A method for predicting the risk of cerebral palsy in children with BIPI

Abstract

The present invention provides a method for predicting the risk of cerebral palsy in children with BIPI. The method comprises: obtaining a data set; the data set is a brain MRI image of a child with BIPI; building an automatic lesion segmentation model for the child with BIPI; obtaining a lesion segmentation result; constructing a KD grading prediction model; obtaining a Kidokoro score result; constructing a lesion feature extraction network model; automatically extracting the image features of the brain MRI image of the child with BIPI corresponding to the data set; constructing a cerebral palsy risk prediction model for the child with BIPI, and using the quantitatively converted GMs evaluation result as the output of the prediction model; integrating the image features of the brain MRI image corresponding to the data set, the quantitatively converted Kidokoro score result, and the perinatal clinical index data into a combination, and using the combination as the input of the prediction model; obtaining a cerebral palsy risk prediction result for the child with BIPI. The cerebral palsy risk prediction model for the child with BIPI of the present invention incorporates multidimensional factors to comprehensively evaluate the risk of cerebral palsy in the child with BIPI, innovatively improving the accuracy and practicality of early prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of cerebral palsy risk prediction, and in particular to a method for predicting the risk of cerebral palsy in children with BIPI. Background Art

[0002] Brain Injury in Premature Infants (BIPI) refers to varying degrees of cerebral ischemia and / or hemorrhagic and inflammatory damage in premature infants caused by various pathological factors during the perinatal and intrapartum periods, resulting in clinically manifested symptoms and signs. Severe cases can lead to neurological sequelae or even death. As the overall survival rate of premature infants increases, the incidence of cerebral palsy in children with BIPI has increased annually. Research statistics in my country indicate that cerebral palsy occurs in 29.13‰ of newborns due to complications of prematurity. Prevention and early intervention are key to reducing the risk of cerebral palsy, but early diagnosis is often difficult.

[0003] The degree of brain injury in children with BIPI is closely related to the risk and prognosis of cerebral palsy. The degree of brain injury in children with BIPI is positively correlated with clinical prognosis, but the lack of corresponding clinical signs and indicators for determining the degree of brain injury in children with BIPI has become a major clinical problem. Studies by Kang O, MNB, Kekky L, and others have confirmed that the MRI scoring system can quantitatively determine the degree of brain injury and assess brain developmental maturity, and has a good predictive value for the long-term prognosis of children with BIPI, including cognitive, motor, and language outcomes. The traditional MRI scoring system only defines the scoring criteria for grading the degree of brain injury in full-term infants with hypoxic-ischemic encephalopathy (HIE). The two most commonly used scoring systems are the Barkovich score and the National Institute of Child Health and Human Development Neonatal Research Network (NICHD NRN) score. These two scoring systems have significant limitations and only analyze damage to the brain's WM and deep gray matter structures. Damage to other areas is not included in the scoring criteria. The Kidokoro (KD) scoring system not only covers a comprehensive range of injury areas and provides detailed scoring assessments of the patterns of each injury area, but also incorporates the development of cortical folds, myelination maturity, and volume changes of important tissue structures after injury into the scoring system. The most critical thing is that the validation group of the KD scoring system is applicable to premature infants. Therefore, the KD scoring system is more comprehensive, objective, and applicable to a wider range of groups in clinical application. In addition, the accuracy of the KD scoring system in determining the degree of injury has been clinically verified and recognized.

[0004] Although the KD scoring system has high clinical application value, the high complexity of the KD scoring system and the high time-consuming nature of manual scoring pose a great challenge to scoring users. The KD scoring system requires a senior pediatric neuroimaging expert who is professional and familiar with the scoring criteria to complete, and the brain area structure measurement work in the KD scoring system is time-consuming. It takes 10-15 minutes to complete a complete scoring of a child. Faced with the huge clinical workload, the clinical popularization of the KD scoring system is difficult to achieve and is limited to children with special conditions and urgent needs. The emergence of artificial intelligence makes it possible to break through this clinical bottleneck. At present, although there are more and more studies on premature infants, there have been no reports on the construction of an intelligent grading AI model for the degree of injury based on the scoring system.

[0005] Furthermore, neurodevelopmental assessment and follow-up are the primary clinical methods for the diagnosis of cerebral palsy and the evaluation of the efficacy of interventions for neurodevelopmental abnormalities. Traditional clinical assessment tools include the Brazelton Neonatal Behavioral Score, the Amiel-Tison Neonatal Neurological Function Test, the Bayley Scales of Infant Development, the 20-Item Neonatal Behavioral Neurological Assessment (NBNA), and many others. However, these assessment tools suffer from limitations such as lengthy time consumption, inconsistent accuracy, and delayed diagnosis. In particular, diagnosis is often not confirmed until the child is two years old, limiting the optimal timing for early rehabilitation. General movements (GMs) assessment, an internationally recognized new movement assessment method, is clinically recommended for its accuracy, simplicity, non-invasiveness, and high predictive power. However, although the GMs assessment method is highly accurate, its results only rely on the clinical signs that have already appeared in the child, and require professional GMs assessors to conduct the assessment. In addition, the earliest time to obtain a high-risk warning for cerebral palsy and movement disorders in children through clinical GMs assessment is 5 months after birth, and for many children the time is extended to 1 year. Although GMs predicts cerebral palsy earlier than other assessment tools, the best time for early intervention of brain damage in premature infants is within 1 month after birth. During this period, the child's brain plasticity is more likely to occur and the clinical intervention results are best.

[0006] Therefore, in view of the series of defects in the relevant technologies for predicting the risk of cerebral palsy in children with BIPI, we actually need a method for early prediction of the risk of cerebral palsy in children with BIPI in clinical practice. Summary of the Invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0008] Therefore, the purpose of the present invention is to propose a method for predicting the risk of cerebral palsy in children with BIPI.

[0009] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a method for predicting the risk of cerebral palsy in children with BIPI, and the method for predicting the risk of cerebral palsy in children with BIPI comprises: step S1: acquiring a data set; wherein, the data set is a brain MRI image of a child with BIPI; the lesion regions of interest in the brain MRI image have been marked; step S2: preprocessing the data set; step S3: randomly dividing the preprocessed data set into a first training set and a first test set; step S4: building an automatic lesion segmentation model for children with BIPI; wherein, the automatic lesion segmentation model for children with BIPI is a first deep learning segmentation model; step S5: inputting the first training set into the segmentation model for Training and optimization, saving the model parameters corresponding to the first training set; step S6: inputting the first test set into the trained segmentation model, and automatically obtaining the lesion segmentation result corresponding to the first test set; step S7: scoring the brain MRI images of the BIPI children in the data set according to the Kidokoro scoring system standard to obtain the Kidokoro scoring result corresponding to the data set; step S8: constructing a KD grading prediction model; wherein the KD grading prediction model is a second deep learning segmentation model; step S9: using the lesion segmentation result of the data set as the input of the KD grading prediction model, and using the Kidokoro scoring result corresponding to the data set as the Label, and randomly divide the lesion segmentation results of the data set into a second training set and a second test set; step S10: input the second training set into the KD grading prediction model for training and optimization, and save the model parameters corresponding to the second training set; step S11: input the second test set into the trained KD grading prediction model, and automatically obtain the Kidokoro score result corresponding to the second test set; step S12: construct a lesion feature extraction network model; wherein the lesion feature extraction network model is a convolutional neural network; step S13: when the lesion segmentation results of the data set are input into the lesion feature extraction network model, the brain MRI image of the BIPI child corresponding to the data set is automatically extracted image features; step S14: obtaining perinatal clinical indicator data of BIPI children corresponding to the data set; step S15: classifying and grading the perinatal clinical indicator data; step S16: obtaining the GMs evaluation results corresponding to the data set; step S17: performing quantitative conversion on the Kidokoro score results corresponding to the data set, the classified and graded perinatal clinical indicator data, and the GMs evaluation results; step S18: constructing a cerebral palsy risk prediction model for BIPI children, and using the quantitatively converted GMs evaluation results as the output of the cerebral palsy risk prediction model for BIPI children; wherein the cerebral palsy risk prediction model for BIPI children is a third deep learning segmentation model;Step S19: Integrate the image features of the brain MRI images of BIPI children corresponding to the dataset, the quantified Kidokoro score results, and the quantified perinatal clinical indicator data into a combination, and randomly divide the combination into a third training set and a third test set; Step S20: Input the third training set into the BIPI children's cerebral palsy risk prediction model for training and optimization, and save the model parameters; Step S21: Input the third test set into the trained BIPI children's cerebral palsy risk prediction model to automatically obtain the BIPI children's cerebral palsy risk prediction results corresponding to the third test set.

[0010] Preferably, the sequence of brain MRI images includes one of the following or a combination thereof: T1WI, T2WI, DWI, and T2-FLAIR.

[0011] Preferably, the automatic lesion segmentation model for BIPI children is the Attention-UNet deep learning segmentation network; the KD grading prediction model is CNN; the lesion feature extraction network model is the three-dimensional convolutional neural network 3D-ResNet; and the cerebral palsy risk prediction model for BIPI children is the XGBoost model.

[0012] Preferably, the step S7 specifically includes: step S7.1: automatically segmenting the six structural areas in the brain MRI image corresponding to the data set, and automatically measuring the numerical values ​​corresponding to the six structural areas; wherein the numerical values ​​corresponding to the six structural areas are: biparietal lobe width, interhemispheric distance, corpus callosum thickness, longest diameter of the cerebellar hemisphere, widest diameter of the left and right lateral ventricular bodies, and basal ganglia area; step S7.2: based on the automatic segmentation results of the six structural areas and the automatic measurement results of the numerical values ​​corresponding to the six structural areas, the brain MRI images of the BIPI children in the data set are scored according to the Kidokoro scoring system standards to obtain the Kidokoro score results corresponding to the data set.

[0013] Preferably, the perinatal clinical indicator data include one of the following or a combination thereof: gestational age at birth, birth weight, gender, mode of delivery, APgar 1min score, APgar 5min score, cause of disease, C-reactive protein, and cerebrospinal fluid detection indicators.

[0014] Beneficial effects of the present invention:

[0015] The method provided by the present invention for predicting the risk of cerebral palsy in children with BIPI combines multi-dimensional factor information such as MRI imaging signs, KD grade, perinatal factors, clinical laboratory indicators, etc., and takes GMs evaluation results as the basis to innovatively develop and verify a risk prediction model for early prediction of cerebral palsy and developmental disorders in children with BIPI. The aim is to improve the accuracy and practicality of early prediction, identify high and low risk rates of cerebral palsy at an earlier stage, provide objective and quantifiable support for early clinical intervention and formulation of rehabilitation plans, and provide important help for parents of children with the disease to actively cooperate or reduce excessive anxiety.

[0016] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic flow chart showing a method for predicting the risk of cerebral palsy in children with BIPI according to an embodiment of the present invention is shown;

[0018] Figure 2 A schematic flowchart of an embodiment of the present invention for constructing a KD grading prediction model using a deep learning method based on a T1WI+T2WI combined sequence of MRI is shown;

[0019] Figure 3 A schematic flowchart of a method for constructing a KD grading prediction model based on BIPI lesion characteristics based on a deep learning method according to an embodiment of the present invention is shown;

[0020] Figure 4 A key schematic diagram showing an embodiment of the present invention, wherein an XGBoost model is used to extract lesion features through deep learning;

[0021] Figure 5 A schematic flow chart showing the construction of a cerebral palsy risk prediction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, Figures 1 to 5 As shown, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0024] Figure 1FIG1 is a schematic flow chart showing a method for predicting the risk of cerebral palsy in children with BIPI according to an embodiment of the present invention. Figure 1 As shown, the method for predicting the risk of cerebral palsy in children with BIPI includes: building an automatic lesion segmentation model for children with BIPI; automatically obtaining lesion segmentation results based on the automatic lesion segmentation model for children with BIPI; building a KD grading prediction model; automatically obtaining Kidokoro score results based on the KD grading prediction model; building a lesion feature extraction network model; automatically extracting image features of brain MRI images of children with BIPI based on the lesion feature extraction network model; building a cerebral palsy risk prediction model for children with BIPI, and using the quantitatively converted GMs evaluation results as the output of the cerebral palsy risk prediction model for children with BIPI; integrating the image features of the brain MRI images of children with BIPI corresponding to the data set, the quantitatively converted Kidokoro score results, and the quantitatively converted perinatal clinical indicator data into a combination, and using the combination as the input of the cerebral palsy risk prediction model for children with BIPI; and automatically obtaining the cerebral palsy risk prediction results for children with BIPI.

[0025] In this embodiment, the method provided by the present invention for predicting the risk of cerebral palsy in children with BIPI is innovatively developed and verified by combining multi-dimensional factor information such as MRI imaging signs, KD grade, perinatal factors, clinical laboratory indicators, etc., based on GMs evaluation results. A risk prediction model for early prediction of cerebral palsy and developmental disorders in children with BIPI is designed to improve the accuracy and practicality of early prediction, identify high and low risk rates of cerebral palsy at an earlier stage, provide objective and quantifiable support for early clinical intervention and formulation of rehabilitation plans, and provide important help for parents of children with the disease to actively cooperate or reduce excessive anxiety.

[0026] In one embodiment of the present invention, a method for predicting the risk of cerebral palsy in children with BIPI specifically includes: step S1: obtaining a data set; wherein the data set is a brain MRI image of a child with BIPI; the lesion regions of interest in the brain MRI image have been marked; step S2: preprocessing the data set; step S3: randomly dividing the preprocessed data set into a first training set and a first test set; step S4: building an automatic lesion segmentation model for children with BIPI; wherein the automatic lesion segmentation model for children with BIPI is a first deep learning segmentation model; step S5: inputting the first training set into the segmentation model for training and optimization, and saving the data. The model parameters corresponding to the first training set are obtained; step S6: the first test set is input into the trained segmentation model to automatically obtain the lesion segmentation result corresponding to the first test set; step S7: the brain MRI images of the BIPI children in the data set are scored according to the Kidokoro scoring system standard to obtain the Kidokoro scoring result corresponding to the data set; step S8: a KD grading prediction model is constructed; wherein the KD grading prediction model is a second deep learning segmentation model; step S9: the lesion segmentation result of the data set is used as the input of the KD grading prediction model, and the Kidokoro scoring result corresponding to the data set is used as the The label of the data set, and randomly divide the lesion segmentation results of the data set into a second training set and a second test set; step S10: input the second training set into the KD grading prediction model for training and optimization, and save the model parameters corresponding to the second training set; step S11: input the second test set into the trained KD grading prediction model, and automatically obtain the Kidokoro score result corresponding to the second test set; step S12: construct a lesion feature extraction network model; wherein the lesion feature extraction network model is a convolutional neural network; step S13: when the lesion segmentation results of the data set are input into the lesion feature extraction network model, the lesion feature extraction network model is automatically obtained. Obtain the image features of the brain MRI images of the BIPI children corresponding to the dataset; Step S14: Obtain the perinatal clinical indicator data of the BIPI children corresponding to the dataset; Step S15: Classify and grade the perinatal clinical indicator data; Step S16: Obtain the GMs evaluation results corresponding to the dataset; Step S17: Perform quantitative conversion on the Kidokoro score results corresponding to the dataset, the classified and graded perinatal clinical indicator data, and the GMs evaluation results; Step S18: Construct a cerebral palsy risk prediction model for BIPI children; wherein the cerebral palsy risk prediction model for BIPI children is a third deep learning segmentation model;Step S19: Integrate the image features of the brain MRI images of BIPI children corresponding to the dataset, the quantified Kidokoro score results, the quantified perinatal clinical indicator data, and the quantified GMs assessment results into a combination, and randomly divide the combination into a third training set and a third test set; Step S20: Input the third training set into the BIPI children's cerebral palsy risk prediction model for training and optimization, and save the model parameters; Step S21: Input the third test set into the trained BIPI children's cerebral palsy risk prediction model to automatically obtain the BIPI children's cerebral palsy risk prediction results corresponding to the third test set.

[0027] In one embodiment of the present invention, the sequence of brain MRI images includes one of the following or a combination thereof: T1WI, T2WI, DWI, and T2-FLAIR.

[0028] In one embodiment of the present invention, the automatic lesion segmentation model for children with BIPI is the Attention-UNet deep learning segmentation network; the KD grading prediction model is CNN; the lesion feature extraction network model is the three-dimensional convolutional neural network 3D-ResNet; and the cerebral palsy risk prediction model for children with BIPI is the XGBoost model.

[0029] In one embodiment of the present invention, step S7 specifically includes: step S7.1: automatically segmenting the six structural areas in the brain MRI image corresponding to the data set, and automatically measuring the numerical values ​​corresponding to the six structural areas; wherein the numerical values ​​corresponding to the six structural areas are: biparietal lobe width, interhemispheric distance, corpus callosum thickness, longest diameter of the cerebellar hemisphere, widest diameter of the left and right lateral ventricular bodies, and basal ganglia area; step S7.2: based on the automatic segmentation results of the six structural areas and the automatic measurement results of the numerical values ​​corresponding to the six structural areas, scoring the brain MRI images of the BIPI children in the data set according to the Kidokoro scoring system standards to obtain the Kidokoro score results corresponding to the data set.

[0030] In one embodiment of the present invention, the perinatal clinical indicator data include one of the following or a combination thereof: gestational age at birth, birth weight, gender, mode of delivery, APgar 1min score, APgar 5min score, cause of disease, C-reactive protein, and cerebrospinal fluid detection indicators.

[0031] In this embodiment, the clinical GMs assessment results and the multimodal combination sequence lesion characteristics determined that the risk prediction model for cerebral palsy in children with BIPI was characterized by incorporating multidimensional factors for comprehensive evaluation.

[0032] In one embodiment of the present invention, Figure 5 As shown, the process of constructing the cerebral palsy risk prediction model of the present invention specifically includes: using the GMS score as a label and multimodal lesions as input to train the ResNet network, and converting other features (such as cause classification, etc.) into integer labels as clinical features; after using the GMS score as a label and multimodal lesions as input to train the ResNet network, the network feature extraction part is separated separately after fine-tuning; using a feature extractor to extract lesion features in MRI as deep learning features; combining the MRI deep learning features and clinical prognosis labels of the case, and dividing the sample into a training set, a test set, and a validation set; using the MRI deep learning features and clinical features of the sample as input, and high risk and low risk of clinical prognosis as labels, to construct an XGBoost regression model; using the network search method to search for the optimal hyperparameters, and using this set of hyperparameters to train the XGBoost model; after training, testing is performed on the test set to measure the prediction effect of the high and low risk grading model for clinical prognosis.

[0033] The following is a specific embodiment to illustrate the technical solution of the present invention. The specific implementation steps of the method for predicting the risk of cerebral palsy in children with BIPI are as follows:

[0034] (1) Automatically obtain the lesion segmentation results. Specifically, the following steps are included:

[0035] Step S1: Obtain a dataset; 507 infants diagnosed with BIPI in the Children's Hospital, Obstetrics and Gynecology, and Neonatal Departments of the Women and Children's Hospital of Dalian Women and Children's Medical Center (Group) between March 2019 and March 2023 were selected as research subjects, with a gestational age of 24 to 36 weeks; 307 infants from the Children's Hospital were used as the primary model construction data. The dataset consists of brain MRI images of infants with BIPI; the brain MRI image sequences include one or a combination of the following: T1WI, T2WI, DWI, and T2-FLAIR. T1WI is T1-weighted imaging (T1WI for short), T2WI is T2-weighted imaging (T2WI for short). DWI is diffusion-weighted imaging (DWI for short). T2-FLAIR is T2-weighted fluid-attenuated inversion recovery sequence (T2-FLAIR for short). The lesion regions of interest of the brain MRI image have been marked; step S2: preprocessing the data set; step S3: randomly dividing the preprocessed data set into a first training set and a first test set; step S4: building an automatic lesion segmentation model for children with BIPI; wherein, the automatic lesion segmentation model for children with BIPI is a first deep learning segmentation model; the automatic lesion segmentation model for children with BIPI is an Attention-UNet deep learning segmentation network; step S5: inputting the first training set into the segmentation model for training and optimization, and saving the model parameters corresponding to the first training set; step S6: inputting the first test set into the trained segmentation model, and automatically obtaining the lesion segmentation result corresponding to the first test set.

[0036] In the process of automatically obtaining the lesion segmentation results, the brain MRI image dataset of children with BIPI is first obtained, the dataset is preprocessed, and the preprocessed dataset is randomly divided into a first training set and a first test set; then an automatic lesion segmentation model for children with BIPI is built; the first training set is input into the segmentation model for training and optimization, and the model parameters corresponding to the first training set are saved; the first test set is input into the trained segmentation model, and the lesion segmentation results corresponding to the first test set are automatically obtained, achieving the purpose of using deep learning methods and performing early diagnosis and segmentation of lesions in children with BIPI based on conventional MRI sequences. Specifically, high-dimensional information features that cannot be recognized by the human eye are mined through deep learning technology, the clinical application value of conventional MRI sequences is improved, and the grassroots clinical applicability of the constructed automatic lesion segmentation model for children with BIPI is improved, providing children with BIPI with fast and reliable intelligent auxiliary diagnosis, and also laying a good foundation for imaging diagnostic physicians to improve diagnostic accuracy, reduce diagnostic time, and solve multiple problems such as the examination burden on children with BIPI.

[0037] (2) Automatically obtain the Kidokoro scoring results. Specifically, the following steps are included:

[0038] Step S7: Score the brain MRI images of the BIPI children in the data set according to the Kidokoro scoring system standards to obtain the Kidokoro scoring results corresponding to the data set; Step S8: Construct a KD grading prediction model; wherein the KD grading prediction model is a second deep learning segmentation model; Step S9: Use the lesion segmentation results of the data set as the input of the KD grading prediction model, use the Kidokoro scoring results corresponding to the data set as the label of the data set, and randomly divide the lesion segmentation results of the data set into a second training set and a second test set; Step S10: Input the second training set into the KD grading prediction model for training and optimization, and save the model parameters corresponding to the second training set; Step S11: Input the second test set into the trained KD grading prediction model to automatically obtain the Kidokoro scoring results corresponding to the second test set.

[0039] The step S7 specifically includes: step S7.1: automatically segmenting the six structural areas in the brain MRI image corresponding to the data set, and automatically measuring the numerical values ​​corresponding to the six structural areas; wherein the numerical values ​​corresponding to the six structural areas are: biparietal lobe width, interhemispheric distance, corpus callosum thickness, longest diameter of the cerebellar hemisphere, widest diameter of the left and right lateral ventricular bodies, and basal ganglia area; step S7.2: based on the automatic segmentation results of the six structural areas and the automatic measurement results of the numerical values ​​corresponding to the six structural areas, scoring the brain MRI images of the BIPI children in the data set according to the Kidokoro scoring system standards to obtain the Kidokoro score results corresponding to the data set.

[0040] In this embodiment, before automatically segmenting the six structural regions in the brain MRI image corresponding to the data set and automatically measuring the values ​​corresponding to the six structural regions, it is necessary to manually measure the measurement values ​​of the six important structural regions. The values ​​corresponding to the six structural regions are: biparietal width (FLV), interhemispheric distance (IF), thickness of the corpus callosum (CC) at three locations (genu, body, and splenium), the longest diameter of the cerebellar hemisphere (CV), the widest diameter of the left and right lateral ventricle bodies (LV), and the basal ganglia area (TCV) (Table 1).

[0041] Table 1 Manual measurement methods of distances and areas of six important structural areas

[0042]

[0043] The measurements were performed by two associate senior physicians with 10 years of experience in pediatric MRI diagnosis. Physician 1 first performed manual measurements to obtain the values, and then Physician 2 corrected the measurement site and measured again to obtain the values. The average of the final measurement results of Physicians 1 and 2 was taken as the final measurement value.

[0044] Furthermore, the six structural regions in the brain MRI image corresponding to the dataset are automatically segmented, and the values ​​corresponding to the six structural regions are automatically measured. In a specific embodiment, the following describes the method for automatic segmentation and automatic measurement of the six structural regions. The six structural regions in the brain MRI image corresponding to the dataset are respectively labeled.

[0045] Use 3D-Unet network for segmentation. The specific segmentation and automatic measurement methods are as follows:

[0046] Segmentation method: First, the six structural regions corresponding to the original image are annotated and classified. Data augmentation is performed using intensity normalization, black background removal, rotation, and scaling. Small patches of randomly cropped (64, 64, 32) are input into the 3D-UNET. After downsampling using maxpooling, feature maps are obtained. A feature extraction layer is constructed using 3D convolution, batchnorm3D, and RELU layers. Upsampling is performed using transposed convolution, and skip connections are used in between to provide high-resolution features for the decoder. The final layer uses a 1x1x1 convolution to adjust the number of channels to obtain the segmentation result. The segmentation result and annotations are used to calculate the loss, and the gradient is backpropagated to adjust the network parameters. After training converges, the six structures are finally segmented on the test set.

[0047] Automatic measurement: We used Python's shapely library to write an automatic measurement program that can build computational geometry models of different brain regions and automatically calculate the dimensions of these regions. First, the T2WI and T1WI images required for measurement were preprocessed using slice interpolation (upsampled to 0.7x0.7x1mm 3 resolution) and N4 bias field correction. Using the FMRIB linear image registration tool (FMRIB PPREMO population template), the segmentation was converted to this space using the obtained transformation while maintaining consistency in size and shape. Among them, since the corpus callosum is a relatively thin area, the 97th percentile was selected to determine the distance to the thickest position of each part of the corpus callosum, reflecting a position similar to manual measurement. The automatic measurement process of the 6 structural areas is as follows:

[0048] IF measurement: Use the automatic measurement program to collect all boundary points in the IF structure, traverse and calculate the distance between two points with the same y coordinate, and then compare and select the maximum value as the inter-hemispherical distance (IF).

[0049] LV measurement: An automatic measurement program was used to collect all boundary points in the LV structure. All points on the left and right sides of a single LV structure were traversed separately, and the left and right diameters were measured. The widest diameter of the left LV was calculated as the widest diameter of the left ventricular body, and the widest diameter of the right LV was calculated as the widest diameter of the right ventricular body.

[0050] CV measurement: An automated measurement program was used to collect all boundary points in the cerebellar hemisphere. All points were traversed on both the left and right sides, and the maximum distance from the left to right border of the image was calculated as the longest diameter (CV) of the cerebellar hemisphere. FLV measurement: Similar to CV measurement, an automated measurement program was used to collect all boundary points in the FLV, and the maximum distance from the left to right border of the image was calculated as the biparietal width (FLV).

[0051] CC measurement: Use an automatic measurement program to collect all boundary points of the corpus callosum, and obtain the center point A of the medial knee, the boundary point B close to the lowest side of the body, and the boundary point C close to the lowest side of the splenium through coordinate comparison. The CC body, CC knee, and CC splenium are divided by these three points. By traversing the medial boundary points, the corresponding perpendicular lines between the medial boundary point and the lateral boundary point are found and intersected at the lateral point. The distance between the two points is calculated, and the distance is compared to obtain the body thickness, knee thickness, and splenium thickness.

[0052] Furthermore, based on the automatic segmentation results of the six structural regions and the automatic measurement results of the corresponding numerical values ​​of the six structural regions, the brain MRI images of the children with BIPI in the dataset were scored according to the Kidokoro scoring system to obtain the Kidokoro scoring results corresponding to the dataset. In the following specific embodiments, we will describe the scoring content and scoring criteria of the Kidokoro scoring system.

[0053] The MRI images of all enrolled BIPI children were scored according to the Kidokoro scoring system. The specific scoring items and result criteria are shown in Tables 2.1-2.4:

[0054] The Kidokoro scoring system is divided into four major categories: white matter abnormalities, cortical gray matter (GM) abnormalities, deep gray matter abnormalities, and cerebellar abnormalities. The assessment and scoring details for white matter abnormalities are shown in Table 2.1. The assessment and scoring details for cortical gray matter abnormalities are shown in Table 2.2. The assessment and scoring details for deep gray matter abnormalities are shown in Table 2.3. The assessment and scoring details for cerebellar structural abnormalities are shown in Table 2.4.

[0055] Table 2.1 Assessment content and scoring criteria for white matter abnormalities

[0056]

[0057] Note: FLV denotes the maximum width of the biparietal lobes in the coronal plane; LV denotes the width of the body of the lateral ventricle.

[0058] Table 2.2 Assessment content and scoring criteria for cortical gray matter abnormalities

[0059]

[0060] Note: IF represents the widest diameter of the interhemispheric distance measured by the single-layer axial view of the largest cross-section of the bilateral superior frontal gyrus

[0061] Table 2.3 Assessment content and scoring criteria for deep gray matter abnormalities

[0062]

[0063]

[0064] Note: TCV represents the maximum cross-sectional area of ​​the basal ganglia measured on the axial position with the caudate head and lentiform nucleus maximally visible.

[0065] Table 2.4 Assessment content and scoring criteria for cerebellar structural abnormalities

[0066]

[0067] Note: CV represents the widest diameter of the bilateral cerebellar hemispheres in the axial image.

[0068] The total score of brain WM abnormality is divided into 4 levels: none (total score 0-2 points), mild (total score 3-4 points), moderate (total score 5-6 points), and severe (total score >7).

[0069] The total score of cortical GM abnormality was divided into 4 levels: none (total score 0), mild (total score 1), moderate (total score 2), and severe (total score > 3);

[0070] Deep GM abnormalities and cerebellar abnormalities were each divided into 4 grades: none (total score 0 points), mild (total score 1 point), moderate (total score 2 points), and severe (total score > 3 points);

[0071] The total score of overall brain abnormality is divided into 4 grades: grade 0: normal (total score 0-3 points), grade 1: mild (total score 4-7 points), grade II: moderate (total score 8-11 points) and grade III: severe (total score >12).

[0072] Furthermore, the brain MRI images of BIPI children were scored by two deputy senior physicians with more than 10 years of rich experience in neonatal MRI diagnosis and familiar with the KD scoring rules according to the project content and details of the above-mentioned MRI scoring system. The scoring of each content was independently scored by two experts, and the final score result was based on the average of the scores of the two experts. Each child finally obtained the total score of the four individual contents of brain WM total score, cortical GM total score, depth GM total score, and cerebellum total score. The total score of the four contents was then summarized to obtain the total score of the overall brain abnormality of each child. This value was graded according to the scoring standard, and finally the grade of each child was obtained. This is for the subsequent use of the Kidokoro score results corresponding to the dataset as the label of the input dataset of the KD grading prediction model.

[0073] In a specific embodiment, the KD hierarchical prediction model is 3D-ResNet in CNN, such as Figure 2 and Figure 3As shown in the figure, the input image size is (128, 128, 32). 3D-ResNet contains multiple residual blocks. Each residual block (ResidualBlock) usually contains multiple convolutional layers, pooling layers and batch normalization layers. The first convolution layer of 3D-ResNet is used to process the input data. It usually uses a larger convolution kernel, such as 3x7x7 convolution, with a stride of 1. Inside the residual block, it usually contains multiple convolutional layers. These convolutional layers usually have smaller convolution kernels, such as 3x3x3 convolution. There will be a batch normalization layer after the convolution layer, which helps to accelerate training and help the model converge. Inside the residual block, the RELU activation function is often used to add nonlinearity to the model. The residual block contains skip connections, which are mainly used to solve the gradient disappearance and gradient explosion problems during the training process of deep neural networks, making deep neural networks easier to train. The skip connection adds the residual block input to the output of the last convolution layer. If the input size and output size are different, the size is adjusted through 1x1x1 convolution. After the residual block, a pooling layer is used to reduce the size of the feature map. After multiple residual blocks are stacked, the network uses a fully connected layer for classification.

[0074] The cross entropy loss function is used during training, and the cross entropy loss function is defined as:

[0075]

[0076] Where C is the total number of categories, y ic is an indicator function that is 1 when sample i belongs to category c and 0 otherwise. pic is the model's prediction of the probability that sample i belongs to category c. The Adam optimizer was used during training, with an initial learning rate of 0.001 and a cosine annealing learning rate adjustment method. The learning rate was set high in the early stages of training and gradually decreased in the later stages. This facilitates fine-tuning of model parameters later, improving model training effectiveness while achieving model convergence. The model was trained on the training set data, and performance on the validation set was monitored to adjust hyperparameters. Model accuracy, recall, F1 score, and other metrics were evaluated on the test set.

[0077] Using previously segmented lesions from T1WI single sequences and T1WI+T2WI combined sequences, 307 patients were randomly divided into groups in an 8:1:1 ratio. A 3D-ResNet deep learning network framework, based on the CNN framework, was selected with input image dimensions of (128, 128, 32). The 3D-ResNet model, which includes multiple residual blocks, was used to construct a prediction model for KD grade in premature infants with brain injury. Deep learning methods were found to be more advantageous in constructing the BIPI injury grade KD prediction model, with lesion features from the combined MRI T1WI+T2WI sequence performing best in model construction. This demonstrates the potential of AI technology to improve MRI diagnostic efficacy and reduce scanning costs for patients.

[0078] (3) Automatically extract the image features of the brain MRI images of children with BIPI corresponding to the dataset. Specifically, the following steps are included:

[0079] Step S12: Construct a lesion feature extraction network model; wherein, the lesion feature extraction network model is a convolutional neural network; the lesion feature extraction network model is a three-dimensional convolutional neural network 3D-ResNet; Step S13: When the lesion segmentation result of the data set is input into the lesion feature extraction network model, the image features of the brain MRI image of the BIPI child corresponding to the data set are automatically extracted.

[0080] (4) Step S14: Obtain the perinatal clinical indicator data of the BIPI infant corresponding to the data set.

[0081] Perinatal clinical indicator data include one or a combination of the following: gestational age, birth weight, sex, mode of delivery, APgar 1-minute score, APgar 5-minute score, etiology, C-reactive protein, and cerebrospinal fluid (CSF) test indicators. The cerebral palsy risk prediction model for children with BIPI is characterized by its comprehensive assessment of multiple factors, based on clinical GMs assessment results and multimodal combined sequence lesion characteristics.

[0082] (5) Step S15: Classify and grade the perinatal clinical indicator data.

[0083] The perinatal clinical indicator data were classified and graded, including: the clinical indicators were classified and graded according to the standards of "Expert Consensus on Diagnosis and Prevention of Brain Injury in Premature Infants", "Practical Prematurity and Premature Pediatrics", and "Practical Neonatology", as follows:

[0084] Gender: divided into two categories: male and female.

[0085] Birth weight: Premature infants were divided into 4 groups according to their birth weight: extremely low birth weight group (ELBWI): birth weight ≤ 1000g; very low birth weight group (VLBW): birth weight ≤ 1500g; low birth weight group (LBW): 1500 < birth weight < 2500g; normal birth weight group (NBW): birth weight ≥ 2500g.

[0086] Gestational age at birth: Gestational age at birth was divided into 4 groups: extremely premature group: gestational age at birth ≤ 28 weeks; early premature group: 28 weeks < gestational age at birth ≤ 31 weeks; mid-preterm group: 32 weeks ≤ gestational age at birth < 35 weeks; late premature group: 35 weeks ≤ gestational age at birth < 37 weeks.

[0087] Mode of delivery: divided into two categories: vaginal delivery and cesarean section.

[0088] Apgar score: The score results of two time periods are evaluated and graded according to the following standards:

[0089] 1min (Diagnosis and grading of asphyxia) Grade I: 0-3 points Grade II: 4-7 points Grade III: 8-10 points.

[0090] 5 minutes (to determine the resuscitation effect and prognosis) Grade I: 0-3 points Grade II: 4-7 points Grade III: 8-10 points

[0091] Etiology classification: The etiology of BIPI is divided into four categories: Category A (cerebral hypoxia-ischemia and hemodynamic disorder): including severe intrauterine distress and birth asphyxia, hypercapnia or hypocapnia, circulatory failure / shock, hypotension / hypertension / abnormal blood pressure fluctuations, heart failure, respiratory failure, mechanical ventilation, severe dehydration, hypothermia, intrauterine growth restriction, severe or complex congenital heart disease; Category B (infection and inflammatory response): including chorioamnionitis, intrauterine infection, postnatal infection, Infection, necrotizing enterocolitis, etc.: Category C (blood system diseases): including coagulation abnormalities, moderate to severe anemia, polycythemia-hyperviscosity, antithrombin deficiency, and plasminogen deficiency; Category D (obstetric high-risk factors): including thrombosis or amniotic fluid embolism, maternal complications / complications and bad habits (hypertension, heart disease, diabetes, severe anemia, smoking, drug abuse), and history of abnormal delivery (emergency cesarean section, placental abruption, forceps / vacuum delivery, shoulder dystocia, precipitous labor, and prolonged labor).

[0092] C-reactive protein (CRP): Group A (normal) 0-15 mg / L, Group B (abnormal) >15 mg / L.

[0093] Cerebrospinal fluid test (ADA, sugar, LDH): divided into 2 groups, Group A (normal) ADA < 40 (U / L), LDH < 240 U / L, sugar < 4.5 mmol / L; Group B (abnormal) ADA > 40 (U / L), LDH > 240 U / L, sugar > 4.5 mmol / L.

[0094] (6) Step S16: Obtain the GMs evaluation result corresponding to the data set.

[0095] Specifically, all children in the dataset underwent motor development assessments by specialized GMs assessors at a rehabilitation center. Furthermore, standardized videotaped assessments were used; recordings were made once for the first writhing phase between 3 days and full term, and once for the second restless writhing phase between 2 and 5 months of age. Those assessed as spasticity-synchrony (CS) and restless movement deficiency (F-) in the first two phases were reassessed 12 months after full term. Furthermore, two rehabilitation professionals with certification from the GM Trust training program and qualified assessment qualifications participated in the assessments. Each assessment was initially conducted by one assessor (therapist), followed by a second, more experienced assessor (rehabilitation therapist). The assessors used visual Gestalt perception to assess each GM performance, distinguishing between normal and abnormal subcategories for each phase. The auditory signals in the recordings were muted during the assessments. If multiple abnormal recordings were identified or if assessment difficulties arose, the perception was recalibrated using standard normal GMs recordings. The assessment of a single recording takes approximately 5 minutes.

[0096] The following describes the results of the GMs. Phase 1 (the writhing phase): Assessment results are categorized as normal or abnormal. Normal results are: normal writhing (N); abnormal results are: monotonous whole-body movements (PR), spastic-synchronous whole-body movements (CS), and chaotic whole-body movements (Ch). PR refers to a monotonous sequence of continuous movement components, with movements of different body parts lacking their normal complexity. CS refers to rigid movements with a loss of normal fluidity, with all limb and trunk muscles contracting and relaxing almost simultaneously. Ch refers to large movements of all limbs, chaotic sequences, loss of fluidity, and sudden, disjointed movements.

[0097] Stage 2 (restlessness stage): The assessment results are divided into normal and abnormal. Normal is: normal restless movement (NF), abnormal is: lack of restless movement (F-), abnormal restless movement (AF). NF is a small-amplitude, medium-speed movement that occurs throughout the neck, trunk and limbs, in all directions, with variable movement acceleration. In awake infants, this movement persists except when they are fussy and crying, and can exist at the same time as other movements. The frequency of restless movement changes with age. F- refers to the fact that no restless movement has been observed within weeks to months after full term, which is called "lack of restless movement". AF is moderately or significantly exaggerated in terms of movement amplitude, speed and unsteadiness.

[0098] Normal results in the first stage are considered well-developed, while abnormal results require follow-up to the second stage. Normal results in the second stage are considered well-developed, while abnormal results suggest cerebral palsy and motor delay. F-positive results suggest cerebral palsy, while AF-positive results suggest motor delay. The second stage results are used as the final outcome, and GMs are categorized into high- and low-risk groups. The low-risk group includes normal results (N and NF), while the high-risk group includes abnormal results (F- and AF).

[0099] (7) Step S17: Quantitatively convert the Kidokoro score results, the classified and graded perinatal clinical index data, and the GMs evaluation results corresponding to the data set.

[0100] The Kidokoro score results, the classified and graded perinatal clinical index data, and the GMs assessment results corresponding to the data set were quantitatively converted, specifically including:

[0101] Nine types of perinatal clinical index data were collected from infants with BIPI, including gestational age, birth weight, sex, mode of delivery, APgar 1-min score, APgar 5-min score, etiology, C-reactive protein (CRP), and cerebrospinal fluid tests (ADA, sugar, and LDH). According to the classification standards of the previous study, the data of 9 types of perinatal clinical indicators were classified and sorted, so that these indicators could be converted into positive integers as preparatory input variables of the XGBoost model. The four categories of gestational age at birth (A, B, C, D) were converted into positive integers 0, 1, 2, 3; the four categories of birth weight (A, B, C, D) were converted into positive integers 0, 1, 2, 3; the two categories of gender (A, B) were converted into positive integers 0, 1; the two categories of delivery mode (A, B) were converted into positive integers 0, 1; the three categories of APgar (1min and 5min) scores (A, B, C) were converted into positive integers 0, 1, 2; the four categories of main causes of disease (A, B, C, D) were converted into positive integers 0, 1, 2, 3; the two categories of C-reactive protein (CRP) (A, B) were converted into positive integers 0, 1; and the two categories of cerebrospinal fluid test (A, B) were converted into positive integers 0, 1. The four categories of the Kidokoro score (KD) for whole-brain abnormality (normal, mild damage, moderate damage, and severe damage) corresponding to the dataset are converted to positive integers 0, 1, 2, and 3. The GMS prognostic grading (low risk, high risk) of the GMs assessment results are converted to positive integer labels of 0 and 1, which are used as the final output low-risk and high-risk labels.

[0102] (8) Automatically obtain the cerebral palsy risk prediction results for children with BIPI. The specific steps include:

[0103] Step S18: Construct a cerebral palsy risk prediction model for children with BIPI, and use the quantitatively converted GMs evaluation results as the output of the cerebral palsy risk prediction model for children with BIPI; wherein, the cerebral palsy risk prediction model for children with BIPI is a third deep learning segmentation model; the cerebral palsy risk prediction model for children with BIPI is an XGBoost model; Step S19: Integrate the image features of the brain MRI images of BIPI children corresponding to the data set, the quantitatively converted Kidokoro score results, and the quantitatively converted perinatal clinical indicator data into a combination, and randomly divide the combination into a third training set and a third test set; Step S20: Input the third training set into the cerebral palsy risk prediction model for children with BIPI for training and optimization, and save the model parameters; Step S21: Input the third test set into the trained cerebral palsy risk prediction model for children with BIPI, and automatically obtain the cerebral palsy risk prediction results for children with BIPI corresponding to the third test set.

[0104] Table 3 Results of XGBoost model construction for multimodal cerebral palsy risk prediction of each combination sequence

[0105]

[0106] Table 3 shows the results of constructing the multimodal cerebral palsy risk prediction XGBoost model for each combination sequence. From Table 3, we can see that the multimodal cerebral palsy risk prediction XGBoost model constructed by combining T1WI+T2WI sequences is the optimal model.

[0107] After using the optimal T1WI+T2WI combined sequence segmentation of lesions completed in previous studies as input, we adopted a key strategy: freezing the parameters of the model classification layer (fully connected layer) and continuing to fine-tune it on the training set. After convergence, we carefully separated the feature extraction part of the model, including multi-layer convolution, downsampling pooling, and Batch Normalization (BN) layers. In other words, we intercepted the fully connected layer of the model, which is the part to the left of Figure FC (the part to the left of the dotted line), as shown in Figure 6. Figure 4 As shown in the figure, after separation, we obtained a separate feature extractor component, which we saved as a model. This component can simultaneously extract image features from MRI lesions. We applied this feature extractor to 307 segmented lesions, extracted features, and saved the extracted feature map matrix as a .npy file for use in the XGboost (Extreme Gradient Boosting) model. During this process, the lesion features were efficiently extracted using the deep learning feature extractor and used as one of the key features in XGBoost.

[0108] Table 4 Ablation experiment parameter results of the optimal combination T1WI+T2WI sequence cerebral palsy risk prediction model

[0109]

[0110]

[0111] Table 4 shows the ablation experimental parameter results of the optimal combination T1WI+T2WI sequence cerebral palsy risk prediction model. From Table 4, we can see that the cerebral palsy prediction model constructed with the participation of imaging symptoms, clinical features, and KD features has the best performance.

[0112] Table 5 shows the ablation experiment parameter results for the optimal combined T1WI+T2WI sequence cerebral palsy risk prediction model. As shown in Table 5, the model's effectiveness decreased after single-element ablation and continued with dual-element ablation. The Delong test showed statistically significant differences between the models after each element ablation and the initial model, indicating that single-dimensional indicators are flawed in assessing cerebral palsy prognosis and that a multidimensional approach should be considered. Furthermore, the Z-score values ​​for the remaining models constructed using only clinical and MRI imaging signs were relatively low, at 2.71 and 3.23, respectively, compared to the model constructed using a combination of multidimensional indicators. This indicates that clinical indicators have the greatest impact on the model, followed by MRI imaging signs, which are key indicators for BIPI cerebral palsy risk prediction.

[0113] Table 5 Delong test results of the model constructed by each factor variable after XGboost model ablation

[0114]

[0115] In summary, the method for predicting the risk of cerebral palsy in children with BIPI provided by the present invention combines multi-dimensional factor information such as MRI imaging signs, KD grade, perinatal factors, clinical laboratory indicators, etc., and takes GMs evaluation results as the basis to innovatively develop and verify a risk prediction model for early prediction of cerebral palsy and developmental disorders in children with BIPI. The purpose is to improve the accuracy and practicality of early prediction, identify high and low risk rates of cerebral palsy at an earlier stage, provide objective and quantifiable support for early clinical intervention and formulation of rehabilitation plans, and provide important help for parents of children with the disease to actively cooperate or reduce excessive anxiety.

[0116] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for predicting the risk of cerebral palsy in children with BIPI, characterized in that: include: Step S1: Acquire a data set; wherein the data set is a brain MRI image of a child with BIPI; the lesion regions of interest in the brain MRI image are all marked; Step S2: preprocessing the data set; Step S3: randomly divide the preprocessed data set into a first training set and a first test set; Step S4: Building an automatic segmentation model for lesions in children with BIPI; wherein the automatic segmentation model for lesions in children with BIPI is a first deep learning segmentation model; Step S5: inputting the first training set into the segmentation model for training and optimization, and saving the model parameters corresponding to the first training set; Step S6: inputting the first test set into the trained segmentation model to automatically obtain the lesion segmentation result corresponding to the first test set; Step S7: scoring the brain MRI images of the BIPI children in the dataset according to the Kidokoro scoring system standard to obtain the Kidokoro scoring result corresponding to the dataset; Step S8: constructing a KD hierarchical prediction model; wherein the KD hierarchical prediction model is a second deep learning segmentation model; Step S9: using the lesion segmentation results of the data set as input to the KD grading prediction model, using the Kidokoro score results corresponding to the data set as labels of the data set, and randomly dividing the lesion segmentation results of the data set into a second training set and a second test set; Step S10: inputting the second training set into the KD hierarchical prediction model for training and optimization, and saving the model parameters corresponding to the second training set; Step S11: inputting the second test set into the trained KD grading prediction model to automatically obtain the Kidokoro score result corresponding to the second test set; Step S12: constructing a lesion feature extraction network model; wherein the lesion feature extraction network model is a convolutional neural network; Step S13: when the lesion segmentation result of the data set is input into the lesion feature extraction network model, the image features of the brain MRI image of the BIPI child corresponding to the data set are automatically extracted; Step S14: Obtaining perinatal clinical indicator data of BIPI infants corresponding to the data set; Step S15: classifying and grading the perinatal clinical indicator data; Step S16: Obtaining the GMs evaluation result corresponding to the data set; Step S17: performing quantitative conversion on the Kidokoro score results, the classified and graded perinatal clinical index data, and the GMs assessment results corresponding to the data set; Step S18: constructing a BIPI children's cerebral palsy risk prediction model, and using the quantitatively converted GMs assessment results as the output of the BIPI children's cerebral palsy risk prediction model; wherein the BIPI children's cerebral palsy risk prediction model is a third deep learning segmentation model; Step S19: Integrate the image features of the brain MRI images of the BIPI children corresponding to the data set, the quantified converted Kidokoro score results, and the quantified converted perinatal clinical indicator data into a combination, and randomly divide the combination into a third training set and a third test set; Step S20: inputting the third training set into the BIPI children's cerebral palsy risk prediction model for training and optimization, and saving the model parameters; Step S21: inputting the third test set into the trained BIPI children's cerebral palsy risk prediction model to automatically obtain the BIPI children's cerebral palsy risk prediction result corresponding to the third test set.

2. The method for predicting the risk of cerebral palsy in children with BIPI according to claim 1, characterized in that: The sequence of brain MRI images includes one of the following or a combination thereof: T1WI, T2WI, DWI, and T2-FLAIR.

3. The method for predicting the risk of cerebral palsy in children with BIPI according to claim 1, characterized in that: The automatic segmentation model for lesions in children with BIPI is the Attention-UNet deep learning segmentation network; The KD classification prediction model is CNN; The lesion feature extraction network model is a three-dimensional convolutional neural network 3D-ResNet; The cerebral palsy risk prediction model for children with BIPI is the XGBoost model.

4. The method for predicting the risk of cerebral palsy in children with BIPI according to claim 1, characterized in that: The step S7 specifically includes: Step S7.1: Automatically segmenting six structural regions in the brain MRI image corresponding to the dataset and automatically measuring the values ​​corresponding to the six structural regions; wherein the values ​​corresponding to the six structural regions are: biparietal lobe width, interhemispheric distance, corpus callosum thickness, longest diameter of the cerebellar hemisphere, widest diameter of the left and right lateral ventricular bodies, and basal ganglia area; Step S7.2: Based on the automatic segmentation results of the six structural regions and the automatic measurement results of the corresponding numerical values ​​of the six structural regions, the brain MRI images of the BIPI children in the dataset are scored according to the Kidokoro scoring system standards to obtain the Kidokoro score results corresponding to the dataset.

5. The method for predicting the risk of cerebral palsy in children with BIPI according to claim 4, characterized in that: The perinatal clinical indicator data include one or a combination of the following: gestational age at birth, birth weight, gender, mode of delivery, APgar 1min score, APgar 5min score, cause of disease, C-reactive protein, and cerebrospinal fluid detection indicators.

Images