A method, system, device and terminal for identifying neonatal hyperbilirubinemia

Abstract

The present application belongs to the technical field of medical image processing and deep learning, and discloses a method, system, device and terminal for identifying neonatal hyperbilirubinemia, which selects appropriate images from collected magnetic resonance image data as templates, performs registration operation on the remaining samples and the templates, extracts effective regions from the registered data, divides a plurality of small regions in the effective regions and extracts relevant features of the small region images, builds a graph convolutional neural network model for classification, analyzes the relevant performance of the graph convolutional neural network model by calculating corresponding indexes, combines the attention mechanism in the graph convolutional neural network to locate the model region, optimizes the classification performance by using the located region, and completes the prediction task of neonatal hyperbilirubinemia. The present application reduces the inter-observer error caused by direct reading by using a deep learning model and a supervised learning method with labeled data, and optimizes the identification performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing and deep learning, and particularly relates to a new-born high-bilirubinemia identification method, system, device and terminal. BACKGROUND

[0002] At present, new-born high-bilirubinemia is a disease that new-borns often suffer from after birth. If not treated in time, it can induce serious diseases such as high-bilirubin encephalopathy, causing irreversible damage to the brain organs and seriously endangering the life and health of new-borns. If new-born high-bilirubinemia can be intervened and treated in time, serious consequences can be avoided, and the survival and life quality of new-borns can be greatly improved.

[0003] The ways to diagnose new-born high-bilirubinemia mainly include the following three kinds: (1) checking the degree of jaundice: that is, judging whether the skin and eyes are obviously yellow by visual observation; (2) determining the level of bilirubin: that is, detecting the blood sample by drawing the blood of the new-born; (3) checking the liver function and other indicators: that is, determining by checking the serum choline and glutamic-oxalacetic transaminase and other indicators. In addition, medical staff also need to understand the medical history and family history of the new-born in order to determine whether there are other risk factors, such as the mother suffering from hepatitis B and other diseases or the new-born being asphyxiated at birth. However, there are respective problems in the above three clinical test methods: due to the existence of observer error, the method of observing the severity of jaundice for disease diagnosis may have inconsistent conclusions among different doctors; the method of generally using serum bilirubin numerical level determination has the disadvantages of difficult diagnosis, long detection time and multiple sampling; the method of checking liver function needs to determine the results according to a large number of indicators, and the whole detection process is relatively cumbersome. The existence of these problems brings great difficulties to the identification and diagnosis of new-born high-bilirubinemia.

[0004] With the continuous progress of technology, various types of medical imaging technology provide technical support for the diagnosis of new-born high-bilirubinemia. Equally important is that imaging technology can also be used to monitor the treatment effect of new-born high-bilirubinemia. X-ray photography, CT, ultrasonic imaging and other technologies can monitor the degree of skin jaundice of new-borns in phototherapy in real time, helping doctors to adjust treatment parameters. In recent years, some researchers have begun to use MRI (magnetic resonance imaging) technology to detect new-born high-bilirubinemia. MRI technology uses the interaction of magnetic fields and radio waves to produce high-definition images, and the imaging effect of soft tissues, bones and brain tissues is better than that of X-ray, ultrasonic and other imaging technologies.

[0005] The application of deep imageomics methods in the differential diagnosis of neonatal hyperbilirubinemia is still in the exploratory and development stage, and there is room for improvement in the network structure. In most related studies, there is a problem of a small number of samples of neonatal hyperbilirubinemia patients, making it difficult to obtain a deep learning differential diagnosis model with good performance indicators. In addition, when using MRI to assist in predicting neonatal hyperbilirubinemia, the specific manifestations of the disease in imaging are not combined, and the pathological characteristics are not explored. These problems existing in the field of neonatal hyperbilirubinemia research need to be further solved and improved in the future.

[0006] Through the above analysis, the problems and defects of the prior art are:

[0007] (1) The existing blood test method for detecting neonatal hyperbilirubinemia is traumatic, and the method for determining the serum bilirubin value level has the disadvantages of difficult diagnosis, long detection time, and the need for multiple sampling.

[0008] (2) The method of checking liver function needs to determine the results according to a large number of indicators, and the entire detection time is long.

[0009] (3) The method of observing the severity of jaundice for disease diagnosis may have inconsistent conclusions among different doctors, and there is a large observer error.

[0010] (4) The method of using deep imageomics to predict neonatal hyperbilirubinemia is still in the exploratory and development stage, and the existing prediction model has low prediction accuracy for the disease.

[0011] (5) When using MRI to assist in predicting neonatal hyperbilirubinemia, the specific manifestations of the disease in imaging are not combined, which is not conducive to finding the specific manifestations of neonatal hyperbilirubinemia in imaging and the lesion area. SUMMARY

[0012] In view of the problems existing in the prior art, the present application provides a neonatal hyperbilirubinemia differential method, system, device and terminal, especially a neonatal hyperbilirubinemia differential method, system, device and terminal based on a graph convolutional neural network.

[0013] The application is implemented as follows: a new-born high bilirubinemia identification method, the new-born high bilirubinemia identification method comprising: selecting a suitable image as a template from collected magnetic resonance image data, performing a registration operation on the remaining samples and the template; extracting an effective area from the registered data, dividing a plurality of small areas in the effective area and extracting relevant features of small area images; building a graph convolutional neural network model for classification, analyzing the relevant performance of the graph convolutional neural network model by calculating corresponding indicators; combining the attention mechanism in the graph convolutional neural network to locate the model area, optimizing the classification performance by using the located area, and completing the prediction task of new-born high bilirubinemia.

[0014] Further, the new-born high bilirubinemia identification method comprises the following steps:

[0015] Step one: according to certain rule standards, screening suitable sample data; combining the clinical information and image picture information of the sample data, selecting the same magnetic resonance sequence data between different data samples as the sample data set for subsequent experiments according to the relevant standards and the shooting quality of the image;

[0016] Step two: performing a registration operation on the sample data and extracting the region of interest; for the collected N experimental data, selecting a sample meeting the requirements in shape orientation, size and gray value as the template data for registration, and taking the remaining N-1 samples as the data to be registered; under the relevant registration software, selecting a suitable registration method to match one by one with the template data, and finally obtaining N registration completed data; selecting M central slices in the sequence and extracting the central effective area;

[0017] Step three: selecting a plurality of small areas in the region of interest and extracting features; randomly selecting small areas with a size of SxS from the N effective area data, and the number of small areas selected in each example is H; extracting features of the small areas in each sample to obtain the reconstructed feature vector;

[0018] Step four: building a graph convolutional neural network classification model and analyzing the model prediction indicators; according to certain rules, taking the small area as the node of the graph convolutional neural network, constructing the connection relationship between the nodes and edges to obtain the graph data structure; then selecting appropriate loss function and optimizer to train the model, and finally analyzing the model performance according to the calculation of the corresponding indicators;

[0019] Step five, determine the area of interest of the model in combination with the attention mechanism; through the attention mechanism of the graph convolutional neural network, combine local information and global information, select the K largest weights in the H small areas divided as the explanation of the model prediction result, and perform visualization operation in the original graph; for the selected K parts, perform classification prediction through the classifier, and analyze the change of performance indicators.

[0020] Further, the screening of suitable sample data according to certain rule standards in step one includes:

[0021] (1) In combination with relevant clinical case information and medical knowledge, age and pathological characteristics conforming to the standard are selected from a large number of samples as suitable objects to be processed;

[0022] (2) Each suitable object to be processed contains different sequence of magnetic resonance data, and a unified modality type is selected to obtain N sample data.

[0023] Further, the registration operation and extraction of the region of interest of the sample data in step two includes:

[0024] (1) The data screened and under the same magnetic resonance modality is read into the related software, and one example of data meeting the size and shape requirements is selected as the registration reference template;

[0025] (2) The remaining N-1 examples of magnetic resonance data are taken as the objects to be registered, and the objects to be registered and the selected reference template are registered respectively to obtain N-1 examples of new data after registration;

[0026] (3) The data type of the N-1 examples of new data after registration is three-dimensional voxel format, and the same M slices are extracted from each registered image, totaling M×(N-1) slice data.

[0027] Further, the selection of a plurality of small areas in the region of interest and the extraction of features in step three includes:

[0028] (1) For the N examples of data samples after registration, H small areas with a size of S×S are selected in each example;

[0029] (2) The small areas with a size of S×S obtained are encoded to obtain the reconstructed feature vector.

[0030] Further, the construction of the graph convolutional neural network classification model and the analysis of the model prediction indicators in step four include:

[0031] (1) Construct a graph convolutional neural network according to the relationship between the same category and the difference between different categories;

[0032] (2) Divide the sample data into training set, validation set and test set, select the corresponding loss function and optimizer;

[0033] (3) Train the model, and evaluate the model performance by calculating the corresponding indicators in the test set.

[0034] Further, the combined graph attention mechanism in step five determines the region that the model focuses on, including:

[0035] (1) According to the attention mechanism in the graph convolutional neural network, select the K regions with the largest weights in the H small regions that have been divided;

[0036] (2) Take the K regions as the graph convolutional neural network model to predict the result, and perform visualization operation in the original graph;

[0037] (3) Take the K effective regions processed by the graph convolutional neural network as more representative parts in the sample category, use the feature vector processed by the graph convolutional neural network, and combine the logistic regression method in machine learning to perform binary classification, and obtain the classification result.

[0038] Another object of the present application is to provide a new-born high bilirubinemia identification system applying the new-born high bilirubinemia identification method, which comprises:

[0039] An image registration module is used to select appropriate images as templates from the collected magnetic resonance image data, and perform registration operation on the remaining samples and the templates;

[0040] An image feature extraction module is used to extract effective regions from the registered data, divide a plurality of small regions in the effective regions, and extract relevant features of the small region images;

[0041] A network model construction module is used to build a graph convolutional neural network model for classification, and analyze the related performance of the graph convolutional neural network model by calculating the corresponding indicators;

[0042] A model region positioning module is used to position the model region by combining the attention mechanism in the graph convolutional neural network, and optimize the classification performance by using the positioned region.

[0043] Another object of the present application is to provide a computer device, which comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the new-born high bilirubinemia identification method.

[0044] Another object of the present application is to provide an information data processing terminal for realizing the new-born high bilirubinemia identification system.

[0045] In combination with the above technical solutions and the technical problems solved, the technical solutions to be protected by the present application have the following advantages and positive effects:

[0046] First, in view of the technical problems existing in the prior art and the difficulty in solving the problems, the present application closely combines the technical solutions to be protected and the results and data obtained during the research and development process, and analyzes in detail and depth how the technical solutions solve the technical problems and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0047] The new-born high bilirubinemia differential method of the present application selects appropriate images as templates from a large amount of collected magnetic resonance image data, performs registration operation on the remaining samples, extracts effective regions from the registered data, divides a plurality of small parts in the effective regions and extracts relevant features of the images in the parts, builds a graph convolutional neural network model for classification, analyzes the relevant performance of the model by calculating corresponding indicators, locates the model attention area in combination with the attention mechanism in the graph convolutional neural network, indicates the specific performance of the new-born high bilirubinemia in the imaging, and optimizes the classification performance by using the located area.

[0048] The present application finds a template suitable for all images, which is conducive to formulating a unified standard. After registration of the remaining images with the template, all data can be fully and effectively utilized, the observer error caused by direct film reading is avoided, and the accuracy of the differential is better improved. The present application selects and builds a suitable deep learning model, which can achieve higher relevant indicators in this task, so that the classification result has higher accuracy. At the same time, according to the attention mechanism in the graph convolutional neural network, the model interested area can be found out to explain the prediction result of the model. The present application uses relevant mechanisms and methods, which is conducive to making relevant explanations for the conclusions made by the model, and locating the effective area on the image, which can provide reference information for new-born high bilirubin children and is conducive to the follow-up diagnosis.

[0049] Second, from the perspective of the product as a whole, the technical solutions to be protected by the present application have the following technical effects and advantages:

[0050] The new-born high bilirubinemia identification method of the application can process a large number of magnetic resonance images, adopts a method combining deep learning and imageomics, and efficiently and accurately predicts new-born high bilirubinemia. Compared with traditional blood testing or observation methods, the prediction accuracy is higher. The application provides a research idea for image evaluation while predicting new-born high bilirubinemia, and has important reference value for studying brain tissue of patients, quantifying the degree and position of bilirubin deposition and other problems.

[0051] Third, as the creative evidence of the claims of the application, it is also embodied in the following important aspects:

[0052] (1) The expected income and commercial value of the technical scheme of the application after transformation are:

[0053] New-born high bilirubinemia is a disease commonly suffered by new-born babies after birth. The application of the new-born high bilirubinemia prediction method proposed by the application is expected to provide effective means for early diagnosis, accurate treatment and personalized medical treatment of the disease. The expected income includes but is not limited to the following aspects: improving the diagnostic accuracy and effectiveness of new-born high bilirubinemia, reducing the misdiagnosis rate and missed diagnosis rate, improving the treatment effect and survival rate; realizing personalized medical treatment, tailoring the treatment plan for patients according to the prediction results, avoiding unnecessary treatment and drug side effects; increasing the confidence and satisfaction of patients in the treatment of new-born high bilirubinemia, and improving the brand value and reputation of the hospital.

[0054] The commercial value of the application mainly includes: wide market prospect. According to statistics, about 600,000 to 2 million new-born babies suffer from high bilirubinemia in China every year, which seriously threatens the health and safety of new-born babies. Therefore, the prediction method of new-born high bilirubinemia has broad market prospect and potential commercial value; technical leading. The prediction and diagnosis of high bilirubinemia based on the application have high precision and reliability, and have obvious technical advantages and leading position compared with other new-born high bilirubinemia identification methods; cooperation opportunity. The application can provide cooperation opportunities for pharmaceutical enterprises and medical institutions to jointly develop and promote related products and services, and realize the win-win business model.

[0055] (2) The technical scheme of the application fills the domestic and foreign industry technical blank:

[0056] The application uniquely focuses on the identification and image evaluation of new-born high bilirubinemia, takes new-born high bilirubinemia as the prediction task, and evaluates the performance form of high bilirubinemia of different degrees on the image, which is a method not involved in the past research. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings described below only represent some of the embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0058] Figure 1 is a flow chart of the method for identifying neonatal hyperbilirubinemia provided by the embodiments of the present application;

[0059] Figure 2 is a schematic diagram of the overall experimental process provided by the embodiments of the present application;

[0060] Figure 3 is a flow chart of the small area selection method provided by the embodiments of the present application;

[0061] Figure 4 is a schematic diagram of the structure of the graph convolutional neural network module provided by the embodiments of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0063] In view of the problems existing in the prior art, the present application provides a method, system, device and terminal for identifying neonatal hyperbilirubinemia, which will be described in detail below in combination with the drawings.

[0064] As shown in Figure 1 , the method for identifying neonatal hyperbilirubinemia provided by the embodiments of the present application comprises the following steps:

[0065] S101, selecting a suitable image from the collected magnetic resonance image data as a template, and performing registration operation on the remaining samples and the template;

[0066] S102, extracting the effective area of the registered data, dividing a plurality of small areas in the effective area and extracting the related features of the small area images;

[0067] S103, building a graph convolutional neural network model for classification, and analyzing the related performance of the graph convolutional neural network model by calculating the corresponding indicators;

[0068] S104, positioning the model area in combination with the attention mechanism in the graph convolutional neural network, and optimizing the classification performance by using the positioning area.

[0069] As shown in Figure 2As shown, the new-born high bilirubinemia differential method provided by the embodiment of the application specifically comprises the following steps:

[0070] (1) According to certain rule standards, suitable sample data are screened, combined with sample data clinical information and image picture information, and data of the same magnetic resonance sequence between different data samples are selected as sample data sets for subsequent experiments according to relevant standards and image shooting quality. A total of N sample data can be collected. According to certain rule standards, suitable sample data are screened as follows:

[0071] (1.1) Combined with relevant clinical case information and medical knowledge, cases with age and pathological characteristics meeting the standards are screened from a large number of samples as suitable processing objects;

[0072] (1.2) Each suitable processing object contains magnetic resonance data of different sequences, and a unified modality type needs to be selected to obtain N sample data.

[0073] (1.3) The N data samples are divided into a high bilirubinemia group and a normal control group.

[0074] (2) The sample data are subjected to registration operation and the region of interest is extracted; for the N experimental data collected in the first step, one sample meeting the requirements in shape orientation, size and gray value is selected as the template data for registration, and the remaining N-1 samples are taken as the data to be registered. In the relevant registration software, a suitable registration method is selected to match the template data one by one, and finally N registration completed data are obtained. M slices in the center of the sequence are selected, and the central effective region is extracted.

[0075] The sample data are subjected to registration operation and the region of interest is extracted, and the following steps are performed:

[0076] (2.1) The data screened in the first step and under the same magnetic resonance modality are read into the relevant software, and one data meeting the requirements in size and shape is selected as the registration reference template;

[0077] (2.2) The remaining N-1 magnetic resonance data are taken as the data to be registered, and they are respectively registered with the selected reference template to obtain N-1 new data after registration;

[0078] (2.3) The N-1 new data after registration are in three-dimensional voxel format, and a corresponding number of slices need to be extracted from each example for subsequent operation. The same M slices are extracted from each registered image, and a total of Mx(N-1) slice data are obtained.

[0079] (3) Selecting a plurality of small regions in the region of interest and extracting features thereof, randomly selecting small regions of SxS in size from the N examples of data in which the effective regions are extracted in step two, the number of small regions selected in each example being H, for each small region in each example sample, extracting features thereof to obtain a reconstructed feature vector.

[0080] In which a plurality of small regions are selected in the region of interest and features thereof are extracted, as follows:

[0081] (3.1) For the N examples of data samples in which registration is completed, H small regions of SxS are selected in each example;

[0082] (3.2) The small regions of SxS obtained are encoded to obtain a reconstructed feature vector.

[0083] (4) Building a graph convolutional neural network classification model, analyzing the model prediction indicators, according to certain rules, taking the small regions as nodes of the graph convolutional neural network, constructing the connection relationship between the nodes and the edges, obtaining the graph data structure, then selecting appropriate loss functions and optimizers to train the model, and finally analyzing the performance of the model according to the calculation of the corresponding indicators.

[0084] In which a graph convolutional neural network classification model is built, and the model prediction indicators are analyzed, as follows:

[0085] (4.1) Building a graph convolutional neural network according to the relationship between the same category and the difference between different categories;

[0086] (4.2) Dividing the sample data into a training set, a validation set and a test set, and selecting appropriate loss functions and optimizers;

[0087] (4.3) Training the model, and evaluating the performance of the model by calculating the corresponding indicators in the test set.

[0088] (5) Combining the graph attention mechanism, finding the region that the model focuses on, and further improving the classification performance, through the attention mechanism of the graph convolutional neural network, fully combining local information and global information, selecting the K largest weights from the H small regions divided previously, and taking them as the explanation of the model prediction result, and performing visualization operation in the original graph. At the same time, the K selected parts can become the feature region representing this category, which is classified and predicted by the classifier, and the change of the performance indicators is analyzed.

[0089] In which the graph attention mechanism is combined to find the region that the model focuses on and further improve the classification performance, as follows:

[0090] (5.1) According to the attention mechanism in the graph convolutional neural network, the K regions with the largest weights in the previously divided H small regions are selected;

[0091] (5.2) The K regions are used as an explanation for the prediction result of the graph convolutional neural network model, and a visualization operation is performed on the original image;

[0092] (5.3) The K effective regions processed by the graph convolutional neural network can be regarded as more representative parts in the sample category, and a further classification result can be obtained by using the feature vector processed by the graph convolutional neural network and combining the logistic regression method in machine learning for binary classification.

[0093] The new-born high-bilirubinemia identification system provided by the embodiment of the application comprises:

[0094] An image registration module is configured to select a suitable image as a template from the collected magnetic resonance image data, and perform a registration operation on the remaining samples and the template.

[0095] An image feature extraction module is configured to extract effective regions from the registered data, divide a plurality of small regions in the effective regions, and extract relevant features of the small region images.

[0096] A network model construction module is configured to build a graph convolutional neural network model for classification, and analyze the performance of the graph convolutional neural network model by calculating corresponding indicators.

[0097] A model region positioning module is configured to position a model region by combining an attention mechanism in the graph convolutional neural network, and predict the new-born high-bilirubinemia by using the positioned region.

[0098] In order to prove the creativity and technical value of the technical scheme of the application, this part is an application embodiment of the technical scheme of the claim on a specific product or related technology. The new-born high-bilirubinemia prediction method provided by the application can be used for identifying and diagnosing high-bilirubinemia, so as to help doctors to formulate a more personalized treatment plan for patients.

[0099] Embodiment 1: A patient is found to have abnormal indicators during blood sampling test, and after further examination and diagnosis, it is determined that the patient has high-bilirubinemia. The doctor inputs the brain MR image of the patient collected into the system, and the system gives a prediction result of high-bilirubinemia, and according to the result, the doctor can formulate a corresponding treatment plan.

[0100] Example 2: Two radiologists have different opinions on the diagnosis of neonatal hyperbilirubinemia based on the brain MR image of a certain patient, and the patient cannot be tested for blood indicators in a short time. The MRI image is sent to the present application, and the prediction result of the patient's hyperbilirubinemia provided by the present application can provide a reference for resolving the disagreement between doctors.

[0101] As a preferred embodiment, the method for identifying neonatal hyperbilirubinemia provided by the embodiment of the present application specifically comprises the following steps:

[0102] (1) According to certain rule standards, suitable sample data is screened, and the specific process is as follows:

[0103] (1a) The data used in the embodiment of the present application is the brain MR image of a newborn, combined with specific medical record information, removing premature infant cases, and only using full-term birth newborns as data samples, a total of 606 data samples are obtained.

[0104] (1b) The selected image contains T1, T2, DWI and other MR sequences, and T1 sequence is selected as the experimental object, which contains 273 cases of hyperbilirubinemia samples and 333 cases of normal control group samples, each containing 15 two-dimensional slice numbers.

[0105] (2) The sample data is registered and the region of interest is extracted, and the specific process is as follows:

[0106] (2a) According to the shape and size, 1 sample is selected as the reference image, and the size of the image is 512x512x15, and the remaining 605 data are used as the registered image;

[0107] (2b) Using image processing software 3D-Slicer and MITK for medical image processing, the method of rigid registration is selected to obtain registered three-dimensional voxel data, a total of 606 examples, each consisting of 15 two-dimensional MR image slices;

[0108] (2c) The registered 606 three-dimensional voxel data samples are further processed, and the middle 5 layers of the 15 two-dimensional slices are selected for subsequent experimental data;

[0109] (2d) For the extracted two-dimensional slices, the center of the image is taken as the reference point, and the region of interest with a size of 128x128 is divided.

[0110] (3) As shown in Figure 3 , a number of small regions are selected in the region of interest, and their features are extracted, and the specific process is as follows:

[0111] (3a) randomly selecting 25 small regions in the extracted 128*128 size region of interest, each small region being 16*16 in size;

[0112] (3b) encoding each small region to obtain 25 1*1600 feature vectors.

[0113] (4) As shown in the figure, the application analyzes the model prediction index by building a graph convolutional neural network classification model, and the specific process is as follows: Figure 4

[0114] (4a) 606 data samples are divided into training set, validation set and test set according to the ratio of 8:1:1, and the number of samples in the three sets is 481, 62 and 63 respectively;

[0115] (4b) The adjacency matrix in the graph convolutional neural network is constructed, and the Hamming distance between images is taken as the basis for the difference between the two, and in the two types of small regions, the calculation is carried out between the two types of data according to the principle of maximum inter-class difference and minimum intra-class difference;

[0116] (4c) The above 25 feature vectors are taken as the nodes of the network, and 5 adjacent nodes adjacent to them are found according to the principle in (4a) to construct a graph structure; the constructed graph structure is first passed through a convolutional layer, and a 1*128 feature vector is output;

[0117] (4d) Then pass through a pooling layer with a parameter setting of 0.8, output a graph structure containing 20 nodes, and calculate the sum of all feature vectors and record it as y';

[0118] (4e) Then pass through a convolutional layer again to ensure that the dimension of the output feature vector is still 1*128;

[0119] (4f) Again pass through a pooling layer with a parameter setting of 0.9, output a graph structure containing 18 nodes, and calculate the sum of all feature vectors and record it as z';

[0120] (4g) Add z' and y' to obtain a 1*128 feature vector a';

[0121] (4h) Pass through two linear layers again, and finally make a binary classification on the input;

[0122] (4i) The PyTorch framework version used in the experiment is 1.2.0, the Python version is 3.7, the number of training rounds in the network is set to 200, the learning rate parameter is set to 0.00005, and the regularization parameter is set to 0.012.

[0123] ​(5) Combine the graph attention mechanism to find the area of interest of the model and visualize the explanation. The specific process is as follows:

[0124] (5a) Select the round with the highest accuracy in the validation set and save its weight for testing the test set;

[0125] (5b) Calculate the corresponding indicators in the test set;

[0126] (5c) According to the graph attention mechanism, a set of training weights W1, W2...W 25 and the feature vector X1, X2...X 25 of each node, the global feature vector X can be represented as follows:

[0127]

[0128] (5d) Select the first K representative nodes as representative nodes, and the weight X' of each node can be represented as follows:

[0129]

[0130] (5e) Label the small area corresponding to the first K feature vectors on the original graph, which can more intuitively show the performance of neonatal hyperbilirubinemia on the MR image;

[0131] (5f) The K value here is selected as 5, and the first five representative feature vectors are added to obtain a 1x128 feature vector. The newly obtained feature vector has the same label as before, and the label and feature vector are input into the logistic regression model to calculate the classification accuracy. The relevant indicators for evaluating the performance of the model include the receiver operating characteristic curve (ROC) and the area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV). The corresponding calculation formulas are as follows:

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] Wherein, TP represents the number of positive samples predicted as positive samples, TN represents the number of negative samples predicted as negative samples, FP represents the number of positive samples predicted as negative samples, FN represents the number of negative samples predicted as positive samples, P represents the number of all positive samples, and N represents the number of all negative samples; the ACC, SEN, SPE, PPV, NPV, AUC and the like are all within the range of 0 to 1. The closer the value is to 1, the better the prediction performance of the model. The four prediction results in the test set are shown in Table 1.

[0138] Table 1: Prediction experiment results

[0139] Number of experiments ACC AUC SEN SPE PPV NPV 1 0.76 0.84 0.76 0.80 0.79 0.78 2 0.75 0.83 0.67 0.90 0.86 0.76 3 0.68 0.74 0.60 0.71 0.72 0.78 4 0.71 0.78 0.76 0.70 0.75 0.71

[0140] From the experimental results, the effect of the present application is good, and compared with other diagnosis methods, the present application has higher prediction and discrimination accuracy for neonatal hyperbilirubinemia, and has important significance for the diagnosis and treatment scheme determination of neonatal hyperbilirubinemia.

[0141] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by using special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by using computer executable instructions and / or included in processor control code, such as provided on a carrier medium, such as a magnetic disk, CD or DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc. They can also be realized by software executed by various types of processors, or by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0142] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any modification, equivalent replacement and improvement within the technical range disclosed by the present application, which is made by those skilled in the art within the spirit and principle of the present application, should be covered within the protection scope of the present application.

Claims

1. A method for discriminating between a newborn with hyperbilirubinemia, characterized in that, The method for identifying neonatal hyperbilirubinemia comprises the following steps: The method for identifying neonatal hyperbilirubinemia comprises the following steps: Step one, according to certain rules and standards, suitable sample data are screened; combined with the clinical information and image information of the sample data, the same kind of magnetic resonance sequence data between different data samples is selected as the sample data set for subsequent experiments according to the relevant standards and the shooting quality of the image; Step two, the sample data is registered and the region of interest is extracted; for the collected N experimental data, a sample meeting the requirements in shape orientation, size and gray value is selected as the template data for registration, and the remaining N-1 samples are used as the data to be registered; in the relevant registration software, the appropriate registration method is selected to match the template data one by one, and finally N registration completed data are obtained; M slices in the center of the sequence are selected, and the center effective region is extracted; Step three, a plurality of small regions are selected in the region of interest and features are extracted; from the N extracted effective region data, a small region with a size of SxS is randomly selected, and the number of small regions selected in each sample is H; the features of each small region in the sample are extracted to obtain the reconstructed feature vector; Step four, a graph convolutional neural network classification model is built, and the model prediction index is analyzed; according to certain rules, the small regions are used as the nodes of the graph convolutional neural network, the connection relationship between the nodes and the edges is constructed, and the graph data structure is obtained; then, appropriate loss function and optimizer are selected to train the model, and finally the model performance is analyzed according to the calculation of the corresponding index; Step five, the region focused by the model is determined by combining the graph attention mechanism; through the attention mechanism of the graph convolutional neural network, the local information and the global information are combined, the K regions with the largest weight in the H small regions are selected as the explanation of the prediction result of the model, and the visualization operation is performed on the original image; the K parts are classified and predicted by the classifier, and the change of the performance index is analyzed; Step five of combining the graph attention mechanism to determine the region focused by the model comprises: (1) according to the attention mechanism in the graph convolutional neural network, the K regions with the largest weight in the H small regions are selected; (2) the K regions are used as the prediction result of the graph convolutional neural network model, and the visualization operation is performed on the original image; (3) The K effective regions processed by the graph convolutional neural network are regarded as more representative parts in the sample category, and a binary classification is performed by combining the feature vectors processed by the graph convolutional neural network and a logistic regression method in machine learning to obtain a classification result.

2. The method for differentiating the cause of neonatal hyperbilirubinemia according to claim 1, wherein The screening of appropriate sample data according to certain rule standards in step one includes: (1) Combining relevant clinical case information and medical knowledge, age and pathological characteristics conforming to the standard are screened from a large number of samples to obtain cases as appropriate processing objects; (2) Each appropriate processing object contains magnetic resonance data of different sequences, and a unified modality type is selected to obtain N sample data.

3. The method for differentiating the type of neonatal hyperbilirubinemia according to claim 1, wherein The registration operation and the extraction of the region of interest in the sample data in step two include: (1) The data screened and under the same magnetic resonance modality are read into the related software, and one example of data meeting the size and shape requirements is selected as the reference template for registration; (2) The remaining N-1 examples of magnetic resonance data are taken as the objects to be registered, and the objects to be registered and the selected reference template are registered respectively to obtain N-1 examples of new data after registration; (3) The data type of the N-1 examples of new data after registration is three-dimensional voxel format, and M slices are extracted from each registered image, totaling M×(N-1) slice data.

4. The method for differentiating the degree of hyperbilirubinemia of a newborn according to claim 1, wherein The selection of a number of small regions in the region of interest and the extraction of features in step three include: (1) For the N examples of data samples after registration, H small regions with a size of S×S are selected in each example; (2) The small regions with a size of S×S are encoded to obtain the reconstructed feature vectors.

5. The method for differentiating the degree of the neonatal hyperbilirubinemia according to claim 1, wherein The building of the graph convolutional neural network classification model and the analysis of the model prediction index in step four include: (1) The graph convolutional neural network is constructed according to the relationship between the same category and the difference between different categories; (2) The sample data is divided into a training set, a validation set and a test set, and a corresponding loss function and an optimizer are selected; (3) The model is trained, and the performance of the model is evaluated by calculating the corresponding index in the test set.

6. A system for identifying neonatal hyperbilirubinemia using the method according to any one of claims 1 to 5, characterized in that, The newborn hyperbilirubinemia identification system includes: An image registration module for selecting appropriate images as templates from the collected magnetic resonance image data, and performing registration operation on the remaining samples and the templates; An image feature extraction module for extracting effective regions from the registered data, dividing a number of small regions in the effective regions, and extracting relevant features of the small region images; A network model construction module for building a graph convolutional neural network model for classification, and analyzing the performance of the graph convolutional neural network model by calculating the corresponding index; A model region positioning module for positioning the model region by combining the attention mechanism in the graph convolutional neural network, and optimizing the classification performance by using the positioned region.

7. A computer device, comprising: The computer device includes a memory and a processor, and the memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the newborn hyperbilirubinemia identification method according to any one of claims 1-5.

8. An information data processing terminal, characterized by The information data processing terminal is used to realize the new-born baby hyperbilirubinemia identification system as claimed in claim 6.

Images