Newborn encephalopathy prediction model and method thereof

Abstract

The invention discloses a neonatal encephalopathy prediction model and a method thereof, and relates to the technical field of neonatal encephalopathy prediction. Comprising the following steps: S1, collecting image data of normal newborns and diseased newborns, including whole body images and MRI image data; s2, generating a T1 weighted image, a T2 weighted image and a DWI image for the MRI image data; according to the method, a large number of images of normal newborns and two sick newborns (bilirubin encephalopathy and hypoxia ischemic encephalopathy) are collected according to the proportion of 1: 1: 1, and the male and female proportion is 1: 1, so that data deflection is avoided, the generalization ability of the model is improved, and the diversity and representativeness of a data set are ensured; meanwhile, by generating a T1 weighted image, a T2 weighted image and a DWI image, brain tissue information under multiple perspectives is provided, the features of different encephalopathy can be more accurately recognized, deep learning training is carried out through a U-Net model, accurate and rapid prediction can be provided, and a powerful tool is provided for early diagnosis of neonatal encephalopathy.

Description

Technical Field

[0001] The present invention relates to the technical field of neonatal encephalopathy prediction, and in particular to a neonatal encephalopathy prediction model and method thereof. Background Art

[0002] Neonatal bilirubin encephalopathy is an acute disease caused by abnormally high bilirubin levels in the neonatal period, which causes bilirubin to pass through the blood-brain barrier and deposit in nerve cells, thereby causing nervous system dysfunction. The disease is usually caused by intrauterine asphyxia, hypoxia, infection, acidosis, hemolysis and other factors, and manifests as symptoms such as drowsiness, poor response, and abnormal muscle tone. In severe cases, it can lead to death or leave sequelae such as athetosis, hearing impairment, and intellectual disability. Treatment requires a combination of light therapy, drug therapy, exchange transfusion therapy and other methods, and close monitoring of changes in the disease. The key to prevention is early identification and treatment of high-risk factors, including prenatal examinations, monitoring of bilirubin levels, proper feeding and active treatment of primary diseases, so as to reduce the incidence of the disease and reduce sequelae;

[0003] Hypoxic-ischemic encephalopathy of newborn (HIE) is a disease caused by brain hypoxia and reduced blood flow due to various reasons, which leads to neonatal brain damage. This disease is one of the important complications of neonatal asphyxia and one of the common causes of neurological disability in children. Its main pathological manifestations are edema, softening, necrosis and hemorrhage of brain tissue. The causes of neonatal hypoxic-ischemic encephalopathy are diverse, including fetal distress, birth asphyxia, perinatal infection, congenital heart disease and maternal diseases;

[0004] Both encephalopathy may show abnormalities of the nervous system in terms of symptoms, such as drowsiness, poor response, abnormal muscle tone and convulsions, but their causes, pathogenesis and treatment methods are different. Neonatal bilirubin encephalopathy is mainly caused by excessive bilirubin levels, while neonatal hypoxic-ischemic encephalopathy is related to perinatal brain tissue ischemia and hypoxia, so it needs to be predicted, identified and treated as soon as possible. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a neonatal encephalopathy prediction model and method.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A neonatal encephalopathy prediction model comprises the following steps:

[0008] S1: Collect image data of normal newborns and sick newborns, including whole body images and MRI image data;

[0009] S2: Generate T1-weighted images, T2-weighted images and DWI images for MRI image data;

[0010] S3: preprocessing the generated image and the surrounding image, including denoising and normalization;

[0011] S4: Use the U-Net model to perform deep learning training on the preprocessed images and build a prediction model;

[0012] S5: Introduce the attention mechanism in the U-Net model to focus on important areas in the image;

[0013] S6: Train, validate and test the model;

[0014] S7: Deploy the tested TensorFlow to hospital equipment.

[0015] Preferably, in the image data collected in step S1, the ratio of images of normal newborns, newborns with neonatal bilirubin encephalopathy and newborns with neonatal hypoxic-ischemic encephalopathy is 1:1:1, and the ratio of male to female newborns is 1:1.

[0016] Furthermore: in the step S2, the MRI image data is used to generate T1-weighted images, T2-weighted images and DWI images by adjusting parameters, and the high signals of different weighted images are used to comprehensively predict whether the newborn is sick.

[0017] Furthermore: In the S3 step, the total variation regularization denoising method is used to denoise the T1-weighted image, T2-weighted image and DWI image, retaining the high signal area, and the non-local mean denoising method is used to process the child's body image, retaining the pathological characteristics of yellow skin and pale cyanosis.

[0018] As a preferred solution of the present invention: in the step S3, the denoised T1-weighted image, T2-weighted image and DWI image are normalized by homogeneous standardization.

[0019] As a further solution of the present invention: in the S4 step, semantic segmentation of the MRI image and the neonatal body image is performed by using a labeling tool, different tissue structures in the image are labeled, key points such as the lesion center can be specifically labeled, and corresponding labels are classified and assigned, and a label image (mask) is subsequently generated for model training.

[0020] As a further solution of the present invention: in the step S5, the channels of the feature map of the high signal area of ​​the child in the feature map are weighted to highlight important channel information, and the spatial position of the high signal area of ​​the child in the feature map is weighted to highlight important spatial information.

[0021] Based on the above scheme: During the model training process in step S6, the Adam optimizer is used, the learning rate is 0.0009, the batch size is 64, the Epochs is 500, and L2 regularization is used to prevent overfitting.

[0022] Based on the above solution: the model deployed in the hospital equipment in step S7 can provide prediction services through TensorFlow Serving and can be accessed through RESTful API or gRPC interface.

[0023] A method for using a neonatal encephalopathy prediction model comprises the following steps:

[0024] R1: After the birth of a newborn baby at risk of encephalopathy, MRI images and whole-body images of the newborn are taken to obtain prediction data:

[0025] R2: Generate T1-weighted images, T2-weighted images and DWI images from the obtained MRI images, and use software to process the generated T1-weighted images, T2-weighted images, DWI images and neonatal peripheral images;

[0026] R3: Import the processed image into the loaded TensorFlow model for prediction;

[0027] R4: Combine the model output with medical expertise to determine whether the newborn is at risk of disease and the type of encephalopathy that may occur.

[0028] The beneficial effects of the present invention are:

[0029] 1. A neonatal encephalopathy prediction model and method thereof, by collecting a large number of images of normal newborns and two diseased newborns (bilirubin encephalopathy, hypoxic-ischemic encephalopathy) in a ratio of 1:1:1, and a male-female ratio of 1:1, avoiding data skew, improving the generalization ability of the model, and ensuring the diversity and representativeness of the data set; at the same time, by generating T1-weighted images, T2-weighted images and DWI images, providing brain tissue information from multiple perspectives, which helps to more accurately identify the characteristics of different encephalopathy, using U-Net model for deep learning training, can provide accurate and fast predictions, and provide a powerful tool for the early diagnosis of neonatal encephalopathy.

[0030] 2. A neonatal encephalopathy prediction model and method, which can accurately identify and segment the lesion area through the U-Net deep learning model. In addition, by introducing the channel attention module and the spatial attention module, the model can focus more on key areas and features when processing images, such as high signal areas such as the globus pallidus, thereby improving the accuracy of judging the type of encephalopathy.

[0031] 3. A neonatal encephalopathy prediction model and method, which can help the medical team make decisions faster, provide more timely treatment for newborns, and greatly reduce the diagnosis time; and with the increase of usage time, as well as the accumulation of data and the advancement of technology, the deep learning model can be continuously optimized and updated, providing more reliable encephalopathy prediction services. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the process of building a neonatal encephalopathy prediction model proposed by the present invention;

[0033] Figure 2 It is a flow chart of a method for using a neonatal encephalopathy prediction model proposed by the present invention. DETAILED DESCRIPTION

[0034] The technical solution of this patent is further described in detail below in conjunction with specific implementation methods.

[0035] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0036] Embodiment 1:

[0037] A prediction model method for neonatal encephalopathy, such as Figure 1 As shown, the following steps are included:

[0038] S1: Data collection: First, a large number of image data of normal newborns and sick newborns were collected. These image data included the whole-body image data and MRI (magnetic resonance imaging) image data of newborns. In order to ensure the diversity and representativeness of these data sets, the ratio of normal newborns, newborns with neonatal bilirubin encephalopathy and newborns with neonatal hypoxic-ischemic encephalopathy in these image data was 1:1:1, and the ratio of male to female newborns from which these image data came was ensured to be 1:1.

[0039] MRI, or magnetic resonance imaging, is an advanced medical imaging technology. It uses a strong magnetic field and high-frequency pulses to resonate the nuclei inside the human body and reconstruct images through computers. In MRI images, both encephalopathy will affect partial image changes of the globus pallidus and structural changes of the ventricles, but bilirubin encephalopathy usually manifests as symmetrical high signals in bilateral globus pallidus, while hypoxic-ischemic encephalopathy may present extensive, fuzzy-bordered high-signal shadows involving multiple brain regions;

[0040] S2: Generate T1-weighted images, T2-weighted images and DWI (diffusion-weighted imaging) images; Generate T1-weighted images and T2-weighted images by adjusting the parameters such as repetition time (TR), echo time (TE), flip angle, slice thickness, slice spacing, and field of view in the original MRI image data;

[0041] T1-weighted images can provide better observation of brain anatomical structures. In T1-weighted images, patients with acute bilirubin encephalopathy have symmetrical high signals in the bilateral globus pallidus, while hypoxic-ischemic encephalopathy can cause abnormal morphology and structure of brain tissue and produce corresponding abnormal signals.

[0042] T2-weighted images are mainly used to observe tissue water content and spin speed, and have a good display effect for detecting lesions such as edema, inflammation and masses. Brain tissue edema and inflammation caused by bilirubin encephalopathy appear as high-signal areas in T2-weighted images. Edema caused by hypoxic-ischemic encephalopathy also appears as high-signal areas, accompanied by high-signal abnormalities in the basal ganglia and thalamus and signal abnormalities in white matter damage.

[0043] DWI images are MRI images based on MRI sequences. By applying diffusion-sensitive gradients in three mutually perpendicular directions, the diffusion movement state of water molecules in tissues is detected, and the apparent diffusion coefficient (ADC) value of each pixel is calculated accordingly. Finally, the DWI image reflecting the restricted diffusion of water molecules is obtained through natural logarithm operation. The neuronal apoptosis caused by bilirubin encephalopathy has little effect on the diffusion of water molecules in damaged tissues, and is not obvious in the DWI image; while the cytotoxic edema and nerve fiber damage caused by hypoxic-ischemic encephalopathy are clearly displayed on the DWI image; through the generated T1-weighted images, T2-weighted images and DWI images, the characteristics of encephalopathy in newborns can be extracted, and it can be further determined whether the child has bilirubin encephalopathy or hypoxic-ischemic encephalopathy;

[0044] S3: Preprocess the image. First, denoise the image. Use the total variation regularization denoising method to denoise the T1-weighted image, T2-weighted image and DWI image. The formula is: The high signal area in the image is retained during denoising, so that the high signal area shown in the three images is different, which can be used to confirm whether the newborn has encephalopathy and what specific encephalopathy it has;

[0045] Then the denoised image is normalized by using homogeneous normalization, combining the minimum-maximum normalization and Z-score normalization characteristics to subtract the average value from the intensity value of the original image and then divide it by the maximum range of the image intensity.

[0046] At the same time, the whole body images of the children were processed by the non-local mean denoising method. The skin, mucous membranes and sclera of children with bilirubin encephalopathy were obviously yellowed. The skin would continue to turn yellow and would appear golden yellow under certain lighting conditions. The skin of children with hypoxic-ischemic encephalopathy would become pale and cyanotic locally due to insufficient blood supply and poor local blood circulation. In some cases, the skin would also turn yellow. The non-local homogeneous denoising method retained the pathological characteristics of yellow skin and pale cyanosis, which was more conducive to the diagnosis of the patient's illness.

[0047] S4: Model construction, using the U-Net model to perform deep learning training on MRI images and neonatal peripheral images. First, semantic segmentation of MRI images and neonatal peripheral images is performed through annotation tools, and different tissue structures are annotated in the images. Key points such as the center of the lesion can be specifically annotated and classified and assigned corresponding labels. Subsequently, a labeled image (mask) is generated for model training;

[0048] Then the labeled images (masks) are randomly divided into training set, validation set and test set in a ratio of 8:1:1. When dividing the labeled images (masks), one data sample belongs to only one set.

[0049] Use TensorFlow to build the U-Net model, define convolution blocks according to requirements, use convolution blocks and pooling layers to build the encoder part of U-Net, which is responsible for extracting the features of the label image (mask); use convolution blocks and upsampling layers to build the decoder part of U-Net, which is responsible for restoring the feature map to the size of the original image;

[0050] Add a skip connection between the encoder and decoder to pass the encoder’s feature map to the decoder, and combine the encoder, decoder, and skip connection to build a complete U-Net model;

[0051] S5: Introduce attention mechanism: add a channel attention module or a spatial attention module after the convolution layer, including a channel attention module: by weighting the channels of the feature map of the high signal area of ​​the child in the feature map, highlight the important channel information;

[0052] Spatial attention module: By weighting the spatial position of the high signal area in the feature map of the child, the important spatial information is highlighted. The designed attention module is added to each convolution block. The spatial attention module is added to the decoder. The attention mechanism allows the model to pay more attention to important areas when processing images, such as the globus pallidus area.

[0053] S6: Model training: using Adam optimizer, learning rate is 0.0009, batch size is 64, Epochs is 500, L2 regularization is used to prevent overfitting, and the loss function uses cross entropy loss and IoU loss;

[0054] Then, the preprocessed data set is imported into the model for training. After each training cycle (epoch), the model is evaluated using the validation set data. The evaluation includes accuracy, precision, and recall rate to prevent overfitting or underfitting. During the training process, the training-evaluation-optimization process can be repeated to obtain the best model configuration.

[0055] After the training is completed and the optimal model configuration is obtained, the accuracy, precision, and recall rate of the final performance of the model are evaluated using the previously divided test set. After the test is completed, a neonatal encephalopathy prediction model that can be deployed and put into practical use is obtained;

[0056] S7: Deployment: Install TensorFlow Serving in hospital equipment, and deploy the model trained with TensorFlow to TensorFlow Serving by mounting a volume. After TensorFlow Serving is started, it will automatically load the model and provide prediction services. The model service can be accessed through the RESTful API or gRPC interface. When in use, use the client code to send requests to TensorFlow Serving and obtain the prediction results.

[0057] The prediction model generated by the above method collects a large number of image data of normal newborns and diseased newborns, and ensures the diversity and representativeness of the data set. The model can learn rich feature information, thereby improving the accuracy of the prediction. At the same time, it combines the advantages of the two types of image information to provide a more comprehensive basis for diagnosis. In addition, the attention mechanism is introduced to enable the model to pay more attention to important areas such as the globus pallidus when processing images, further improving the accuracy of the prediction. The U-Net model is used for deep learning training. The model performs well in the field of medical image segmentation and can accurately identify and segment the lesion area.

[0058] In addition, the model is based on deep learning technology and has strong learning ability and adaptability. After deployment, it can further improve the prediction performance through continuous learning and optimization.

[0059] Embodiment 2:

[0060] A method for using a neonatal encephalopathy prediction model, such as Figure 2 As shown, the following steps are included:

[0061] R1: After the birth of a newborn baby at risk of encephalopathy, MRI images and neonatal peripheral images are taken;

[0062] R2: Generate T1-weighted images, T2-weighted images and DWI images from the obtained MRI images, and use software to process the generated T1-weighted images, T2-weighted images, DWI images and neonatal peripheral images;

[0063] R3: Import the processed image into the loaded TensorFlow model for prediction;

[0064] R4: Combine the model output with medical expertise to determine whether the newborn is at risk of disease and the type of encephalopathy that may occur.

[0065] Using the above method for prediction, MRI images can be taken immediately after the birth of the newborn, which can identify newborns at risk of encephalopathy at an early stage and take timely intervention measures, which may improve the prognosis;

[0066] By generating T1-weighted images, T2-weighted images, and DWI images, and combining them with images of the newborn's body, the model can obtain information on the newborn's brain structure, function, and metabolism from multiple angles, helping to reduce misdiagnosis and missed diagnosis. It helps the medical team make decisions faster and provide more timely treatment for newborns, greatly reducing diagnosis time. As time goes by, data accumulates, and technology advances, the deep learning model can be continuously optimized and updated to provide more reliable brain disease prediction services.

[0067] The above is a preferred specific implementation manner of the present invention, and the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the field within the technical scope disclosed by the present invention in combination with the prior art or public common sense, within the spirit and principle of the present invention, shall be covered by the protection scope of the present invention.

Claims

1. A neonatal encephalopathy prediction model method, characterized in that: The steps include: S1: Collect image data of normal newborns and sick newborns, including whole body images and MRI image data; S2: Generate T1-weighted images, T2-weighted images and DWI images for MRI image data; S3: preprocessing the generated image and the surrounding image, including denoising and normalization; S4: Use the U-Net model to perform deep learning training on the preprocessed images and build a prediction model; S5: Introduce the attention mechanism in the U-Net model to focus on important areas in the image; S6: Train, validate and test the model; S7: Deploy the tested TensorFlow to hospital equipment.

2. A neonatal encephalopathy prediction model method according to claim 1, characterized in that: In the image data collected in step S1, the ratio of normal newborns, newborns with neonatal bilirubin encephalopathy and newborns with neonatal hypoxic-ischemic encephalopathy is 1:1:1, and the ratio of male to female newborns is 1:

1.

3. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: In step S2, the MRI image data is adjusted to generate T1-weighted images, T2-weighted images and DWI images, and the high signals of different weighted images are used to comprehensively predict whether the newborn is sick.

4. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: In the step S3, the total variation regularization denoising method is used to denoise the T1-weighted image, the T2-weighted image and the DWI image, retaining the high signal area, and the non-local mean denoising method is used to process the whole body image of the child, retaining the pathological characteristics of yellow skin and pale cyanosis.

5. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: In the step S3, the denoised T1-weighted image, T2-weighted image and DWI image are normalized by homogeneous standardization.

6. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: In the S4 step, semantic segmentation of the MRI images and neonatal peripheral images is performed through annotation tools, different tissue structures in the images are annotated, key points such as the lesion center can be specifically annotated, and corresponding labels are assigned to them, and a labeled image (mask) is subsequently generated for model training.

7. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: In the step S5, the channels of the feature map of the high signal area of ​​the patient in the feature map are weighted to highlight important channel information, and the spatial position of the high signal area of ​​the patient in the feature map is weighted to highlight important spatial information.

8. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: During the model training process in step S6, the Adam optimizer is used, the learning rate is 0.0009, the batch size is 64, the epochs is 500, and L2 regularization is used to prevent overfitting.

9. A neonatal encephalopathy prediction model method according to claim 2, characterized in that: The model deployed in the hospital device in step S7 can provide prediction services through TensorFlow Serving and can be accessed through a RESTful API or a gRPC interface.

10. A method for using a neonatal encephalopathy prediction model according to any one of claims 1 to 9, characterized in that: The steps include: R1: After the birth of a newborn baby at risk of encephalopathy, MRI images and full-body images of the newborn baby are taken to obtain prediction data: R2: Generate T1-weighted images, T2-weighted images and DWI images from the obtained MRI images, and use software to process the generated T1-weighted images, T2-weighted images, DWI images and neonatal peripheral images; R3: Import the processed image into the loaded TensorFlow model for prediction; R4: Combine the model output with medical expertise to determine whether the newborn is at risk of disease and the type of encephalopathy that may occur.

Images