound sample features respectively, and combines to form a sample feature group corresponding to the patient. Subsequently, the sample feature group is extracted with replacement by the sampling module 42, one at a time, to finally form a certain number of extracted sample feature groups for constructing decision trees. Repeat the process to obtain multiple decision trees containing different extracted sample feature groups. For each decision tree, the node generation module 43 randomly selects a number of features from the image sample features and / or ultrasound sample features as tree nodes, so that different extracted sample feature groups are classified in the decision tree, thereby realizing the construction of the decision tree, and the model combination module 44 finally combines all the decision trees as a random forest model. After assembling the random forest model, the model iteration module 45 also verifies the random forest model based on the ten-fold cross-validation method using the verification feature group to determine whether the current random forest model meets the iteration condition. If not, the model is adjusted and re-constructed to obtain an iterated model, so that the finally output model has better prediction effect.

[0093] In one embodiment, the prediction system further comprises a feature group screening module 5 for selecting available features from a plurality of alternative features in the magnetic resonance image;

[0094] As shown in Figure 6 , the feature group screening module 5 comprises:

[0095] The univariate analysis module 51 receives the magnetic resonance image, the univariate analysis module 51 extracts a plurality of alternative features from the magnetic resonance image, and generates a predictive evaluation result for each alternative feature respectively;

[0096] The univariate analysis module 51 also screens the alternative features according to the predictive evaluation result to obtain a plurality of first screening features;

[0097] The variance inflation measurement module 52 is connected to the univariate analysis module 51, and the variance inflation measurement module 52 generates a variance inflation factor for each first screening feature respectively;

[0098] The variance inflation measurement module 52 further screens the first screening features according to a variance inflation factor to obtain a plurality of second screening features;

[0099] The multivariate analysis module 53 is connected to the variance inflation measurement module 52, and generates available features and feature thresholds of each available feature according to the second screening features.

[0100] Specifically, in order to achieve better prediction effect of placenta condition, in the embodiment, after processing the magnetic resonance image and constructing a plurality of candidate features based on shape features, first-order statistical features, texture features and the like of the placenta region image, the predictability of each candidate feature is analyzed based on the logistic regression method in the univariate analysis module 51, thereby eliminating features with low prediction accuracy and retaining the features as first screening features. Subsequently, the variance inflation measurement module 52 quantizes the collusive relationship between each first screening feature based on the variance inflation factor (VIF), and iteratively eliminates redundant features. In each iteration, the variance inflation factor of each first screening feature is calculated, and the first screening feature with a high VIF value is removed to avoid repeated selection of features, thereby obtaining second screening features. Finally, through the multivariate analysis module 53, the correlation and predictability of each second screening feature are measured based on random forest multivariate analysis, and a corresponding quantitative score is calculated. Based on the quantitative score, the more important second screening features are selected as the image feature group, and the parameters and thresholds of each available feature are determined through grid search, which are used to distinguish the attributes of each node in the random forest model training, thereby achieving a good image feature group generation effect.

[0101] In order to verify the effectiveness of the above technical solutions, in one embodiment, the prediction system constructed is verified through the following process:

[0102] The clinical data of pregnant women who delivered and underwent placenta MRI examination in a hospital from January 2015 to December 2021 is extracted, including magnetic resonance images, ultrasound data and clinical diagnosis conclusions, and the data is further screened into an FGR group and a control group.

[0103] The standard of the FGR group is:

[0104] 1. Single pregnancy; 2. Estimated fetal weight is lower than the tenth percentile of the corresponding gestational age; 3. Ultrasonic examination shows abnormal umbilical artery blood flow velocity; 4. The fetus has no congenital malformation; 5. The actual birth weight of the fetus is lower than the tenth percentile of the corresponding gestational age;

[0105] The standard of the control group is:

[0106] 1. Singleton pregnancy; 2. Full-term delivery of a newborn; 3. No abnormal complications during pregnancy; 4. Complete clinical data.

[0107] Based on the above-mentioned standard, 202 samples aged between 17-40 years old were screened, which were randomly divided into 119 training sets, 40 test sets and 43 validation sets, and the prenatal clinical characteristics of the samples are shown in Table 1:

[0108]

[0109] Table 1: Prenatal clinical characteristics of sample data

[0110] In order to achieve better screening effect of image feature group, in this verification process, based on the above-mentioned image processing method, the shape, intensity and texture features of each magnetic resonance image were extracted, and 960 image features were obtained as candidate features. Subsequently, the prediction accuracy of each candidate feature was obtained after univariate analysis. The candidate features with a prediction accuracy greater than 0.6 were output as the first screening features. Subsequently, a large number of redundant features with a VIF value greater than 10 were removed from the first screening features, and 32 non-redundant second screening features were obtained. Finally, the correlation and predictability of the second screening features were analyzed and valued based on the random forest algorithm, and three features with an importance greater than 0.02 were obtained as the features in the image feature group, i.e. placenta flatness, placenta area volume ratio and placenta region gray matrix feature.

[0111] The placenta flatness refers to the flatness of the placenta in the original image, and this feature belongs to the shape feature of the placenta. This feature describes the flatness of the placenta, i.e. the micro shape feature of the placenta surface. Based on the analysis results obtained by the above-mentioned analysis process, it can be known that there is a significant difference between the placenta flatness of the FGR group and the control group, and therefore it can be added to the image feature group;

[0112] The placenta area volume ratio refers to the area volume ratio of the placenta in the original image, and this feature belongs to the shape feature of the placenta. This feature describes the overall shape of the placenta, i.e. the macro (overall) flatness of the placenta. Based on the analysis results obtained by the above-mentioned analysis process, it can be known that there is a significant difference between the placenta flatness of the FGR group and the control group, and therefore it can be added to the image feature group;

[0113] The placenta region gray matrix feature refers to a gray size region matrix feature of the placenta in the Gaussian-Laplacian filter image, and this feature belongs to the texture feature of the placenta. This feature describes the gray level variation of the placenta in the Gaussian-Laplacian filter image. Based on the analysis results obtained by the above-mentioned analysis process, it can be known that there is a significant difference between the placenta gray level similarity of the FGR group and the normal pregnant women in the Gaussian-Laplacian filter image, and therefore it can be added to the image feature group.

[0114] The important and highly associated imaging features were obtained through the above analysis process to construct the image feature group, and combined with the ultrasound feature group, including fetal weight (EFW), abdominal circumference (AC) and pulsatility index (PI) to construct the joint prediction model. To verify the effectiveness of the model, the above ultrasound feature group was further combined with the logistic regression (LR) model, and the additional model based on 23 shape and texture features in other studies (Dahdouh S, Andescavage N, Yewale S, et al. In vivo placental MRI shape and textural features predict fetal growth restriction and postnatal outcome. J Magn Reson Imaging. Feb 2018; 47(2): 449-458. doi: 10.1002 / jmri.25806, hereinafter 5806 literature) were used as control models. The ROC curves of each model are shown in Figure 4 The performance of each model in the training set, test set and validation set is shown in Table 2.

[0115]

[0116]

[0117] Table 2: Training data of each model

[0118] Table 2 and Figure 7 , Figure 8 , Figure 9 In the above table 2 and Figure 7 , Figure 8 , Figure 9 In the above table 2 and

[0119] Based on the above data, the AUC of the joint model in the test set and the validation set is 0.87 (95% CI: 0.74-0.96) and 0.87 (95% CI: 0.76-0.97), respectively, which is higher than the first prediction model and the second prediction model using only ultrasound indicators. In addition, the first prediction model constructed only based on fetal weight has an AUC of 0.76 (95% CI: 0.56-0.93) and 0.76 (95% CI: 0.60-0.89) in the test set and the validation set, respectively. By comparing the first prediction model and the second prediction model, it can be seen that the second prediction model based on the above three ultrasound indicators shows better performance, with an AUC of 0.82 (95% CI: 0.65-0.94) and 0.82 (95% CI: 0.67-0.93) in the test set and the validation set, respectively. In addition, the third prediction model based on 5806 literature shows comparable performance to the joint model in the test set, indicating that imaging features can effectively diagnose FRG, with an AUC of 0.87 (95% CI: 0.75-0.97); however, the performance of the third prediction model in the validation set is significantly reduced, with an AUC of 0.70 (95% CI: 0.48-0.90), which is significantly worse than the joint model.

[0120] Based on the above comparison results, it can be seen that the joint prediction model based on the image feature group and the ultrasound feature group described in the present application can achieve better prediction effect.

[0121] The above is only the preferred embodiment of the present application, and does not limit the implementation and protection scope of the present application. For those skilled in the art, it should be realized that any equivalent replacement and obvious changes made according to the content of the present application should be included in the protection scope of the present application.

Claims

1. A system for predicting fetal growth restriction, characterized in that, The system comprises: a collection module, which collects magnetic resonance images and ultrasound data for a patient; a feature extraction module connected to the collection module, which extracts an image feature group from the magnetic resonance images and an ultrasound feature group from the ultrasound data; a joint prediction module connected to the feature extraction module, which generates a prediction result according to the image feature group and the ultrasound feature group; the image feature group comprises placenta flatness, placenta area-volume ratio, and placenta region gray matrix features; the ultrasound feature group comprises fetal weight, abdominal circumference, and umbilical artery pulsatility index; the feature extraction module comprises an image feature extraction module, which comprises: a wavelet filtering module, which receives the magnetic resonance images, and generates wavelet filtered images according to the magnetic resonance images; a Gaussian Laplacian filtering module connected to the wavelet filtering module, which processes the wavelet filtered images to obtain Gaussian Laplacian filtered images; a feature generation module connected to the wavelet filtering module and the Gaussian Laplacian filtering module, which extracts image features from the wavelet filtered images, the Gaussian Laplacian filtered images, and the magnetic resonance images, respectively; a feature calculation module, which generates the image feature group according to the image features; the feature generation module comprises: a shape extraction module, which extracts placenta shape from derived images and the magnetic resonance images and adds the placenta shape to the image features; the derived images comprise the wavelet filtered images and the Gaussian Laplacian filtered images; a first-order statistics extraction module, which extracts first-order statistics features from the derived images and the magnetic resonance images and adds the first-order statistics features to the image features; a texture feature extraction module, which extracts placenta texture from the derived images and the magnetic resonance images and adds the placenta texture to the image features.

2. The prediction system of claim 1, wherein, The feature extraction module comprises a pulsatility index calculation module, which comprises: a first flow rate generation module, which generates systolic flow rate and diastolic flow rate according to the ultrasound data; a second flow rate generation module, which generates average flow rate according to the ultrasound data; a pulsatility index generation module connected to the first flow rate generation module and the second flow rate generation module, respectively, which generates the pulsatility index according to the systolic flow rate, the diastolic flow rate, and the average flow rate.

3. The prediction system of claim 1, wherein, The joint prediction module generates the prediction result according to the image feature group and the ultrasound feature group using a random forest model; the prediction system further comprises a model training module, which comprises: a sample feature generation module configured to receive sample data and extract image sample features corresponding to the image feature set and ultrasound sample features corresponding to the ultrasound feature set from the sample data, and combine the image sample features and the ultrasound sample features to form a sample feature set; a sampling module connected to the sample feature generation module, the sampling module configured to sample from a plurality of the sample feature sets, and generate a decision tree based on the sampled sample feature set; a node generation module connected to the sampling module, the node generation module configured to generate a tree node based on the image sample features and / or the ultrasound sample features of the sample feature set for each of the decision trees; a model combination module connected to the node generation module, the model combination module configured to generate the random forest model based on the decision trees; a model iteration module connected to the sample feature generation module, the sampling module, the node generation module, and the model combination module, respectively, the model iteration module configured to validate the random forest model based on a validation feature set corresponding to the sample feature set, and determine whether to output the current random forest model.

4. The prediction system of claim 1, wherein, The prediction system further comprises a feature set screening module configured to select usable features from a plurality of candidate features in the magnetic resonance image; the feature extraction module is configured to extract the image feature set from the magnetic resonance image according to the usable features; the feature set screening module comprises: a univariate analysis module configured to receive the magnetic resonance image, extract a plurality of the candidate features from the magnetic resonance image, and generate a predictive evaluation result for each of the candidate features, respectively; the univariate analysis module is further configured to screen the candidate features according to the predictive evaluation results to obtain a plurality of first screening features; a variance inflation measurement module connected to the univariate analysis module, the variance inflation measurement module configured to generate a variance inflation factor for each of the first screening features, respectively; the variance inflation measurement module is further configured to screen the first screening features according to the variance inflation factors to obtain a plurality of second screening features; a multivariate analysis module connected to the variance inflation measurement module, the multivariate analysis module configured to generate the usable features and a feature threshold value for each of the usable features based on the second screening features.

Images