Posterior fossa tumor hydrocephalus prediction method and system, medium and computer equipment

Abstract

The invention relates to a child posterior fossal tumor postoperative hydrocephalus prediction method and system based on artificial intelligence, a storage medium and computer equipment. The invention provides a method for predicting hydrocephalus after a child posterior fossal tumor operation based on artificial intelligence, and the method comprises the steps: S1, obtaining and preprocessing a preoperative MRI image, and unifying the data of the preoperative MRI image to a standard state; s2, automatically segmenting a key region, positioning a rear fossa tumor focus and an associated region, and automatically segmenting the key region in the rear fossa tumor focus and the associated region through a deep learning model; s3, multi-dimensional feature extraction: extracting multi-dimensional features from the segmentation result of the step S2; s4, performing XGBoost prediction analysis and verification, quantitatively evaluating and predicting the risk of hydrocephalus after the fossa tumor operation of the child through an XGBoost machine learning model, and outputting a hydrocephalus prediction result in a quantitative mode. The invention further discloses a prediction system, a storage medium and computer equipment.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence (AI), and more particularly to a prediction method, prediction system, storage medium, and computer device for predicting the risk of postoperative hydrocephalus in children with posterior fossa tumors using deep learning and machine learning technologies based on MRI images of the brain. Background Technology

[0002] Posterior fossa tumors are a common type of intracranial tumor in children, and hydrocephalus is a particularly common and serious postoperative complication. Hydrocephalus refers to the abnormal accumulation of cerebrospinal fluid in the ventricles, leading to ventricular enlargement and potentially increased intracranial pressure, severely impacting the child's prognosis and quality of life. Currently, clinicians mainly rely on experience and preoperative imaging characteristics to assess the risk of postoperative hydrocephalus in children, lacking objective and quantitative predictive tools.

[0003] In existing technologies, the prediction of postoperative hydrocephalus after posterior fossa tumor surgery mainly suffers from the following problems:

[0004] (1) The prediction method relies too much on the doctor's subjective experience and lacks quantitative indicators and objective evaluation standards;

[0005] (2) Most existing studies focus on a single factor (such as tumor volume or location) and neglect the comprehensive analysis of multi-dimensional features;

[0006] (3) There is a lack of specific predictive models for pediatric posterior fossa tumors, and adult models (if available) are difficult to apply directly to pediatric patients with specific posterior fossa tumor characteristics.

[0007] (4) Traditional machine learning methods have limited effectiveness in processing high-dimensional medical image data, and the feature extraction process is cumbersome and highly subjective.

[0008] (5) Existing prediction systems lack clinical applicability and are difficult to integrate into doctors' diagnosis and treatment processes to provide immediate decision support.

[0009] Therefore, with the development and changes in technology and applications, it is of great significance to develop an artificial intelligence-based method and system for predicting postoperative hydrocephalus in children with posterior fossa tumors. This system can automatically extract multi-dimensional features from preoperative MRI images and perform comprehensive analysis and prediction by embedding pre-trained deep learning and machine learning algorithms. This will improve the diagnosis and treatment effect and the prognosis of children, and at the same time make it possible to overcome other technical defects and problems in existing technologies.

[0010] The information included in this background section of the specification of this invention, including any references cited herein and any descriptions or discussions thereof, is included for technical reference purposes only and is not intended to limit the scope of the invention. Summary of the Invention

[0011] The present invention is proposed in view of the foregoing and other further ideas. The present invention aims to solve the above-mentioned technical deficiencies and other technical problems.

[0012] The technical problems to be solved by this invention include: in order to solve the technical problems and defects of existing methods for predicting postoperative hydrocephalus in children with posterior fossa tumors, such as strong subjectivity, limited accuracy and precision, lack of quantitative indicators and multi-dimensional analysis, this invention proposes an artificial intelligence-based method, system and storage medium for predicting postoperative hydrocephalus in children with posterior fossa tumors.

[0013] According to one aspect of the present invention, the technical solution of the artificial intelligence-based method for predicting postoperative hydrocephalus in children with posterior fossa tumors includes at least the following key steps:

[0014] Step 1: Preoperative MRI image acquisition and preprocessing

[0015] - Receives brain MRI image data in DICOM format.

[0016] Image preprocessing includes voxel resampling, intensity normalization, and noise filtering.

[0017] - Convert the processed data into a standard format for easier subsequent analysis.

[0018] Step 2: Automatic segmentation of key regions

[0019] - Automatic segmentation of posterior fossa tumor regions using deep learning models

[0020] - Achieve automatic identification and quantification of hydrocephalus areas

[0021] -Automatic detection and segmentation of seepage areas

[0022] -Segmentation results were post-processed morphologically to optimize boundary accuracy.

[0023] Step 3: Multi-dimensional feature extraction

[0024] - Radiomics feature extraction, including:

[0025] (i) Morphological characteristics (including multiple items, such as 14 items): volume, surface area, sphericity, etc.;

[0026] (ii) Intensity characteristics (including multiple items, such as 17 items): mean, variance, skewness, kurtosis, etc.;

[0027] (iii) Texture features (including multiple features, such as 73 features): GLCM, GLRLM, GLSZM and other matrix features

[0028] - Quantitative feature extraction of hydrocephalus

[0029] -Feature extraction and quantification of brain midline shift

[0030] Step 4: Multi-model fusion predictive analysis

[0031] - Input the features extracted in steps 2 and 3 into the trained XGBoost machine learning model.

[0032] - Perform normalization and other processing on the features.

[0033] - Output hydrocephalus risk prediction results

[0034] Step 5: Visualization and Output of Prediction Results

[0035] - Generate intuitive forecast reports

[0036] -Visualize key impact features

[0037] - Output standardized result file

[0038] According to another aspect of the present invention, the architecture of the artificial intelligence-based prediction system for postoperative hydrocephalus in children with posterior fossa tumors includes, for example, the following core modules: a data management module, responsible for the import, preprocessing, and management of MRI image data; an automatic segmentation module, enabling accurate segmentation of the posterior fossa tumor and related regions; a feature extraction module, extracting multi-dimensional feature information from the segmentation results; a brain midline annotation module, enabling semi-automatic quantitative analysis of brain midline deviation; a prediction analysis module, integrating multiple features to achieve accurate prediction of postoperative hydrocephalus risk; and a user interaction module, providing an intuitive and user-friendly interface and result display functions.

[0039] More specifically, according to one aspect of the present invention, an artificial intelligence-based method for predicting postoperative hydrocephalus in children with posterior fossa tumors is provided. The prediction method includes the following steps: Step S1: Receiving preoperative MRI image data of the child's brain and preprocessing the MRI image data. The preprocessing includes voxel resampling, intensity standardization, and noise filtering of the MRI image data, unifying the MRI image data to a uniform spatial resolution, intensity range, and standardized data format, providing a standardized data foundation for subsequent automatic segmentation of key regions and multi-dimensional feature extraction; Step S2: Automatic segmentation of key regions, locating the posterior fossa tumor lesion and related regions, and using deep learning... The model automatically segments multiple key regions within the posterior fossa tumor lesion and associated areas; Step S3: Multi-dimensional feature extraction, used to extract multi-dimensional, quantifiable image features from the segmentation results of Step S2, which reflect the imaging characteristics of the posterior fossa tumor lesion and associated areas, and quantitatively analyzes the image features, wherein the image features include the volume, shape, signal intensity distribution, and spatial location of the posterior fossa tumor lesion and associated areas; and Step S4: XGBoost prediction analysis and validation, which uses the XGBoost machine learning model to quantitatively assess and predict the risk of postoperative hydrocephalus in children with posterior fossa tumors, and outputs the prediction results of hydrocephalus in a quantitative manner.

[0040] According to an embodiment of the present invention, the method further includes step S5: visualization and output of the prediction results, used to intuitively present the results of quantitative assessment of hydrocephalus.

[0041] According to an embodiment of the present invention, step S2 further includes: automatic segmentation of the posterior fossa tumor region; joint segmentation of the ventricular effusion region and the exudative region, which includes automatic identification, quantification and segmentation of the ventricular effusion region, and automatic detection and segmentation of the exudative region; morphological post-processing of the segmentation results to optimize boundary accuracy; step S3 further includes: radiomics feature extraction, hydrocephalus quantification feature extraction and brain midline shift feature extraction; step S4 further includes: inputting the features extracted in step S3 into the XGBoost machine learning model, and outputting the probability of predicting the hydrocephalus in a quantified manner, wherein step S4 further includes extracting the feature importance of the XGBoost model and identifying the key features that have the greatest impact on the prediction.

[0042] According to an embodiment of the present invention, in step S2, the automatic segmentation of the posterior fossa tumor region adopts an improved U-Net deep learning network architecture as a deep learning model, wherein the encoder uses a ResNet50 backbone network to extract features and incorporates an attention mechanism; the joint segmentation of the ventricular effusion region and the exudation region adopts a 3D UNet network with a deep supervision mechanism as a deep learning model.

[0043] According to an embodiment of the present invention, in step S3, the radiomics feature extraction includes: extracting morphological features, including volume, surface area, and sphericity; extracting intensity features, including mean, variance, skewness, and kurtosis; extracting texture features, including matrix features GLCM, GLRLM, and GLSZM; the hydrocephalus quantification feature extraction includes: calculating ventricular volume based on segmentation results; quantifying the volume of the hydrocephalus region and the exudation region; calculating the relative proportions of different parts of the ventricle; the brain midline offset feature extraction includes: performing semi-automatic annotation of the brain midline; calculating the average, maximum, and coefficient of variation of the brain midline offset; and analyzing the differences in brain midline offset at different levels.

[0044] According to another aspect of the present invention, an artificial intelligence-based prediction system for postoperative hydrocephalus in children with posterior fossa tumors is provided, comprising: a data management module configured for importing, preprocessing, and managing preoperative MRI image data, unifying the preoperative MRI image data into a standardized data standard through voxel resampling, intensity standardization, and format conversion, providing a consistent data foundation for subsequent processing; an automatic segmentation module configured for locating the posterior fossa tumor lesion and associated regions, and automatically segmenting multiple key regions within the posterior fossa tumor lesion and associated regions using a deep learning model; and a feature extraction module configured for extracting multi-dimensional, quantifiable image features from the automatic segmentation results, reflecting the posterior fossa... The system includes: imaging features of the tumor lesion and associated regions, and quantitative analysis of these imaging features, including the volume, shape, signal intensity distribution, and spatial location of the posterior fossa tumor lesion and associated regions; a brain midline annotation module configured to achieve semi-automatic brain midline annotation through automatic detection algorithms combined with human interaction, and to perform precise quantitative measurement and analysis of brain midline deviation; a predictive analysis module configured for multi-model feature fusion, using the XGBoost machine learning model to quantitatively assess and predict the risk of postoperative hydrocephalus in children with posterior fossa tumors, and to output the predicted hydrocephalus results in a quantitative manner; and a user interaction module configured to provide an interface for visualizing the predicted results.

[0045] According to an embodiment of the present invention, the data management module is configured to: receive preoperative MRI image data of a child's brain and perform preprocessing on the MRI image data, the preprocessing including voxel resampling, intensity standardization, and noise filtering of the MRI image data, unifying the MRI image data to a uniform spatial resolution, intensity range, and standardized data format, providing a standardized data foundation for subsequent automatic segmentation of key regions and multi-dimensional feature extraction; the automatic segmentation module is configured to perform: automatic segmentation of the posterior fossa tumor region; joint segmentation of the ventricular effusion region and the exudative region, including automatic identification, quantification, and segmentation of the ventricular effusion region, and automatic detection and segmentation of the exudative region; morphological postprocessing of the segmentation results to optimize boundary accuracy; the feature extraction module is configured to perform: radiomics feature extraction, hydrocephalus quantification feature extraction, and brain midline shift feature extraction; the predictive analysis module is configured to: input the features extracted by the feature extraction module into the XGBoost machine learning model, and output the probability of predicting the hydrocephalus in a quantified manner, wherein the predictive analysis module is further configured to extract the feature importance of the XGBoost model and identify the key features that have the greatest impact on the prediction.

[0046] According to one embodiment of the present invention, the automatic segmentation module is configured to use an improved U-Net deep learning network architecture as a deep learning model to perform automatic segmentation of the posterior fossa tumor region, and to use a 3D UNet network with a deep supervision mechanism as a deep learning model to perform joint segmentation of the ventricular effusion region and the exudation region.

[0047] According to an embodiment of the present invention, the radiomics feature extraction includes: extracting morphological features, including volume, surface area, and sphericity; extracting intensity features, including mean, variance, skewness, and kurtosis; extracting texture features, including matrix features GLCM, GLRLM, and GLSZM; the hydrocephalus quantification feature extraction includes: calculating ventricular volume based on segmentation results; quantifying the volume of the hydrocephalus region and the exudation region; calculating the relative proportions of different parts of the ventricles; the brain midline offset feature extraction includes: performing semi-automatic annotation of the brain midline; calculating the mean, maximum, and coefficient of variation of the brain midline offset; and analyzing the differences in brain midline offset at different levels.

[0048] According to another aspect of the present invention, a storage medium is provided that stores a computer program, which, when executed by a processor, implements the artificial intelligence-based method for predicting postoperative hydrocephalus in children with posterior fossa tumors as described above.

[0049] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the prediction method as described above.

[0050] Compared with the prior art, the present invention can achieve at least the following beneficial technical effects:

[0051] (1) The prediction accuracy is significantly improved: By integrating multi-dimensional information such as tumor radiomics features, hydrocephalus quantitative features, midline shift features and tumor type, the prediction accuracy reaches more than 90%, which is significantly higher than the traditional method of about 75%.

[0052] (2) Objective quantitative assessment: It abandons the traditional assessment method that relies on doctors' subjective experience and provides objective quantitative assessment results based on a large amount of data training, making the prediction process more standardized and repeatable.

[0053] (3) Personalized risk assessment: Based on the specific imaging and clinical characteristics of each patient, targeted risk prediction is provided, rather than simple group statistical results, which is more in line with the concept of precision medicine.

[0054] (4) Clinical decision support: Through clear risk level classification and intuitive visualization, it provides clinicians with powerful decision support tools, which helps to develop more reasonable surgical plans and postoperative management strategies.

[0055] (5) High degree of automation: The system achieves a high degree of automation from image input to prediction result output, which greatly reduces manual operation steps, improves work efficiency, and reduces human error.

[0056] (6) Good scalability: The system adopts a modular design, which can flexibly integrate new features and models, and facilitates functional expansion and performance optimization according to clinical needs.

[0057] (7) Convenient clinical application: The system has a user-friendly interface and is easy to operate. It is easy to integrate into the existing hospital information system to achieve seamless clinical application.

[0058] (8) Evidence-based medicine support: The prediction results are based on a large amount of data analysis and machine learning algorithms, providing doctors with more powerful evidence-based medicine support and enhancing the persuasiveness of doctor-patient communication.

[0059] Further embodiments of the present invention can achieve other advantageous technical effects not listed hereafter, which may be partially described below and can be expected and understood by those skilled in the art after reading the present invention.

[0060] The summary portion of this invention is intended to introduce, in a simplified form, the concepts and options that will be further described in the "Detailed Description" section to aid the reader in understanding the invention. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter. All of the foregoing features are to be understood as merely exemplary, and further features and purposes regarding structures and methods can be gleaned from this disclosure. A more complete display of the features, details, utility, and advantages of the invention is provided in the following written description of various embodiments of the invention, illustrated in the accompanying drawings, and defined in the appended claims. Therefore, the many limiting interpretations of the summary cannot be understood without further reading of the entire specification, claims, and drawings. Attached Figure Description

[0061] The above-described features and advantages of these embodiments, as well as other features and advantages, and the ways in which they are implemented, will become more apparent and the embodiments of the invention will be better understood by referring to the following description in conjunction with the accompanying drawings, in which:

[0062] Figure 1 The illustration shows a schematic diagram of the steps of a method for predicting postoperative hydrocephalus in children with posterior fossa tumors based on artificial intelligence according to an embodiment of the present invention. The method includes preprocessing of images and data, annotation, feature extraction, optimization and XGBoost predictive analysis and verification, as well as visualization and output of the final prediction results, etc. Detailed Implementation

[0063] In the following description of the accompanying drawings and detailed embodiments, details of one or more embodiments of the invention will be set forth. Other features, objects, and advantages of the invention will become apparent from these descriptions, drawings, and claims.

[0064] It should be understood that the illustrated and described embodiments are not limited in application to the details of the construction and arrangement of the components set forth in the following description or illustrated in the drawings. The illustrated embodiments may be other embodiments and can be implemented or performed in various ways. The examples are provided by way of explanation rather than limitation of the disclosed embodiments. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made to the embodiments of the invention without departing from the scope or spirit of the invention. For example, features illustrated or described as part of one embodiment may be used with another embodiment to still produce another embodiment.

[0065] Similarly, it should be understood that the phrases and terms used in this document are for descriptive purposes and should not be considered restrictive. For example, the use of “including,” “contains,” or “has,” and their variations, is intended to include, in an open-ended sense, the items listed thereafter, their equivalents, and any additional items.

[0066] The present invention will be explained and described in more detail below with reference to embodiments of the artificial intelligence-based method and system for predicting postoperative hydrocephalus in children with posterior fossa tumors, and in conjunction with the accompanying drawings.

[0067] Hardware environment

[0068] The following is a non-limiting example of the hardware environment required for the prediction method and system for pediatric posterior fossa tumor surgery based on artificial intelligence, as well as for the operation of the present invention.

[0069] The prediction method and system of this invention can run on ordinary PCs / laptops / servers, etc., to implement the prediction method of this invention. An example configuration of the PC is as follows: CPU: Intel Core i5 13600KF or equivalent processor; Memory: 16GB or more; Storage: 500GB SSD; Operating System: Ubuntu 20.04LTS or Windows 10 / 11 Professional.

[0070] Software environment

[0071] An example of the software environment for the prediction method and prediction system of the present invention is as follows.

[0072] The prediction system of this invention can be developed based on Python 3.10.10, and its core dependencies include: PyTorch 2.0.0+cu118: a deep learning framework used for training and inference of neural network models; PyQt5: for building graphical user interfaces; SimpleITK and Nibabel: for medical image processing; PyRadiomics: for radiomics feature extraction; and scikit-learn and XGBoost: for machine learning model building and prediction. (See attached...) Figure 1 As shown, one embodiment of the artificial intelligence-based method for predicting postoperative hydrocephalus in children with posterior fossa tumors according to the present invention includes the following steps.

[0073] Step 1: Preoperative MRI image acquisition and preprocessing

[0074] Preoperative MRI image acquisition and preprocessing lay the foundation for quantitative assessment. The prediction system first receives brain MRI image data in DICOM format. To ensure the accuracy of subsequent analysis, the images undergo a series of preprocessing steps, including voxel resampling, intensity normalization, and noise filtering, to eliminate differences and interference caused by different scanning parameters, unifying the image data to a standardized state and providing a high-quality data foundation for subsequent key region segmentation and feature extraction.

[0075] This step includes the following detailed implementation process:

[0076] 1.1 Data Input and Validation: The system supports batch import of DICOM format MRI data, automatically checks the integrity and validity of DICOM sequences, and extracts and saves key sequence information (such as patient ID, scan parameters, etc.).

[0077] 1.2 Image Preprocessing

[0078] Image preprocessing generally includes:

[0079] Voxel resampling: MRI images with different scanning parameters are uniformly resampled to a standard voxel size of 1×1×1mm to ensure data consistency.

[0080] Intensity standardization: The Z-score method is used to standardize gray values, mapping pixel values ​​to a distribution with a mean of 0 and a standard deviation of 1.

[0081] Noise reduction: A nonlocal mean filtering algorithm is applied to effectively remove noise while preserving edge information.

[0082] Off-field correction: The N4 algorithm is used to correct the off-field inhomogeneity of MRI images, thereby improving segmentation accuracy.

[0083] Skull removal: Using a deep learning model to remove the skull region of the brain, leaving only the brain region.

[0084] 1.3 Data Organization and Storage

[0085] Data organization and storage generally include:

[0086] Save the processed data in a standard format (NIfTI format) for easy subsequent processing.

[0087] Build data indexes to optimize data access speed.

[0088] Implement an automatic data backup mechanism to prevent accidental data loss.

[0089] Step 2: Automatic segmentation of key regions

[0090] The automatic segmentation of key regions can accurately locate posterior fossa tumor lesions and related areas, mainly including:

[0091] Posterior fossa tumor segmentation: An improved U-Net deep learning network architecture is employed. The encoder integrates a ResNet50 backbone network to enhance feature extraction capabilities, and an attention mechanism is introduced to improve the model's ability to focus on the tumor region, thereby effectively improving segmentation accuracy. The segmentation network performs end-to-end processing on the input MRI images, outputting a high-precision binarized tumor mask. The segmentation results undergo morphological post-processing operations (e.g., but not limited to small region removal and hole filling) to further optimize the tumor region boundaries and ensure the clinical applicability of the segmentation results.

[0092] Joint segmentation of ventricular effusion and exudative regions: A 3D UNet network with deep supervision is used, simultaneously optimized for features of the ventricular system and exudative regions. Through a multi-scale feature fusion architecture, it adapts to the differences in morphological features of different regions and employs a combined loss function strategy (Dice loss function, cross-entropy loss function, and boundary-weighted loss function) to achieve multi-class segmentation of ventricular effusion and exudative regions, accurately identifying these two closely related regions in a single step. The model outputs a three-channel prediction map (background, ventricular effusion region, and exudative region), facilitating subsequent quantitative analysis of these two regions separately.

[0093] This segmentation step can automatically segment multiple key regions using a deep learning model, specifically including:

[0094] 2.1 Segmentation of posterior fossa tumor regions

[0095] An improved U-Net deep learning network architecture is adopted, in which the encoder uses a ResNet50 backbone network; an attention mechanism is added to enhance the model's focus on the tumor region; accurate segmentation of the tumor region is achieved, and a binary mask is generated; the segmentation results are post-processed by morphological operations to remove small holes and false positive regions.

[0096] 2.2 Combined segmentation of ventricular hydrocephalus and exudative areas

[0097] A 3D UNet network employing a deep supervision mechanism is optimized for both the ventricular system and exudative regions. It achieves multi-class segmentation, simultaneously identifying ventricular effusion and exudative regions. A multi-scale feature fusion architecture is designed to adapt to the morphological differences in different regions. The following combined loss function strategy is used:

[0098] The Dice loss function reduces the impact of class imbalance, as follows:

[0099]

[0100] in:

[0101] N: Total number of pixels or total number of samples

[0102] p_i: The predicted probability value of the i-th pixel (usually between 0 and 1).

[0103] g_i: The true label value of the i-th pixel (usually 0 or 1)

[0104] ε (epsilon): A smoothing term to prevent the denominator from being zero; it is usually taken as a very small positive number (e.g., 1e-7).

[0105] ∑: Summation symbol

[0106] Numerator: 2 × (intersection of predicted and actual values) + smoothing term

[0107] Denominator: Sum of predicted values ​​+ Sum of actual values ​​+ Smoothing term.

[0108] Dice Loss measures the degree of overlap between the predicted result and the true label. The smaller the value, the higher the overlap and the better the segmentation effect. When the prediction is completely correct, Dice Loss = 0; when the prediction is completely wrong, Dice Loss is close to 1.

[0109] Cross-entropy loss (CE Loss) enhances pixel-level classification capabilities, as follows:

[0110]

[0111] in:

[0112] N: Total number of samples

[0113] i: Sample index, from 1 to N

[0114] g_i: The true label value of the i-th sample (0 or 1, in the case of binary classification).

[0115] p_i: The probability value (between 0 and 1) that the i-th sample is predicted as positive.

[0116] log: natural logarithm function

[0117] The minus sign (-) makes the loss function positive because the log probability is usually negative.

[0118] 1 / N: Calculate the average over all samples.

[0119] When g_i = 1 (the true label is positive): the loss is -log(p_i), and the closer the predicted probability p_i is to 1, the smaller the loss.

[0120] When g_i = 0 (the true label is negative): the loss is -log(1-p_i). The closer the predicted probability p_i is to 0, the smaller the loss.

[0121] Cross-entropy loss can effectively punish incorrect predictions. The loss is 0 when the predicted probability is exactly the same as the true label, and the loss increases sharply when the prediction is wrong.

[0122] Boundary-weighted loss (CE Loss) improves boundary recognition accuracy, as follows:

[0123]

[0124] in:

[0125] N: Total number of samples (usually the total number of pixels)

[0126] i: Sample index, from 1 to N

[0127] w_i: The weight coefficient of the i-th sample, used to adjust the importance of different regions.

[0128] g_i: The true label value (0 or 1) of the i-th sample.

[0129] p_i: The probability value (between 0 and 1) that the i-th sample is predicted as positive.

[0130] log: natural logarithm function

[0131] The minus sign (-) makes the loss function positive.

[0132] 1 / N: Calculate the average over all samples.

[0133] Furthermore, the model outputs a three-channel prediction map, which includes: background, hydrocephalus area, and exudation area.

[0134] Post-segmentation processing: Volume calculation and morphological analysis of the ventricular effusion region to quantify the degree of hydrocephalus; volume measurement and spatial distribution feature extraction of the exudation region; analysis of the relative positional relationship and interaction features of the two regions.

[0135] It provides 3D visualization of the segmentation results, intuitively showing the spatial distribution of each region.

[0136] Simultaneously calculate morphological parameters for both regions, including quantitative parameters such as volume, surface area, and maximum diameter.

[0137] 2.3 Validation and Optimization of Segmentation Results

[0138] It provides a visual representation of the segmentation results for easy manual verification; it enables interactive segmentation correction, allowing professionals to fine-tune the results; and it automatically saves the segmentation results for subsequent feature extraction.

[0139] Step 3: Multi-dimensional feature extraction

[0140] Multi-dimensional feature extraction steps can deeply mine quantitative indicators.

[0141] Radiomics Feature Extraction: Morphological features (such as volume, surface area, sphericity, etc., 14 items), intensity features (such as mean, variance, skewness, kurtosis, etc., 17 items), and texture features (based on matrix features such as GLCM, GLRLM, and GLSZM, a total of 73 items) were extracted from tumor regions, hydrocephalus regions, and exudative regions. These features describe the imaging characteristics of each region from different perspectives. For example, volume and surface area can directly reflect the size and morphological changes of the region, while texture features can reflect the microstructure and heterogeneity of the tissue, providing rich quantitative information for assessing the risk of hydrocephalus.

[0142] Quantitative feature extraction of hydrocephalus: Based on the segmentation results, the volume of the ventricles is directly calculated, the volume of hydrocephalus and exudation areas is quantified, and the relative proportions of different parts of the ventricles are calculated. The above quantitative indicators intuitively reflect the degree of accumulation and structural expansion characteristics of cerebrospinal fluid in the ventricular system, providing a key medical parameter basis for the automated and quantitative assessment of hydrocephalus risk.

[0143] Midline shift feature extraction: This function achieves semi-automatic annotation of the brain midline, calculates the average, maximum, and coefficient of variation of midline shift, and analyzes the differences in midline shift at different levels. Midline shift is an important objective parameter reflecting the spatial displacement of brain tissue and dynamic changes in intracranial pressure caused by hydrocephalus and other intracranial space-occupying lesions. Through automatic measurement and quantitative analysis of midline shift, the impact of hydrocephalus on brain tissue structure and intracranial pressure balance can be assessed more accurately.

[0144] More specifically, step 3 (multi-dimensional feature extraction) performs comprehensive feature extraction on the segmentation results, including:

[0145] 3.1 Radiomics Feature Extraction

[0146] Implement a standardized feature extraction process using the PyRadiomics library.

[0147] 3.1.1 Morphological features, describing the geometry and spatial properties of the region of interest (ROI) (e.g., 14 items): including volume, surface area, surface area to volume ratio, sphericity, maximum diameter, etc.

[0148] Volumetric characteristics:

[0149]

[0150] Where v i Let N be the volume of the i-th voxel, and N be the total number of voxels within the ROI.

[0151] Surface area characteristics:

[0152]

[0153] Among them, a j Let M be the area of ​​the j-th surface grid, and M be the total number of surface grids.

[0154] Sphericity characteristics:

[0155]

[0156] This metric quantifies the similarity between the ROI and a sphere of the same volume; the closer the value is to 1, the closer it is to a sphere.

[0157] Maximum diameter characteristics:

[0158]

[0159] Where, p i and p j Let be any two points on the boundary of the ROI.

[0160] Compressibility characteristics:

[0161]

[0162] This metric measures the compactness of the ROI; a higher value indicates a more compact shape, reflecting the spatial aggregation characteristics of the lesions.

[0163] 3.1.2 Intensity characteristics, describing the statistical distribution characteristics of voxel intensity values ​​within the ROI (e.g., 17 items): including mean, variance, skewness, kurtosis, energy, entropy, etc.

[0164] Intensity features, first-order statistics are as follows:

[0165] Mean:

[0166]

[0167] in, X i represents the intensity value of the i-th voxel, and N represents the total number of voxels within the ROI. The average level of voxel intensity within the ROI is a fundamental indicator for describing the overall brightness of the lesion.

[0168] variance:

[0169]

[0170] Where μ is the mean, and the denominator is estimated using N-1 unbiased values. The dispersion of the intensity values ​​reflects the uniformity of intensity within the lesion.

[0171] Skewness:

[0172]

[0173] Where σ is the standard deviation, and the cube retains the sign information. It describes the asymmetry of the intensity distribution; positive skewness indicates that there are more high-intensity pixels, and negative skewness indicates that there are more low-intensity pixels.

[0174] Kuroshi:

[0175]

[0176] The fourth power eliminates the influence of sign, resulting in the standardized fourth-order central moment. It measures the sharpness of the intensity distribution; kurtosis indicates a more concentrated distribution, while hypokurtosis indicates a flatter distribution.

[0177] Energy characteristics:

[0178]

[0179] Among them, X i This represents the voxel intensity value, and squaring ensures non-negativity. It reflects the overall magnitude of the intensity values ​​within the ROI; high energy values ​​typically correspond to high-intensity regions.

[0180] Entropy characteristics:

[0181]

[0182] Where p(i) is the probability of gray value i appearing, and N g The number of gray levels after discretization.

[0183] 3.1.3 Texture features describe the texture pattern of an image by analyzing the spatial relationships between pixels (e.g., 73 items), including features based on the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size region matrix (GLSZM), adjacent gray-level dependency matrix (GLDM), and adjacent gray-level tone difference matrix (NGTDM), etc.

[0184] Gray-level Co-occurrence Matrix (GLCM) Features:

[0185] Define a GLCM with distance d and angle θ as $P(i,j|d,θ)$, whose main characteristics include:

[0186] Contrast:

[0187] Where P(i,j) is the probability value of position (i,j) in GLCM, and (ij) 2 This is the square of the grayscale difference. It measures the magnitude of the grayscale difference between adjacent pixels; high contrast indicates drastic texture changes.

[0188] Correlation:

[0189] Where μ x μ y Let σ be the mean of the row and the mean of the column, respectively. x σ y This represents the corresponding standard deviation. It describes the degree of linear correlation between adjacent pixels, reflecting the regularity and directionality of the texture.

[0190] energy:

[0191] In this context, the square of P(i,j) emphasizes elements with high probabilities. It reflects the uniformity of the texture, with high energy values ​​indicating relatively simple and regular texture patterns.

[0192] entropy:

[0193] The negative sign ensures a positive entropy value, and log is typically the natural logarithm. It quantifies the complexity and randomness of the texture; a high entropy value indicates a more complex and varied texture pattern.

[0194] Gray run length matrix (GLRLM) characteristics:

[0195] Define a run length matrix P(i, j|θ), where i is the gray value and j is the run length:

[0196] Short trips emphasize:

[0197] Where P(i,j) are elements of the run-length matrix, i is the gray value, j is the run length, and j is the gray value. 2 The denominator helps to suppress long runs. It emphasizes the contribution of short continuous runs; high values ​​indicate fine texture and frequent grayscale changes.

[0198] Longer travel itineraries emphasize:

[0199] Where, j 2 In molecules, this emphasizes the importance of long, continuous strokes. High values ​​indicate a rough texture and the presence of large, uniform regions.

[0200] Uneven travel length:

[0201] in, The number of all runs of length j is represented by a square operation, which amplifies the non-uniformity. This measures the uniformity of the distribution of runs of different lengths, reflecting the regularity of the texture pattern.

[0202] Gray-scale region matrix (GLSZM) features:

[0203] Small area emphasis:

[0204] Where P(i,j) are elements of the region size matrix, i is the gray value, j is the region size, and j is the gray value. 2 The denominator suppresses large regions. It highlights the contribution of small connected regions, and high values ​​indicate that the texture has a fine local structure.

[0205] Large-area emphasis:

[0206] Where, j 2 In molecular analysis, the importance of large regions is emphasized. High values ​​highlight the importance of highly connected regions, with high values ​​indicating the presence of large homogeneous regions.

[0207] Adjacent gray-level dependency matrix (GLDM) features and adjacent gray-level tonal difference matrix (NGTDM) features further enrich the dimensions of texture description.

[0208] The feature extraction process is performed on the tumor region, hydrocephalus region, and exudation region segmented by the deep learning model, respectively, to generate a comprehensive feature matrix.

[0209] 3.2 Quantitative Feature Extraction of Hydrocephalus

[0210] The hydrocephalus quantitative feature extraction includes, for example, the following implementations: calculating the ventricular volume based on the segmentation results; quantifying the volume of hydrocephalus and exudative regions; and calculating the relative proportions of different parts of the ventricles.

[0211] A non-limiting specific example of the relevant content of the hydrocephalus quantitative feature extraction can be found in Table 1 below.

[0212] Table 1: Quantitative Feature Extraction of Hydrocephalus

[0213]

[0214]

[0215] 3.3 Extraction of brain midline shift features

[0216] Brain midline shift feature extraction generally includes: implementing semi-automatic brain midline annotation; providing an intuitive interactive interface with support for shortcut key operations; calculating the average, maximum, and coefficient of variation of midline shift; and analyzing the differences in midline shift at different levels.

[0217] 3.4 Preprocessing of Feature Data (Original Feature Values)

[0218] Feature normalization: The Z-score method is used to normalize the original feature values ​​to a uniform scale, eliminating the influence of different feature dimensions and value ranges, so that each feature can participate in subsequent modeling and analysis on the same scale, as follows:

[0219]

[0220] in,

[0221] x′ represents the normalized eigenvalues;

[0222] x represents the original eigenvalue;

[0223] μ represents the mean of this feature across all samples;

[0224] σ represents the standard deviation of this feature across all samples.

[0225] Feature selection: Recursive feature elimination (RFE) and L1 regularization are used to select the subset of features with the highest predictive value.

[0226] Multicollinearity handling: Identifies and processes highly correlated features, reducing feature redundancy.

[0227] Missing value handling: Multiple imputation is used to fill in any missing feature values ​​that may exist. Step 4: XGBoost Predictive analysis and validation

[0228] The multi-model fusion predictive analysis step can comprehensively and quantitatively assess risk. The extracted multi-dimensional features (including radiomics features, hydrocephalus quantitative features, and midline shift features) are integrated and input into a pre-trained XGBoost machine learning model. The model uses complex algorithms to comprehensively analyze and weight these features, considering the interactions and correlations between them, and finally outputs a predicted probability value (0-100%) of hydrocephalus risk, quantitatively clarifying the likelihood of postoperative hydrocephalus. Simultaneously, the feature importance of the XGBoost model can be extracted and visualized, identifying the key features with the greatest impact on the prediction results, further explaining the basis of the quantitative assessment results.

[0229] This step can achieve high-precision predictions using machine learning models such as XGBoost:

[0230] 4.1 Feature Integration and Preprocessing

[0231] All features extracted in steps 2 and 3 are integrated, including tumor features, hydrocephalus and exudative region features, radiomics features, and midline shift features.

[0232] Feature normalization: Normalizes feature values ​​to a uniform scale.

[0233] Feature selection: Recursive feature elimination (RFE) and L1 regularization are used to select the subset of features with the highest predictive value.

[0234] Multicollinearity handling: Identifies and processes highly correlated features, reducing feature redundancy.

[0235] Missing value handling: Multiple imputation is used to fill in any missing feature values ​​that may exist.

[0236] 4.2 Analysis and Output of Prediction Results

[0237] The trained XGBoost model is used to predict test samples. An example of the model parameters is shown below:

[0238] -'n_estimators':384,

[0239] -'max_depth':4,

[0240] -'learning_rate':0.1344084822001895,

[0241] -'subsample':0.5456014064805899,

[0242] -'colsample_bytree':0.9378492661146669,

[0243] -'gamma':0.21656170238806297

[0244] The generated predicted probability value (0-100%) represents the risk of postoperative hydrocephalus.

[0245] Extract and visualize the feature importance of the XGBoost model to identify the key features that have the greatest impact on the prediction results.

[0246] SHAP value analysis is generated to interpret the characteristic contribution of individual prediction results.

[0247] Output standardized forecast reports, including forecast probabilities, risk levels, and key influencing factors.

[0248] 4.3 Model Evaluation and Validation

[0249] The model's performance is comprehensively evaluated and validated on independent test sets, as shown below:

[0250] -Accuracy: The overall accuracy of predictions.

[0251] -Sensitivity: The ability to correctly identify when hydrocephalus will actually occur.

[0252] - Specificity: The ability to correctly identify cases where hydrocephalus will not actually occur.

[0253] -AUC-ROC value: A comprehensive indicator of the model's ability to distinguish between different models.

[0254] - Precision-Recall Curve: Evaluate model performance at different thresholds.

[0255] Step 5: Visualization and Output of Prediction Results

[0256] The visualization and output steps for prediction results can intuitively present the quantitative evaluation results.

[0257] Risk prediction report generation: Generates a standardized report containing basic patient information, imaging features, and prediction results, clearly displaying the probability and level of hydrocephalus risk, allowing doctors to understand the patient's hydrocephalus risk status at a glance.

[0258] Feature visualization: Generates XGBoost feature importance bar charts to intuitively display the main features affecting prediction results; provides SHAP value waterfall charts to explain the feature contribution of prediction results for individual patients; and also provides a 3D visualization interface to display segmentation results and key regions, as well as a graphical display of midline offset, enabling doctors to intuitively understand the basis and results of quantitative assessment from multiple dimensions.

[0259] This step allows for a visual display of the prediction results:

[0260] 5.1 Risk prediction report generation, including: generating a standardized report containing basic patient information, imaging features and prediction results; clearly displaying the probability and level of hydrocephalus risk; highlighting key influencing factors identified by the XGBoost model; and including model reliability information, such as the confidence interval of the predicted probability.

[0261] 5.2 Feature visualization, including: generating XGBoost feature importance bar charts to intuitively display the main features affecting the prediction results; providing SHAP value waterfall charts to explain the feature contribution of prediction results for individual patients; providing a 3D visualization interface to display segmentation results and key regions; and realizing a graphical display of midline offset.

[0262] 5.3 Results export function, including: support for exporting prediction results to CSV format for easy subsequent analysis; provide NIFTI format export of segmentation results; support for generating comprehensive reports in PDF format, including all analysis results and visualization charts.

[0263] Implementation and Deployment of Prediction Systems

[0264] The prediction system of this invention adopts a modular software architecture, with each functional module encapsulated as an independent class and interacting through standardized interfaces. The system interface can be implemented using, for example, the PyQt5 framework, providing an intuitive and user-friendly interactive experience. The main interfaces may include: a login interface, a main function menu, an automatic segmentation interface, a feature extraction interface, a midline annotation interface, and a prediction analysis interface.

[0265] Regarding system deployment, hospitals can choose between local deployment or server deployment, depending on their actual needs. Local deployment is suitable for single-machine use, allowing the software to run directly on the doctor's workstation; server deployment, on the other hand, supports concurrent access by multiple users and centralized data management, making it more suitable for the needs of large medical institutions.

[0266] The prediction system of this invention also provides a comprehensive error handling mechanism, with full pre-setup and processing for various possible abnormal situations. These include data input errors, segmentation model loading failures, and feature extraction anomalies, all of which have corresponding error prompts and recovery mechanisms to ensure stable system operation.

[0267] Operational procedures of the prediction system

[0268] A non-limiting example of the exemplary operation flow of the prediction system of the present invention is as follows.

[0269] 1) Log in to the system

[0270] - Launch the software and enter the login interface.

[0271] - Enter username and password (default is empty)

[0272] -After successful verification, you will enter the main interface.

[0273] 2) Data import and splitting

[0274] - Select "Automatic Segmentation" on the main interface.

[0275] - Import patient's MRIDICOM data

[0276] - Select the segmentation model type

[0277] - Perform segmentation to obtain masks for posterior fossa tumors, hydrocephalus areas, and exudative areas.

[0278] 3) Feature extraction

[0279] - Select the "Radiomics Analysis" function

[0280] - The system automatically loads the segmentation results from the previous step.

[0281] - Perform feature extraction process

[0282] - Generate a CSV file containing all radiomics features.

[0283] 4) Centerline labeling

[0284] - Select the "Centerline Annotation Tool" function

[0285] - Mark the midline of the brain at each level on the interactive interface.

[0286] - The system automatically calculates the centerline offset parameters.

[0287] - Save the annotation results.

[0288] 5) Predictive Analysis

[0289] - Select the "Predictive Analytics Module" function

[0290] -The system automatically integrates the results data from the first three steps.

[0291] - Load optimized XGBoost model parameters

[0292] -Execute predictive analytics

[0293] - View prediction results, risk assessment, and feature importance analysis.

[0294] 6) Exporting Results

[0295] -View forecast report

[0296] - Export analysis results as needed

[0297] -Save or print the forecast report.

[0298] Through the detailed embodiments described above, the prediction method and system of the present invention achieve artificial intelligence-based prediction of postoperative hydrocephalus in children with posterior fossa tumors, providing objective and accurate risk assessment results and strong support for clinical decision-making. Meanwhile, the system's modular design and user-friendly interface ensure its usability and adaptability in real-world medical environments.

[0299] Beneficial technical effects

[0300] In summary, through the series of steps in the prediction method and the specific configuration of the prediction system of this invention, the prediction method and system of this invention achieve precise quantitative assessment of the risk of postoperative hydrocephalus in children with posterior fossa tumors. This provides clinicians with objective, accurate, and intuitive decision support information, helping to develop more personalized and targeted treatment plans and improve the prognosis of children. The beneficial technical effects achievable by this invention include, but are not limited to, those listed below.

[0301] High prediction accuracy: By integrating multi-dimensional information, the prediction accuracy reaches over 90%, significantly higher than the approximately 70% of traditional methods.

[0302] Objective quantitative assessment: Abandoning the traditional assessment method that relies on doctors' subjective experience, it provides objective quantitative assessment results based on training with a large amount of data.

[0303] Personalized risk assessment: Based on the patient's specific imaging and clinical characteristics, it provides targeted risk predictions, which is in line with the concept of precision medicine.

[0304] Clinical decision support: Provides clinicians with powerful decision support tools to help develop more reasonable surgical plans and postoperative management strategies.

[0305] High degree of automation: It achieves a high degree of automation from image input to prediction result output, improving work efficiency and reducing human error.

[0306] Excellent scalability: The modular design allows for flexible integration of new features and models, facilitating feature expansion and performance optimization.

[0307] Convenient for clinical applications: User-friendly interface, easy to operate, and easy to integrate into existing hospital information systems.

[0308] Evidence-based medicine support: Predictive outcomes are based on extensive data analysis and machine learning algorithms, enhancing the persuasiveness of doctor-patient communication.

[0309] Application scenarios

[0310] This invention is mainly applied to the prediction of hydrocephalus risk after surgery for pediatric posterior fossa tumors. It can be widely used in various medical institutions, physical examination facilities, and medical research institutes in the fields of neurosurgery and pediatrics, providing decision support for clinicians, improving diagnosis and treatment outcomes, and improving the prognosis, growth, and quality of life of children.

[0311] The foregoing description of several embodiments of the invention has been provided for illustrative purposes. This foregoing description is not intended to be exhaustive, nor is it intended to limit the invention to the precise features and / or forms disclosed. It will be apparent to those skilled in the art that many modifications and variations can be made in light of the foregoing teachings, all of which fall within the scope of this invention. The scope of the invention is defined only by the appended claims.

Claims

1. A method for predicting postoperative hydrocephalus in children with posterior fossa tumors based on artificial intelligence, characterized in that, The prediction method includes the following steps: Step S1: Receive MRI image data of the child's brain before surgery and perform preprocessing on the MRI image data. The preprocessing includes voxel resampling, intensity standardization and noise filtering of the MRI image data, and unifying the MRI image data to a uniform spatial resolution, intensity range and standardized data format, so as to provide a standardized data foundation for subsequent automatic segmentation of key areas and multi-dimensional feature extraction. Step S2: Automatic segmentation of key regions, locating the posterior fossa tumor lesion and related regions, and automatically segmenting multiple key regions in the posterior fossa tumor lesion and related regions using a deep learning model; Step S3: Multi-dimensional feature extraction, used to extract multi-dimensional, quantifiable image features from the segmentation results of Step S2. These features reflect the imaging characteristics of the posterior fossa tumor lesion and associated regions, and are quantitatively analyzed. The image features include the volume, shape, signal intensity distribution, and spatial location of the posterior fossa tumor lesion and associated regions. Step S4: XGBoost Predictive Analysis and Validation, which uses the XGBoost machine learning model to quantitatively assess and predict the risk of hydrocephalus after posterior fossa tumor surgery in children, and outputs the prediction results of hydrocephalus in a quantitative manner.

2. The prediction method according to claim 1, characterized in that, The method further includes step S5: visualization and output of the prediction results, used to intuitively present the results of quantitative assessment of hydrocephalus.

3. The prediction method according to claim 1 or 2, characterized in that, Step S2 further includes: automatic segmentation of the posterior fossa tumor region; joint segmentation of the ventricular effusion region and the exudative region, which includes automatic identification, quantification and segmentation of the ventricular effusion region, and automatic detection and segmentation of the exudative region; and morphological post-processing of the segmentation results to optimize boundary accuracy. Step S3 further includes: radiomics feature extraction, hydrocephalus quantitative feature extraction, and brain midline shift feature extraction; and Step S4 further includes: inputting the features extracted in step S3 into the XGBoost machine learning model to output the probability of predicting hydrocephalus in a quantitative manner. Step S4 also includes extracting the feature importance of the XGBoost model and identifying the key features that have the greatest impact on the prediction.

4. The prediction method according to claim 3, characterized in that, In step S2, the automatic segmentation of the posterior fossa tumor region uses an improved U-Net deep learning network architecture as a deep learning model, wherein the encoder uses a ResNet50 backbone network to extract features and incorporates an attention mechanism; the joint segmentation of the ventricular effusion region and the exudation region uses a 3D UNet network with a deep supervision mechanism as a deep learning model.

5. The prediction method according to claim 3, characterized in that, In step S3, the image omics feature extraction includes: extracting morphological features, including volume, surface area, and sphericity; extracting intensity features, including mean, variance, skewness, and kurtosis; and extracting texture features, including matrix features GLCM, GLRLM, and GLSZM. The hydrocephalus quantitative feature extraction includes: calculating the ventricular volume based on the segmentation results; quantifying the volume of the hydrocephalus region and the exudation region; and calculating the relative proportions of different parts of the ventricle. The brain midline offset feature extraction includes: performing semi-automatic annotation of the brain midline; calculating the average, maximum and coefficient of variation of the brain midline offset; and analyzing the differences in brain midline offset at different levels.

6. An artificial intelligence-based predictive system for postoperative hydrocephalus in children with posterior fossa tumors, characterized in that, include: The data management module is configured for importing, preprocessing and managing preoperative MRI image data. It unifies the preoperative MRI image data into a standardized data standard through voxel resampling, intensity standardization and format conversion, so as to provide a consistent data foundation for subsequent processing. An automatic segmentation module is configured to locate posterior fossa tumor lesions and associated regions, and to automatically segment multiple key regions in the posterior fossa tumor lesions and associated regions using a deep learning model. A feature extraction module is configured to extract multi-dimensional, quantifiable image features from the results of the automatic segmentation. These features reflect the imaging characteristics of the posterior fossa tumor lesion and associated regions. The module also performs quantitative analysis on these image features, which include the volume, shape, signal intensity distribution, and spatial location of the posterior fossa tumor lesion and associated regions. The brain midline annotation module is configured to combine automatic detection algorithms with human interaction to achieve semi-automatic annotation of the brain midline, and to automatically calculate and output quantitative characteristic indicators of brain midline deviation based on the results of the semi-automatic annotation, so as to assist in subsequent risk assessment and predictive analysis. The predictive analytics module is configured for multi-model feature fusion, quantitatively assessing and predicting the risk of postoperative hydrocephalus in children with posterior fossa tumors using an XGBoost machine learning model, and outputting the predicted hydrocephalus results in a quantitative manner; and The user interaction module is configured to provide an operation interface to provide a visual display of the prediction results.

7. The prediction system according to claim 6, characterized in that, The data management module is configured to receive MRI image data of children before brain surgery and perform preprocessing on the MRI image data. The preprocessing includes voxel resampling, intensity standardization and noise filtering of the MRI image data, and unifying the MRI image data to a uniform spatial resolution, intensity range and standardized data format, so as to provide a standardized data foundation for subsequent automatic segmentation of key areas and multi-dimensional feature extraction. The automatic segmentation module is configured to perform: automatic segmentation of the posterior fossa tumor region; joint segmentation of the ventricular effusion region and the exudative region, which includes automatic identification, quantification and segmentation of the ventricular effusion region, and automatic detection and segmentation of the exudative region; and morphological post-processing of the segmentation results to optimize boundary accuracy. The feature extraction module is configured to perform: radiomics feature extraction, hydrocephalus quantification feature extraction, and brain midline shift feature extraction; and The predictive analysis module is configured to: input the features extracted by the feature extraction module into the XGBoost machine learning model, and output the probability of predicting hydrocephalus in a quantitative manner. The predictive analysis module is further configured to extract the feature importance of the XGBoost model and identify the key features that have the greatest impact on the prediction.

8. The prediction system according to claim 6 or 7, characterized in that, The automatic segmentation module is configured to use an improved U-Net deep learning network architecture as a deep learning model to perform automatic segmentation of the posterior fossa tumor region, and to use a 3D UNet network with a deep supervision mechanism as a deep learning model to perform joint segmentation of the ventricular effusion region and the exudation region.

9. The prediction system according to any one of claims 7-8, characterized in that, The image omics feature extraction includes: extracting morphological features, including volume, surface area, and sphericity; extracting intensity features, including mean, variance, skewness, and kurtosis; and extracting texture features, including matrix features GLCM, GLRLM, and GLSZM. The hydrocephalus quantitative feature extraction includes: calculating the ventricular volume based on the segmentation results; quantifying the volume of the hydrocephalus region and the exudation region; and calculating the relative proportions of different parts of the ventricle. The brain midline offset feature extraction includes: performing semi-automatic annotation of the brain midline; calculating the average, maximum and coefficient of variation of the brain midline offset; and analyzing the differences in brain midline offset at different levels.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence-based method for predicting postoperative hydrocephalus in children with posterior fossa tumors as described in any one of claims 1 to 5.

11. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the prediction method as described in any one of claims 1 to 5.

Images