Immune globulin A nephropathy risk prediction method and system based on ultrasound

Abstract

The invention relates to an immunoglobulin A nephropathy risk prediction method and system based on ultrasound, and the method comprises the steps: obtaining a kidney longest axis ultrasound image of a sample patient, determining a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generating a boundary contour; a target area image in the kidney longest axis ultrasonic image is preprocessed, and image omics feature extraction processing is carried out; dimensionality reduction and screening are carried out on the extracted radiomics features to obtain target radiomics features, radiomics scores are calculated, and the radiomics scores are combined with preset independent risk variables to obtain combined features; constructing a prediction model and performing training; and inputting the combined features corresponding to the target patient into the trained prediction model to obtain a corresponding risk type prediction result. The method realizes accurate prediction of immunoglobulin A nephropathy, can be used in clinical scenes where renal biopsy cannot be performed or provides information guidance for replacement of renal biopsy, and avoids delay of optimal treatment opportunity.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic image recognition, and in particular to an ultrasound-based immunoglobulin A nephropathy risk prediction method and system. Background Art

[0002] Immunoglobulin A nephropathy (IgAN) is the most common primary glomerular disease worldwide and one of the most common glomerular diseases in children and adolescents. Unlike the variable course of the disease discovered in adulthood, IgAN often presents as a benign disease in childhood, with clinical manifestations of rapid disappearance of paroxysmal macroscopic hematuria or persistent mild microscopic hematuria. In many mild cases, some only show asymptomatic microscopic hematuria with persistent microproteinuria, without any other clinical manifestations. It is difficult to detect directly through clinical characteristics, and previous studies have shown that the prognosis of children with IgAN is not optimistic. Therefore, early detection is crucial for the management and prognosis of childhood IgAN disease.

[0003] In the clinical treatment of IgAN, renal biopsy is the gold standard. However, as an invasive examination, renal biopsy may cause various complications such as perinephric hematoma, pain and infection. In addition, compared with adults, children are more likely to have fear and rejection of renal puncture biopsy. In addition, the conservative ideas of some parents and their overprotection of their children may also lead to the inability to perform renal biopsy on children and obtain pathology in time, which hinders the early detection of patients and the best treatment time, affecting the long-term prognosis of children.

[0004] Therefore, there is an urgent need for an effective method to help prevent IgAN at an early stage, which can be applied in situations where renal biopsy cannot be performed or provide information guidance for alternative renal biopsy. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an ultrasound-based immunoglobulin A nephropathy risk prediction method and system in view of the deficiencies of the above-mentioned prior art.

[0006] The technical solution of the present invention to solve the above technical problem is as follows: A method for predicting the risk of immunoglobulin A nephropathy based on ultrasound, comprising the following steps:

[0007] Acquire a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region;

[0008] Preprocessing the target area image in the longest axis ultrasound image of the kidney, and performing radiomics feature extraction processing on the preprocessed target area image;

[0009] Performing dimensionality reduction and screening on the extracted radiomics features to obtain target radiomics features, and calculating radiomics scores, and combining the radiomics scores of the target radiomics features with preset independent risk variables to obtain combined features;

[0010] Building a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types;

[0011] The combined features corresponding to the target patient are input into the trained prediction model to obtain the corresponding risk type prediction result.

[0012] On the basis of the above technical solution, the present invention can also be improved as follows:

[0013] Further: the step of obtaining the longest axis ultrasound image of the kidney of the sample patient, determining the kidney tissue area in the longest axis ultrasound image of the kidney as the target area, and generating the boundary contour of the target area specifically comprises the following steps:

[0014] Obtain the kidney ultrasound image in DICOM format and perform analytical conversion to obtain the kidney longest axis ultrasound image in NIFTI format;

[0015] Image recognition processing is performed on the kidney longest axis ultrasound image to determine that the kidney tissue area in the kidney longest axis ultrasound image is the target area, and a boundary contour of the target area is generated.

[0016] The beneficial effect of the above further scheme is that by analyzing and transforming the kidney ultrasound image, the longest axis ultrasound image of the kidney can be accurately obtained, and then the target area of ​​the kidney tissue and the corresponding boundary contour can be obtained according to the image recognition processing, which is more convenient for subsequent preprocessing and feature extraction processing of the target area.

[0017] Further: the preprocessing of the target area image in the longest axis ultrasound image of the kidney and the imaging omics feature extraction processing of the preprocessed target area image specifically include the following steps:

[0018] Normalizing the target area image in the longest axis ultrasound image of the kidney according to a preset normalization ratio;

[0019] Resampling the normalized target region image in image space to obtain a target region image with a uniform voxel size;

[0020] A plurality of filters are used to perform image enhancement processing on the target area image, and radiomics feature extraction processing is performed on the enhanced target area image to obtain radiomics features, wherein the image enhancement processing includes filter enhancement image processing and mathematical transformation enhancement image processing.

[0021] The beneficial effect of the above further scheme is: by normalizing and resampling the target area image in the longest axis ultrasound image of the kidney, a target area image of a pixel size can be obtained, and by performing image enhancement processing on the target area image, it helps to emphasize different features and textures in the image, so as to extract image features.

[0022] Further: the dimensionality reduction and screening of the extracted imaging omics features specifically include the following steps:

[0023] The extracted radiomics features were evaluated for intra-group consistency and the intra-group correlation coefficient was calculated;

[0024] The radiomics features whose intragroup correlation coefficient is not less than a preset threshold are converted into a standard normal distribution, and data dimension reduction and screening are performed to obtain target radiomics features, and the radiomics score is calculated.

[0025] The beneficial effect of the above further scheme is: by performing intra-group consistency evaluation on the radiomics features, the radiomics features with correlation coefficients not less than a preset threshold can be retained to obtain candidate radiomics features, and then dimensionality reduction and screening are performed to finally obtain the target radiomics features, and the radiomics score is calculated to facilitate the subsequent determination of the combined features.

[0026] Further: performing data dimension reduction and screening on the radiomics features converted into a standard normal distribution to obtain target radiomics features, and calculating the radiomics score specifically includes the following steps:

[0027] Performing statistical analysis on the imaging features using a t-test, and only retaining the imaging features with a p-value less than a preset p-threshold;

[0028] Using the preset Pearson correlation coefficient, any two imaging features with correlation coefficients greater than the preset correlation threshold are randomly removed to obtain the target imaging features;

[0029] The least absolute convergence and selection operator (LASSO) regression model was constructed, and the lambda parameter with the minimum mean square error (MSE) was calculated through cross-validation of the minimum criterion, and finally the non-zero coefficient characteristics in the LASSO model were obtained;

[0030] Based on the LASSO model, the radiomics score was calculated using the following formula:

[0031] Radiomics Score = a 1 *F 1 +a 2 *F 2 +…+a i *F i +b;

[0032] Among them, a 1 to a i is the coefficient of the target radiomics feature, F 1 To F i is the target imaging feature; the number of target imaging features is i; and b is the intercept.

[0033] The beneficial effect of the above further scheme is that the dimensionality reduction can be performed by statistically analyzing the radiomics features through the t-test, and redundant features can be eliminated through the Pearson correlation coefficient, and the radiomics score is calculated by the least absolute convergence and selection operator (LASSO) regression model, so as to facilitate the combination with the preset independent risk variables to obtain the combined features.

[0034] Further: the step of constructing a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types specifically includes the following steps:

[0035] The imaging omics features and combined features are respectively input into a plurality of preset machine learning algorithms, and the corresponding area under the ROC curve AUC index is calculated;

[0036] Taking the machine learning algorithm with the largest AUC index as the target machine learning algorithm, and building a prediction model based on the target machine learning algorithm;

[0037] The combined features and the corresponding risk types are input into the trained prediction model for training.

[0038] The beneficial effect of the above further scheme is: by respectively inputting the imaging genomics features and the combined features into a plurality of preset machine learning algorithms, the best machine learning algorithm can be screened out by combining the area under the ROC curve AUC indicator, and then the combined features and the corresponding risk types are input into the trained prediction model for training, and the parameters of the prediction model are determined, so that the trained prediction model has recognition ability.

[0039] Further: the method further comprises the following steps:

[0040] A nomogram is constructed according to the combined characteristics of the target patient, and the risk probability corresponding to the risk type prediction result is determined according to the nomogram.

[0041] The beneficial effect of the above further scheme is: by constructing a nomogram, the classification result can be visualized, so as to determine the risk probability of the risk type prediction result of the combined feature according to the nomogram, and interpret the classification result according to the risk description corresponding to the risk probability.

[0042] The present invention also provides an ultrasound-based immunoglobulin A nephropathy risk prediction system, comprising an image recognition module, a feature extraction module, a combined feature module, a construction training module and a risk prediction module;

[0043] The image recognition module is used to obtain a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region;

[0044] The feature extraction module is used to preprocess the target area image in the longest axis ultrasound image of the kidney, and perform radiomics feature extraction processing on the preprocessed target area image;

[0045] The combined feature module is used to reduce the dimension and screen the extracted radiomics features to obtain target radiomics features, calculate the radiomics score, and combine the radiomics score of the target radiomics feature with a preset independent risk variable to obtain a combined feature;

[0046] The training module is used to construct a prediction model based on the radiomics features and the combined features, and train the prediction model based on the combined features and the corresponding risk types;

[0047] The risk prediction module is used to input the combined features corresponding to the target patient into the trained prediction model to obtain the corresponding risk type prediction result.

[0048] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the ultrasound-based immunoglobulin A nephropathy risk prediction method is implemented.

[0049] The present invention also provides an ultrasound-based immunoglobulin A nephropathy risk prediction device, comprising a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0050] The memory is used to store computer programs;

[0051] The processor is used to implement the steps of the ultrasound-based immunoglobulin A nephropathy risk prediction method when executing the program stored in the memory.

[0052] The beneficial effects of the present invention are as follows: the ultrasound-based immunoglobulin A nephropathy risk prediction method and system of the present invention determines the boundary contour of the target area in the longest axis ultrasound image of the kidney of the sample patient, and then performs feature extraction processing after preprocessing, and obtains the target imaging genomics features and imaging genomics scores based on the extracted imaging genomics features through dimensionality reduction screening, thereby obtaining combined features, and then constructs a prediction model based on the imaging genomics features and the combined features and performs training. The trained prediction model can accurately predict the risk type of the combined features corresponding to the target patient, thereby achieving accurate prediction of immunoglobulin A nephropathy, which can be used in clinical scenarios where renal biopsy cannot be performed or to provide information guidance for alternative renal biopsy, thereby helping to achieve early prediction of patients, avoid delays in the best treatment opportunity, and improve the long-term prognosis of children. This model has a broad application prospect in the field of immunoglobulin A nephropathy prediction in children. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of a process of predicting the risk of immunoglobulin A nephropathy based on ultrasound according to an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of target area identification in a kidney longest axis ultrasound image according to an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of the AUC of the training results and verification results of different machine learning algorithm models provided for the prediction model based on combined feature training according to one embodiment of the present invention;

[0056] Figure 4 A schematic diagram of the AUC of the training results and validation results of the prediction model based on combined feature training, independent risk factors and target radiomics feature training model according to an embodiment of the present invention;

[0057] Figure 5a and Figure 5b They are schematic diagrams of a calibration curve and a clinical decision curve of a prediction model based on combined feature training according to an embodiment of the present invention;

[0058] Figure 6 A schematic diagram of a nomogram of a prediction model based on combined feature training according to an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of the structure of an ultrasound-based immunoglobulin A nephropathy risk prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0061] The present invention provides an ultrasound-based immunoglobulin A nephropathy risk prediction method, which obtains the quantitative features of ultrasound images of the long axis of the kidney through radiomics technology to avoid missing useful image information, thereby improving the accuracy of immunoglobulin A nephropathy risk prediction; then, risk prediction is performed in combination with the radiomics features of the ultrasound image and independent risk factors. Since the existing risk prediction classification results are based on end-to-end results, either there is a risk of disease or there is no risk of disease, its interpretability is poor. Therefore, after risk prediction, the present invention uses a nomogram to obtain the risk probability, and interprets the classification result according to the risk description corresponding to the risk probability, that is, explains why there is a risk of disease or why there is no risk of disease, thereby improving the interpretability of immunoglobulin A nephropathy risk prediction.

[0062] In order to better illustrate the technical solution of the present invention, it will be described in detail from the following embodiments.

[0063] like Figure 1 As shown, a method for predicting the risk of immunoglobulin A nephropathy based on ultrasound comprises the following steps:

[0064] S1: Acquire a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region;

[0065] Here, renal ultrasound examination was performed according to a standardized process, and ultrasound images of the long axis of the kidneys of multiple patients were obtained. The specific procedures were as follows:

[0066] In some embodiments, children with immunoglobulin A nephropathy and normal children who underwent routine ultrasound and renal biopsy in a hospital from June 2017 to March 2023 were retrospectively collected, including 101 children with immunoglobulin A nephropathy and 167 normal children, who were randomly divided into training set and test set in a ratio of 7:3.

[0067] Inclusion criteria included: (1) age <18 years; (1) immunoglobulin A nephropathy confirmed by renal puncture biopsy. Exclusion criteria: (1) age >18 years; (2) low quality of ultrasound images used for feature extraction; (3) interval between ultrasound examination and puncture biopsy was more than 2 weeks; (4) clinical information was missing; (5) secondary glomerulonephritis such as purpura nephritis;

[0068] In some embodiments, a multi-frequency (2-5 MHz) convex array probe (C5-2) is used to perform renal ultrasound examination and obtain ultrasound images, including:

[0069] All children were evaluated in the supine position after an overnight fast. The ultrasound probe was gently positioned over the right abdomen, and coronal ultrasound images of the right kidney were obtained in the maximum cross-sectional plane. All measurements were performed during end-inspiratory apnea. The parameter configuration during the acquisition process was based on the optimal display settings of the ultrasound images, and the acquired ultrasound images were extracted based on the DICOM format.

[0070] S2: preprocessing the target area image in the longest axis ultrasound image of the kidney, and performing radiomics feature extraction processing on the preprocessed target area image;

[0071] S3: performing dimensionality reduction and screening on the extracted radiomics features to obtain target radiomics features, and calculating radiomics scores, and combining the radiomics scores of the target radiomics features with preset independent risk variables to obtain combined features;

[0072] S4: constructing a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types;

[0073] S5: Input the combined features corresponding to the target patient into the trained prediction model to obtain the corresponding risk type prediction result.

[0074] The ultrasound-based immunoglobulin A nephropathy risk prediction method of the present invention determines the boundary contour of the target area in the longest axis ultrasound image of the kidney of the sample patient, and then performs feature extraction processing after preprocessing, and obtains the target imaging genomics features and imaging genomics scores based on the extracted imaging genomics features through dimensionality reduction screening, thereby obtaining combined features, and then constructs a prediction model based on the imaging genomics features and the combined features and trains them. The trained prediction model can accurately predict the risk type of the combined features corresponding to the target patient, thereby achieving accurate prediction of immunoglobulin A nephropathy, and can be used in clinical scenarios where renal biopsy cannot be performed or to provide information guidance for alternative renal biopsy, which is helpful to achieve early prediction of patients, avoid delays in the best treatment time, and improve the long-term prognosis of children. This model has a broad application prospect in the field of immunoglobulin A nephropathy prediction in children.

[0075] In one or more embodiments of the present invention, the steps of obtaining the longest axis ultrasound image of the kidney of the sample patient, determining the kidney tissue area in the longest axis ultrasound image of the kidney as the target area, and generating the boundary contour of the target area specifically include the following steps:

[0076] S11: Obtain a kidney ultrasound image in DICOM format and perform analysis and conversion to obtain a kidney longest axis ultrasound image in NIFTI format;

[0077] S12: Perform image recognition processing on the longest axis ultrasound image of the kidney, determine the kidney tissue area in the longest axis ultrasound image of the kidney as the target area, and generate the boundary contour of the target area. Figure 2 FIG. 1 is a schematic diagram of identifying a target area provided by an embodiment of the present invention.

[0078] In one or more embodiments of the present invention, the preprocessing of the target area image in the longest axis ultrasound image of the kidney and the radiomics feature extraction processing of the preprocessed target area image specifically include the following steps:

[0079] S21: normalizing the target area image in the longest axis ultrasound image of the kidney according to a preset normalization ratio;

[0080] S22: resampling the normalized target region image in image space to obtain a target region image with a uniform voxel size;

[0081] S23: performing image enhancement processing on the target area image using a plurality of filters, and performing radiomics feature extraction processing on the enhanced target area image to obtain radiomics features, wherein the image enhancement processing includes filter enhancement image processing and mathematical transformation enhancement image processing.

[0082] In practice, in some embodiments, the image preprocessing and image enhancement are performed on the ultrasound image of the target area with the boundary contour determined based on the Python open source Pyradiomics package, including:

[0083] Comprehensive image preprocessing mainly includes image normalization to a scale of 100 and image spatial resampling to a uniform voxel size (3mm x 3mm x 3mm). In the present invention, multiple filters are used to enhance the image, including Laplacian of Gaussian (LoG) filter, wavelet transform (Wavelet), local binary pattern (LBP3D), and mathematical transformations such as exponential, square, square root and logarithmic, so as to help emphasize different features and textures in the image for image feature extraction.

[0084] Here, it is worth mentioning that radiomics refers to the high-throughput extraction of radiomics features from medical images through computer image processing technology based on medical image analysis, including: first-order features, texture features and shape features. Among them, texture features include gray-level co-occurrence matrix (GLCM), gray-level dependency matrix (GLDM), gray-level co-occurrence matrix (GLRLM), gray-level size zone matrix (GLSZM) and neighborhood gray-level difference matrix (NGTDM). A total of 1,686 radiomics features were extracted, and all feature definitions were in line with the Image Biomarker Standardization Initiative (IBSI).

[0085] In one or more embodiments of the present invention, the dimensionality reduction and screening of the extracted radiomics features specifically include the following steps:

[0086] S31: Evaluate the intra-group consistency of the extracted radiomics features and calculate the intra-group correlation coefficient;

[0087] S32: converting the radiomics features whose intra-group correlation coefficient is not less than a preset threshold into a standard normal distribution, performing data dimension reduction and screening to obtain target radiomics features, and calculating a radiomics score.

[0088] In practice, two image recognition processes are performed on the same image and radiomics features are extracted to obtain two groups of extraction results. The intraclass correlation coefficient (ICC) is used to evaluate the intraclass consistency of the two groups of extraction results, and the radiomics features with an intraclass correlation coefficient not less than the preset threshold (0.75) are retained to obtain candidate radiomics features.

[0089] By performing intra-group consistency evaluation on the radiomics features, the radiomics features with correlation coefficients not less than a preset threshold can be retained to obtain candidate radiomics features, which are then reduced in dimension and screened to ultimately obtain the target radiomics features. The radiomics scores are then calculated to facilitate the subsequent determination of combined features.

[0090] In one or more embodiments of the present invention, the radiomics features converted into a standard normal distribution are subjected to data dimensionality reduction and screening to obtain target radiomics features, and the radiomics score is calculated, which specifically includes the following steps:

[0091] S321: Performing statistical analysis on the imaging genomics features using a t-test, and only retaining the imaging genomics features whose p-values ​​are less than a preset p-threshold; in this embodiment, the preset p-threshold is set at 0.75;

[0092] S322: using a preset Pearson correlation coefficient, randomly removing any two radiomics features whose correlation coefficients are greater than a preset correlation threshold, to obtain a target radiomics feature; in this embodiment, the preset correlation threshold is set at 0.9;

[0093] S323: construct a least absolute convergence and selection operator (LASSO) regression model, calculate the lambda parameter with the minimum mean square error (MSE) through 10-fold cross validation of the minimum criterion, and finally obtain the non-zero coefficient characteristics in the LASSO model;

[0094] S324: Based on the LASSO model, the radiomics score is calculated, and the calculation formula is as follows:

[0095] Radiomics Score = a 1 *F 1 +a 2 *F 2 +…+a i *F i +b;

[0096] Among them, a 1 to a i is the coefficient of the target radiomics feature, F 1 To F i is the target imaging feature; the number of target imaging features is i; and b is the intercept.

[0097] In some embodiments, the target imaging genomics features refer to Table 1, which is a table of target imaging genomics features and corresponding coefficients provided in this embodiment.

[0098] Table 1 Target imaging omics features and corresponding coefficients

[0099]

[0100]

[0101] Statistical analysis of the imaging omics features through t-test can be performed to reduce the dimension of feature data, and redundant features can be eliminated through the Pearson correlation coefficient. The imaging omics score is calculated using the least absolute convergence and selection operator (LASSO) regression model, so as to facilitate the combination with the preset independent risk variables to obtain the combined features.

[0102] It should be noted that, in the embodiment of the present invention, the preset independent risk variable is screened from the clinical variables collected by the scientific research data platform, and the clinical variables include age, gender, long axis size of the kidney ultrasound report, ratio of the kidney long axis to age, etc. It is worth noting that in view of the influence of age on kidney size, we introduced the variable of the ratio of kidney long axis to age to reflect. The specific screening method is: first, the quantitative variables (age, long axis size of the kidney ultrasound report, ratio of kidney long axis to age) are tested by independent sample t, and the qualitative variables (gender) are tested by chi-square test, and the corresponding clinical variables of the difference index that meets the preset index threshold (P<0.05) are output as risk variables, and then the risk variables are input into the multifactor analysis model, i.e., the binary logistic regression model, and the cross-influencing factors between the risk variables are adjusted, and the clinical variables that meet the preset index threshold (P<0.05) are output as independent risk factors.

[0103] In some embodiments, the screening analysis results of clinical variables in this example show that the P values ​​of gender, age, ultrasound-reported kidney length and kidney long axis / age ratio are all less than 0.05, which are statistically significant. Then, gender, age, ultrasound-reported kidney length and kidney long axis / age ratio are used as inputs of the multifactor analysis model, and the P value is output, among which gender and ultrasound-reported kidney length meet the preset threshold P value less than 0.05, which are independent risk factors for immunoglobulin A nephropathy.

[0104] Of course, in practice, independent risk factors can also be set directly.

[0105] It is worth noting that the classification result is the classification result output by the corresponding prediction model based on the combined features. The algorithm selection of the prediction model is based on the radiomics model obtained after the input target radiomics feature training (the radiomics model is also a prediction model, and the radiomics model is used here to represent the prediction model obtained by the combined feature training) and the effect evaluation result of the prediction model obtained after the combined feature training. Based on the radiomics features and combined features, multiple different machine learning algorithm models are trained and verified respectively, and the algorithm model with the largest AUC index of the verification results is comprehensively evaluated and selected as the algorithm model of the final prediction model.

[0106] In one or more embodiments of the present invention, the step of constructing a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types specifically includes the following steps:

[0107] S41: inputting the imaging omics features and the combined features into a plurality of preset machine learning algorithms respectively, and calculating the corresponding area under the ROC curve AUC index;

[0108] In some embodiments, this example selects three different machine learning algorithm models including: a logistic regression algorithm (LR) model, a random forest algorithm (RF) model, and a support vector machine algorithm (SVM) model.

[0109] S42: taking the machine learning algorithm with the largest AUC index as the target machine learning algorithm, and building a prediction model based on the target machine learning algorithm;

[0110] S43: Inputting the combined features and the corresponding risk types into the trained prediction model for training.

[0111] The AUC results of the training results and verification results of different machine learning algorithm models provided by the radiomics model in this embodiment show:

[0112] Based on the logistic regression algorithm model: the AUC of the training set is 0.793 (95% CI 0.726-0.856), that is: the area under the curve is estimated to be 0.793, with 95% confidence that the true AUC value is between 0.726 and 0.856; the AUC of the test set is 0.788 (95% CI 0.684-0.885). Based on the random forest algorithm model: the AUC of the training set is 0.800 (95% CI 0.736-0.865); the AUC of the test set is 0.811 (95% CI 0.702-0.901). Based on the support vector machine algorithm model: the AUC of the training set is 0.759 (95% CI 0.687-0.830); the AUC of the test set is 0.734 (95% CI 0.614-0.844).

[0113] Figure 3 : This is a schematic diagram of the AUC of the training results and verification results of different machine learning algorithm models provided by the prediction model based on combined feature training in this embodiment. The results show that:

[0114] Based on the logistic regression algorithm model: the AUC of the training set was 0.852 (95% CI 0.794-0.906); the AUC of the test set was 0.893 (95% CI 0.810-0.952). Based on the random forest algorithm model: the AUC of the training set was 0.916 (95% CI 0.876-0.952); the AUC of the test set was 0.872 (95% CI 0.785-0.938). Based on the support vector machine algorithm model: the AUC of the training set was 0.877 (95% CI 0.824-0.925); the AUC of the test set was 0.876 (95% CI 0.785-0.946).

[0115] It is worth noting that the Delong test for algorithm comparison based on imaging omics model on the test set showed that only the logistic regression algorithm model and the support vector machine algorithm model had significant differences (P=0.023); the Delong test for algorithm comparison based on prediction model trained with combined features on the test set showed that there were significant differences between the random forest algorithm model and the support vector machine algorithm model (P=0.023), the logistic regression algorithm model and the support vector machine algorithm model (P=0.002). Comprehensively evaluated, the logistic regression algorithm model showed the highest AUC value. Therefore, the logistic regression algorithm model was selected to construct the final prediction model.

[0116] In addition, in some embodiments, a clinical model based on a logistic regression algorithm model is established for independent risk factors, and compared with an imaging omics model based on a logistic regression algorithm model for target imaging omics features and a prediction model based on a logistic regression algorithm model for combined features, see Figure 4 , is a schematic diagram of the AUC of the clinical model based on independent risk factors, the radiomics model based on target radiomics features, and the prediction model based on combined features provided in this embodiment. The results show that:

[0117] Clinical model based on independent risk factors: AUC value of the training set was 0.784 (95% CI 0.712-0.849); AUC value of the test set was 0.810 (95% CI 0.699-0.898). Radiomics model based on target radiomics features: AUC value of the training set was 0.793 (95% CI 0.726-0.856); AUC value of the test set was 0.788 (95% CI 0.684-0.885). Prediction model based on combined features: AUC value of the training set was 0.852 (95% CI 0.794-0.906); AUC value of the test set was 0.893 (95% CI 0.810-0.952).

[0118] It is worth noting that the Delong test on the test set showed that there were significant differences between the prediction model based on combined features and the imaging omics model based on target imaging omics features and the clinical model based on independent risk factors, P = 0.019 and P = 0.011, respectively. Its AUC index was much higher than that of the imaging omics model and the clinical model (the AUC value in the test set was 0.893, and the AUC value in the training set was 0.852), showing the superiority of the prediction model when combining independent risk factors with target imaging omics features for training.

[0119] In some embodiments, in order to further evaluate the predictive performance and practical value of the prediction model, the prediction model based on the combined feature training is evaluated by the calibration curve and the clinical decision curve. Referring to Figure 5 (a and b), it is a schematic diagram of the calibration curve and the clinical decision curve based on the imaging genomics model provided in the present embodiment. In the calibration curve, the secondary diagonal corresponding to the 45° dashed line represents the probability under ideal conditions, and the closer the prediction curve is to the ideal curve, the better the prediction performance is. The calibration curve shows that the prediction performance of the imaging genomics model is close to the probability under the ideal state, and the prediction performance is good. In the clinical decision curve, no means no intervention at all, the gray curve means all interventions, and the area under the curve represents the net benefit. The clinical decision curve shows that when the threshold probability is 0.1-1.0, the net benefit obtained using the prediction model based on the combined feature training is substantially higher than all interventions and no interventions, suggesting that the constructed prediction model has clinical application value.

[0120] Optionally, in one or more embodiments of the present invention, the method further comprises the following steps:

[0121] S6: constructing a nomogram according to the combined characteristics of the target patient, and determining the risk probability corresponding to the risk type prediction result according to the nomogram.

[0122] Here, a nomogram is drawn to visualize the classification results, that is, the independent risk factors and target radiomics features input into the prediction model based on combined feature training are placed in one graph, scores are assigned according to the coefficients of the independent risk factors and the target radiomics features, and then the sum of the scores of the independent risk factors and the target radiomics features is calculated to predict the probability of developing the disease. This allows the model's prediction results to be understood without reading statistical tables or equations, which simplifies the use of the model and makes it more intuitive.

[0123] See also Figure 6 , is a schematic diagram of a nomogram provided by the prediction model based on combined feature training in this embodiment.

[0124] For example, see Figure 6 For a female child, the ultrasound report shows that the long axis size of the kidney is, first of all, the gender is female, and the score in the nomogram is point A corresponding to the gender, and the score corresponding to point A in the nomogram is 0. The long axis size of the kidney is 10cm, and the score corresponding to point B in the nomogram is 41. The target imaging features are extracted, and the corresponding coefficients in Table 1 are substituted into the target imaging parameter calculation formula to calculate the target imaging score. Assuming that the target imaging score of -1.4 points corresponds to point C, and the score corresponding to point C in the nomogram is 17, then the final total score is A+B+C≈58 points, so it can be obtained that the probability of suffering from IgAN is <0.1, and it can be described as a low-risk population.

[0125] Compared with the prior art, this embodiment extracts the radiomic features of renal images through high-throughput radiomics, which is not limited to the information reported by the ultrasound doctor, but extracts the first-order features, shape features and texture features of renal ultrasound, etc., converts them into quantitative data, and screens the extracted radiomics features to avoid overfitting of the constructed model, and compares algorithms such as logistic regression and random forest to construct the most stable radiomics model, which is more accurate than the current research. Secondly, a nomogram is further drawn on the basis of the constructed combined prediction model, which simplifies the use of the prediction model and visualizes the prediction probability, so that the risk of patients suffering from IgAN can be more intuitively reflected, which is more conducive to early prediction of patients, thereby improving long-term prognosis.

[0126] like Figure 7As shown, the present invention also provides an ultrasound-based immunoglobulin A nephropathy risk prediction system, including an image recognition module, a feature extraction module, a combined feature module, a construction training module and a risk prediction module;

[0127] The image recognition module is used to obtain a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region;

[0128] The feature extraction module is used to preprocess the target area image in the longest axis ultrasound image of the kidney, and perform radiomics feature extraction processing on the preprocessed target area image;

[0129] The combined feature module is used to reduce the dimension and screen the extracted radiomics features to obtain target radiomics features, calculate the radiomics score, and combine the radiomics score of the target radiomics feature with a preset independent risk variable to obtain a combined feature;

[0130] The training module is used to construct a prediction model based on the radiomics features and the combined features, and train the prediction model based on the combined features and the corresponding risk types;

[0131] The risk prediction module is used to input the combined features corresponding to the target patient into the trained prediction model to obtain the corresponding risk type prediction result.

[0132] In one or more embodiments of the present invention, the image recognition module obtains the longest axis ultrasound image of the kidney of the sample patient, determines the kidney tissue area in the longest axis ultrasound image of the kidney as the target area, and generates the boundary contour of the target area as follows:

[0133] Obtain the kidney ultrasound image in DICOM format and perform analytical conversion to obtain the kidney longest axis ultrasound image in NIFTI format;

[0134] Image recognition processing is performed on the kidney longest axis ultrasound image to determine that the kidney tissue area in the kidney longest axis ultrasound image is the target area, and a boundary contour of the target area is generated.

[0135] In one or more embodiments of the present invention, the feature extraction module preprocesses the target area image in the longest axis ultrasound image of the kidney, and performs radiomics feature extraction processing on the preprocessed target area image as follows:

[0136] Normalizing the target area image in the longest axis ultrasound image of the kidney according to a preset normalization ratio;

[0137] Resampling the normalized target region image in image space to obtain a target region image with a uniform voxel size;

[0138] A plurality of filters are used to perform image enhancement processing on the target area image, and radiomics feature extraction processing is performed on the enhanced target area image to obtain radiomics features, wherein the image enhancement processing includes filter enhancement image processing and mathematical transformation enhancement image processing.

[0139] In one or more embodiments of the present invention, the combined feature module performs dimension reduction and screening on the extracted radiomics features in the following specific implementations:

[0140] The extracted radiomics features were evaluated for intra-group consistency and the intra-group correlation coefficient was calculated;

[0141] The radiomics features whose intragroup correlation coefficient is not less than a preset threshold are converted into a standard normal distribution, and data dimension reduction and screening are performed to obtain target radiomics features, and the radiomics score is calculated.

[0142] In one or more embodiments of the present invention, the combined feature module performs data dimension reduction and screening on the radiomics features converted into a standard normal distribution to obtain target radiomics features, and calculates the radiomics score in the following specific implementations:

[0143] Performing statistical analysis on the imaging features using a t-test, and only retaining the imaging features with a p-value less than a preset p-threshold;

[0144] Using the preset Pearson correlation coefficient, any two imaging features with correlation coefficients greater than the preset correlation threshold are randomly removed to obtain the target imaging features;

[0145] The least absolute convergence and selection operator (LASSO) regression model was constructed, and the lambda parameter with the minimum mean square error (MSE) was calculated through cross-validation of the minimum criterion, and finally the non-zero coefficient characteristics in the LASSO model were obtained;

[0146] Based on the LASSO model, the radiomics score was calculated using the following formula:

[0147] Radiomics Score = a 1 *F 1 +a 2 *F 2 +…+a i *F i +b;

[0148] Among them, a1 to a i is the coefficient of the target radiomics feature, F 1 To F i is the target imaging feature; the number of target imaging features is i; and b is the intercept.

[0149] In one or more embodiments of the present invention, the construction of the training module constructs a prediction model based on the radiomics features and the combined features, and trains the prediction model based on the combined features and the corresponding risk types as follows:

[0150] The imaging omics features and combined features are respectively input into a plurality of preset machine learning algorithms, and the corresponding area under the ROC curve AUC index is calculated;

[0151] Taking the machine learning algorithm with the largest AUC index as the target machine learning algorithm, and building a prediction model based on the target machine learning algorithm;

[0152] The combined features and the corresponding risk types are input into the trained prediction model for training.

[0153] Optionally, in one or more embodiments of the present invention, the ultrasound-based immunoglobulin A nephropathy risk prediction system also includes a risk probability module, which is used to construct a nomogram based on the combined characteristics of the target patient, and determine the risk probability corresponding to the risk type prediction result based on the nomogram.

[0154] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the ultrasound-based immunoglobulin A nephropathy risk prediction method is implemented.

[0155] The present invention also provides an ultrasound-based immunoglobulin A nephropathy risk prediction device, comprising a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0156] The memory is used to store computer programs;

[0157] The processor is used to implement the steps of the ultrasound-based immunoglobulin A nephropathy risk prediction method when executing the program stored in the memory.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the risk of immunoglobulin A nephropathy based on ultrasound, characterized in that: The steps include: Acquire a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region; Preprocessing the target area image in the longest axis ultrasound image of the kidney, and performing radiomics feature extraction processing on the preprocessed target area image; Performing dimensionality reduction and screening on the extracted radiomics features to obtain target radiomics features, and calculating radiomics scores, and combining the radiomics scores of the target radiomics features with preset independent risk variables to obtain combined features; Building a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types; The combined features corresponding to the target patient are input into the trained prediction model to obtain the corresponding risk type prediction result.

2. The method for predicting the risk of immunoglobulin A nephropathy based on ultrasound according to claim 1, characterized in that: The step of obtaining the longest axis ultrasound image of the kidney of the sample patient, determining the kidney tissue region in the longest axis ultrasound image of the kidney as the target region, and generating the boundary contour of the target region specifically comprises the following steps: Obtain the kidney ultrasound image in DICOM format and perform analytical conversion to obtain the kidney longest axis ultrasound image in NIFTI format; Image recognition processing is performed on the kidney longest axis ultrasound image to determine that the kidney tissue area in the kidney longest axis ultrasound image is the target area, and a boundary contour of the target area is generated.

3. The method for predicting the risk of immunoglobulin A nephropathy based on ultrasound according to claim 1, characterized in that: The preprocessing of the target area image in the longest axis ultrasound image of the kidney and the radiomics feature extraction processing of the preprocessed target area image specifically include the following steps: Normalizing the target area image in the longest axis ultrasound image of the kidney according to a preset normalization ratio; Resampling the normalized target region image in image space to obtain a target region image with a uniform voxel size; A plurality of filters are used to perform image enhancement processing on the target area image, and radiomics feature extraction processing is performed on the enhanced target area image to obtain radiomics features, wherein the image enhancement processing includes filter enhancement image processing and mathematical transformation enhancement image processing.

4. The method for predicting the risk of immunoglobulin A nephropathy based on ultrasound according to claim 1, characterized in that: The dimensionality reduction and screening of the extracted radiomics features specifically include the following steps: The extracted radiomics features were evaluated for intra-group consistency and the intra-group correlation coefficient was calculated; The radiomics features whose intragroup correlation coefficient is not less than a preset threshold are converted into a standard normal distribution, and data dimension reduction and screening are performed to obtain target radiomics features, and the radiomics score is calculated.

5. The method for predicting the risk of immunoglobulin A nephropathy based on ultrasound according to claim 4, characterized in that: The radiomics features converted into a standard normal distribution are subjected to data dimension reduction and screening to obtain target radiomics features, and the radiomics score is calculated, which specifically includes the following steps: Performing statistical analysis on the imaging features using a t-test, and only retaining the imaging features with a p-value less than a preset p-threshold; Using the preset Pearson correlation coefficient, any two imaging features with correlation coefficients greater than the preset correlation threshold are randomly removed to obtain the target imaging features; The least absolute convergence and selection operator (LASSO) regression model was constructed, and the lambda parameter with the minimum mean square error (MSE) was calculated through cross-validation of the minimum criterion, and finally the non-zero coefficient characteristics in the LASSO model were obtained; Based on the LASSO model, the radiomics score was calculated using the following formula: Radiomics Score=a1*F1+a2*F2+…+a i *F i +b; Among them, a1 to a i is the coefficient of the target radiomics feature, F1 to F i is the target imaging feature; the number of target imaging features is i; and b is the intercept.

6. The method for predicting the risk of immunoglobulin A nephropathy based on ultrasound according to claim 1, characterized in that: The step of constructing a prediction model based on the radiomics features and the combined features, and training the prediction model based on the combined features and the corresponding risk types specifically includes the following steps: The imaging omics features and combined features are respectively input into a plurality of preset machine learning algorithms, and the corresponding area under the ROC curve AUC index is calculated; Taking the machine learning algorithm with the largest AUC index as the target machine learning algorithm, and building a prediction model based on the target machine learning algorithm; The combined features and the corresponding risk types are input into the trained prediction model for training.

7. The ultrasound-based method for predicting the risk of immunoglobulin A nephropathy according to any one of claims 1 to 6, characterized in that: The method further comprises the steps of: A nomogram is constructed according to the combined characteristics of the target patient, and the risk probability corresponding to the risk type prediction result is determined according to the nomogram.

8. An ultrasound-based immunoglobulin A nephropathy risk prediction system, characterized in that: It includes image recognition module, feature extraction module, combined feature module, construction training module and risk prediction module; The image recognition module is used to obtain a kidney longest axis ultrasound image of a sample patient, determine a kidney tissue region in the kidney longest axis ultrasound image as a target region, and generate a boundary contour of the target region; The feature extraction module is used to preprocess the target area image in the longest axis ultrasound image of the kidney, and perform radiomics feature extraction processing on the preprocessed target area image; The combined feature module is used to reduce the dimension and screen the extracted radiomics features to obtain target radiomics features, calculate the radiomics score, and combine the radiomics score of the target radiomics feature with the preset independent risk variable to obtain the combined feature; The training module is used to construct a prediction model based on the radiomics features and the combined features, and train the prediction model based on the combined features and the corresponding risk types; The risk prediction module is used to input the combined features corresponding to the target patient into the trained prediction model to obtain the corresponding risk type prediction result.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the ultrasound-based immunoglobulin A nephropathy risk prediction method according to any one of claims 1 to 7 is implemented.

10. An ultrasound-based immunoglobulin A nephropathy risk prediction device, characterized in that: It includes a communication interface, a memory, a communication bus and a processor, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the steps of the ultrasound-based immunoglobulin A nephropathy risk prediction method as described in any one of claims 1 to 7 when executing the program stored in the memory.

Images