Construction method and system of an AI-based postoperative cerebellar mutism automatic risk assessment model

Abstract

The application relates to the technical field of medical care information science, and discloses a construction method and system of an AI-based postoperative cerebellar mutism automatic risk assessment model, according to MRI image data, tumor region label data obtained by combining a semi-automatic labeling tool and artificially corrected tumor region label data, a segmentation model with high precision and noise robustness is obtained through iterative optimization; high-quality tumor region label data is obtained by using the segmentation model, and then the model is trained to learn the corresponding clinical features and tumor region features of a postoperative cerebellar mutism patient with / without cerebellar mutism by combining MRI image data, clinical data and label data of whether the patient has cerebellar mutism, so that a cerebellar mutism risk assessment model with strong generalization ability is obtained, the model can accurately predict the risk of CMS, provides objective and quantitative decision support for clinicians, and is helpful for preoperative risk stratification, personalized surgical plan making and postoperative early intervention.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and in particular to a method and system for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism. Background Technology

[0002] Cerebellar mutism syndrome (CMS) is a common complication after posterior fossa tumor surgery in children, with an incidence of approximately 25%. It manifests as postoperative language impairment, with an incubation period of 1-7 days and a duration of 1-6 months. Language function is difficult to fully recover, severely impacting the child's cognitive, social, and mental health. Currently, the mechanism of CMS is unclear, and effective predictive tools are lacking, hindering the development of individualized treatment plans.

[0003] The following key bottlenecks exist in predictive research on CMS:

[0004] Data limitations: Cases of pediatric posterior fossa tumors are rare, and the sharing of medical images is limited, resulting in small publicly available datasets and poor model generalization ability. Insufficient data diversity also affects model robustness; traditional segmentation models are prone to overfitting when data is scarce, and decreased segmentation accuracy reduces the reliability of feature extraction.

[0005] Labeling issues: Tumor labeling relies on senior physicians, with each case taking over an hour, and there are significant differences in labeling results among different physicians (Dice coefficient difference of 0.15-0.20). Building large-scale datasets is time-consuming and labor-intensive. During the labeling process, issues such as intra- / inter-observer variability, physician fatigue, and unclear tumor boundaries can introduce noise and affect model performance.

[0006] Model limitations: Existing studies are mostly based on single-modal MRI, without integrating multimodal imaging (DTI, fMRI) and clinical parameters, resulting in low predictive efficacy (AUC < 0.75). Traditional machine learning models struggle to capture the nonlinear relationships of high-dimensional features.

[0007] Insufficient clinical translation: Existing research findings are mostly confined to the research stage, lacking practical predictive tools. Doctors still rely on subjective experience to assess risk, and there is an urgent need for objective, efficient, and automated tools to assist in decision-making, in order to optimize treatment plans and doctor-patient communication. Summary of the Invention

[0008] This invention provides a method and system for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism, in order to address the current lack of highly effective risk assessment tools for postoperative cerebellar mutism.

[0009] This invention provides a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism, comprising:

[0010] Acquire MRI imaging data, clinical data, and labeling data on whether or not the target group has cerebellar mutism. The target group is patients with posterior fossa tumors who have undergone surgery for posterior fossa tumors.

[0011] Based on the MRI image data of the target group, a segmentation model was trained using tumor region label data obtained by a semi-automatic annotation tool and manually corrected tumor region label data, and the final tumor region label data was obtained using the segmentation model.

[0012] Based on the MRI imaging data of the target group, the final tumor region labeling data, clinical data, and labeling data on whether or not the patient has cerebellar mutism, the training model learns the characteristics of the tumor region on the MRI images corresponding to the clinical data of patients with posterior fossa tumors and patients without cerebellar mutism, thus obtaining a risk assessment model for cerebellar mutism syndrome.

[0013] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism includes, based on MRI image data of the target population, training a segmentation model using tumor region label data obtained through a semi-automatic annotation tool and manually corrected tumor region label data, and using the segmentation model to obtain the final tumor region label data, comprising:

[0014] A semi-automatic annotation tool was used to delineate tumor regions in the MRI image data of the target population to obtain the first tumor region label data;

[0015] Based on the MRI image data of the target group and the first tumor region label data, multiple segmentation models are trained and the segmentation model that meets the first preset condition is selected from the multiple segmentation models. Based on the MRI image data of the target group, the second tumor region label data is obtained by using the segmentation model that meets the first preset condition.

[0016] Receive manually corrected second tumor region label data;

[0017] Based on the MRI image data of the target population and the manually corrected second tumor region label data, multiple segmentation models were retrained, and the segmentation models that meet the second preset conditions were selected from the multiple segmentation models.

[0018] Based on the MRI image data of the target population, the final tumor region label data is obtained using a segmentation model that meets the second preset conditions.

[0019] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism is provided. Multiple segmentation models include any one or any combination of the following: nnU-Net (v2, ResEncM configuration) deep learning segmentation model, SegResNet deep learning segmentation model, SwinUNetR deep learning segmentation model, DiNTS deep learning segmentation model, and Auto3DSeg deep learning segmentation model. A first preset condition includes a Dice coefficient not lower than a first preset threshold on the validation set; a second preset condition includes a Dice coefficient greater than a second preset threshold on the validation set; and an average surface distance greater than a third preset threshold.

[0020] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism includes obtaining second tumor region label data based on MRI image data of a target population using a segmentation model that meets first preset conditions, comprising:

[0021] When there is more than one segmentation model that meets the first preset condition, the segmentation results of all segmentation models that meet the first preset condition are averaged using a probability graph to obtain the second tumor region label data.

[0022] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism is provided. The method involves training the model to learn the characteristics of the tumor region on MRI images corresponding to the clinical data of patients with posterior fossa tumors suffering from cerebellar mutism and those without, based on MRI image data of the target population, final tumor region labeling data, clinical data, and labeling data indicating whether or not the patient has cerebellar mutism. This results in a cerebellar mutism risk assessment model, comprising:

[0023] Radiomics feature data were obtained based on the MRI imaging data of the target population and the final tumor region labeling data.

[0024] Normalize the radiomics feature data;

[0025] The clinical data is coded.

[0026] Based on normalized radiomics feature data and coded clinical data, multiple machine learning algorithms were used to train models to learn the characteristics of tumor regions on MRI images corresponding to clinical data of patients with posterior fossa tumors and patients without posterior fossa tumors, resulting in multiple candidate cerebellar mutism syndrome risk assessment models.

[0027] The performance of multiple candidate risk assessment models for cerebellar mutism syndrome was evaluated to obtain the final risk assessment model for cerebellar mutism syndrome.

[0028] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism is provided. The radiomics feature data includes any one or any combination of the following: first-order statistical features, two-dimensional morphological features, and three-dimensional morphological features. The first-order statistical features include any one or any combination of the following: energy, entropy, mean, median, minimum, maximum, range, interquartile range, standard deviation, variance, skewness, kurtosis, root mean square, and 10th percentile. The two-dimensional morphological features include any one or any combination of the following. Combinations: Perimeter, MeshSurface, PixelSurface, Maximum Diameter, Perimeter-SurfaceRatio; 3D morphological features include any one of the following or any combination thereof: MeshVolume, SurfaceArea, Sphericity, Maximum3DDiameter, MajorAxisLength, MinorAxisLength, LeastAxisLength, SurfaceVolumeRatio, Compactness1, Elongation.

[0029] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism is provided. The clinical data includes any one or any combination of the following: age (continuous value), gender (male / female), surgical approach (such as cerebellar hemispherectomy approach, cerebellomedullary fissure approach, etc., specific coding rules are shown in Table 1), tumor histopathological subtype (such as astrocytic lineage tumor, ependymal tumor, medulloblastoma, etc., specific coding rules are shown in Table 2), gender, surgical approach, and pathological type (categorical variables are coded one-hot).

[0030] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism includes normalizing radiomics feature data, which comprises:

[0031] The radiomics feature data are standardized using the first expression, which is:

[0032] Where ∈=1e-6, μ represents the data mean, σ represents the data standard deviation, and x represents the data to be standardized. norm This represents the standardized data;

[0033] The standardized radiomics feature data are normalized to the range of [0,1] by minimum-maximum intensity normalization to obtain normalized radiomics feature data.

[0034] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism includes encoding and processing clinical data, comprising:

[0035] Uncategorical variables in clinical data are coded according to preset rules;

[0036] One-hot encoding is performed on categorical variables in clinical data.

[0037] According to the present invention, an AI-based automated risk assessment model for postoperative cerebellar mutism is constructed using a variety of machine learning algorithms, including any one of the following or any combination thereof: logistic regression, random forest, support vector machine, extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost). The final cerebellar mutism risk assessment model is a model trained using random forest.

[0038] According to the present invention, a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism includes, in which the performance of multiple candidate cerebellar mutism risk assessment models is evaluated to obtain a final cerebellar mutism risk assessment model, comprising:

[0039] The performance of multiple candidate cerebellar mutism risk assessment models was evaluated using performance metrics to obtain the final cerebellar mutism risk assessment model. The performance metrics include any one or any combination of the following: area under the ROC curve (AUC), F1 score, precision, recall, and accuracy.

[0040] This invention also provides a postoperative cerebellar mutism risk assessment system, comprising:

[0041] The test data receiving module is used to: receive MRI image data and clinical data of the test subject from at least one terminal, wherein the test subject is a patient with a posterior fossa tumor;

[0042] The postoperative cerebellar mutism risk assessment module is used to: obtain the postoperative cerebellar mutism risk assessment result of the test subject by using the final postoperative cerebellar mutism risk assessment model obtained by the AI-based automated risk assessment model for postoperative cerebellar mutism constructed by any of the above-mentioned methods based on the MRI image data and clinical data of the test subject.

[0043] The postoperative cerebellar mutism risk assessment result output module is used to send the postoperative cerebellar mutism risk assessment result of the subject to at least one terminal.

[0044] It should be noted that a terminal refers to an input / output device connected to a computer system. Depending on the function, terminals can be divided into various types: smart terminals or intelligent terminals, dumb terminals, interactive terminals or online terminals. Specifically, a terminal can be various mobile communication devices, such as mobile phones and tablets. This article aims to provide users with the function of inputting data and outputting data.

[0045] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the above-described method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism.

[0046] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism as described above.

[0047] The present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute any of the above-described methods for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism.

[0048] This invention provides a method and system for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism. Based on MRI image data of the target population, combined with tumor region label data obtained from a semi-automatic annotation tool and manually corrected tumor region label data, a segmentation model with high accuracy and noise robustness is obtained through iterative optimization. High-quality tumor region label data for the target population is obtained using the segmentation model, and then combined with MRI image data, clinical data, and label data indicating whether or not the patient has cerebellar mutism. The model is trained to learn the clinical characteristics of patients with posterior fossa tumors who have cerebellar mutism and those without, as well as the characteristics of tumor regions on MRI images. This results in a cerebellar mutism risk assessment model with strong generalization ability, capable of accurately predicting CMS risk. This provides objective and quantitative decision support for clinicians, aiding in preoperative risk stratification, personalized surgical planning, and early postoperative intervention, thereby improving patient prognosis. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating the construction method of an AI-based automated risk assessment model for postoperative cerebellar mutism provided by the present invention.

[0051] Figure 2 A flowchart illustrating the data labeling process.

[0052] Figure 3 The diagram illustrates the annotation results of human-machine collaboration. Figure A shows the semi-automatic annotation results of ITK-SNAP; Figure B shows the final annotation results after two rounds of iteration.

[0053] Figure 4 This shows an overview of the Auto3dseg framework.

[0054] Figure 5 The results show a three-dimensional visualization of the tumor.

[0055] Figure 6 yes Figure 2 The enlarged diagram in the lower left shows a performance comparison of multiple candidate risk assessment models for cerebellar mutism syndrome.

[0056] Figure 7 This is a schematic diagram of the structure of a postoperative cerebellar mutism risk assessment system provided by the present invention.

[0057] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0059] Figure 1 This is a flowchart illustrating a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism provided by the present invention. The execution entity of this method can be any applicable terminal-side device or network-side device, such as a device for constructing a postoperative cerebellar mutism risk assessment model.

[0060] See Figure 1 The present invention provides a method for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism, which may include:

[0061] S110. Obtain MRI imaging data, clinical data, and labeling data on whether the target group has cerebellar mutism. The target group is patients with posterior fossa tumors who have undergone surgery for posterior fossa tumors.

[0062] In one embodiment, the MRI imaging data and clinical data of the target population can be obtained from the hospital's medical database, and the label of whether or not the patient has cerebellar mutism can be pre-labeled manually. The clinical data includes any one or any combination of the following: age (continuous numerical value), sex (male / female), surgical approach (e.g., cerebellar hemispherectomy approach, cerebellomedullary cleft approach, etc.), tumor histopathological subtype (e.g., astrocytic lineage tumor, ependymal tumor, medulloblastoma, etc.), and pathological type.

[0063] S120. Based on the MRI image data of the target group, a segmentation model is trained by using tumor region label data obtained through a semi-automatic annotation tool (e.g., ITK-SNAP 3.8.0) and manually corrected tumor region label data, and the final tumor region label data is obtained using the segmentation model.

[0064] See Figures 2 to 5In one embodiment, S120 may include:

[0065] A semi-automatic annotation tool was used to delineate tumor regions in the MRI image data of the target population to obtain the first tumor region label data;

[0066] Based on the MRI image data of the target population and the first tumor region label data, multiple segmentation models are trained using strategies such as 5-fold cross-validation. Segmentation models that meet the first preset conditions are selected from these models. Then, based on the MRI image data of the target population, the second tumor region label data is obtained using the segmentation models that meet the first preset conditions. The multiple segmentation models include any one or any combination of the following: nnU-Net (v2, ResEncM configuration) deep learning segmentation model, SegResNet deep learning segmentation model, SwinUNetR deep learning segmentation model, DiNTS deep learning segmentation model, and Auto3DSeg deep learning segmentation model. The first preset conditions include a Dice coefficient not lower than a first preset threshold (0.75 in this embodiment) on the validation set. Furthermore, when the number of segmentation models meeting the first preset conditions is greater than one, the segmentation results of all segmentation models meeting the first preset conditions are averaged using a probability graph to enhance robustness to label noise, thus obtaining the second tumor region label data.

[0067] The system receives manually corrected second tumor region label data. Experienced radiologists can review and manually correct the second tumor region label data using professional software such as 3D Slicer, paying particular attention to key areas such as tumor boundaries, tumor-brainstem junction, and cystic necrosis areas.

[0068] Based on the MRI image data of the target population and the manually corrected second tumor region label data, multiple segmentation models are retrained (this process can be iterated, for example, 2 iterations). The segmentation model that meets the second preset conditions (e.g., Dice coefficient reaches 0.91±0.03, average surface distance reaches 0.7±0.3mm) is selected from the multiple segmentation models. This model can be used for fast and accurate automatic segmentation of tumor regions (ROI) on new preoperative MRI images.

[0069] Based on the MRI image data of the target population, the final tumor region label data is obtained using a segmentation model that meets the second preset conditions.

[0070] S130. Based on the MRI image data of the target group, the final tumor region label data, clinical data, and label data of whether or not the patient has cerebellar mutism, the training model learns the characteristics of the tumor region on the MRI images corresponding to the clinical data of patients with posterior fossa tumors and patients without cerebellar mutism, and obtains a risk assessment model for cerebellar mutism syndrome.

[0071] In one embodiment, S130 may include:

[0072] Based on the MRI imaging data of the target population and the final tumor region labeling data, radiomics feature data were obtained. These radiomics feature data include any one or any combination of the following: first-order statistical features, two-dimensional morphological features, and three-dimensional morphological features. First-order statistical features include any one or any combination of the following: Energy, Entropy, Mean, Median, Minimum, Maximum, Range, Interquartile Range, Standard Deviation, Variance, Skewness, Kurtosis, Root Mean Square, and 10th Percentile. Two-dimensional morphological features include any one or any combination of the following: Perimeter... r), MeshSurface, PixelSurface, Maximum Diameter, Perimeter Surface Ratio; 3D morphological features include any one or any combination of the following: MeshVolume, SurfaceArea, Sphericity, Maximum 3D Diameter, Major Axis Length, Minor Axis Length, Least Axis Length, SurfaceVolume Ratio, Compactness1, Elongation. These features quantify the distribution characteristics of pixel (or voxel) intensity within the tumor region through grayscale statistics.

[0073] The radiomics feature data are normalized, specifically by standardizing the radiomics feature data using a first expression, which is: Where ∈=1e-6, μ represents the data mean, σ represents the data standard deviation, and x represents the data to be standardized. norm The data is represented by the standardized data. Then, the standardized radiomics feature data is normalized to the range of [0,1] by the minimum-maximum intensity normalization to obtain the normalized radiomics feature data.

[0074] The clinical data is coded. Non-categorical variables in the clinical data are coded according to preset rules. For example, surgical approaches such as cerebellar hemispherectomy and cerebellomedullary fissure approach are coded according to Table 1, and tumor histopathological subtypes such as astrocytic lineage tumors, ependymal tumors, and medulloblastomas are coded according to Table 2. Categorical variables in the clinical data are coded one-hot.

[0075] Table 1

[0076]

[0077]

[0078] Table 2

[0079]

[0080] By fusing normalized radiomics feature data and encoded clinical data, multiple machine learning algorithms were used for system evaluation and hyperparameter optimization (e.g., grid search and 5-fold cross-validation). The models were trained to learn the features of the tumor region on MRI images corresponding to the clinical data of patients with posterior fossa tumors and patients without cerebellar mutism, respectively, resulting in multiple candidate risk assessment models for cerebellar mutism syndrome. Among them, the multiple machine learning algorithms include any one of the following or any combination thereof: logistic regression (log_reg), random forest (rf), support vector machine (svm), extreme gradient boosting (xgb, XGBoost), and adaptive boosting (AdaBoost);

[0081] Performance metrics were used to evaluate the risk assessment models for cerebellar mutism syndrome, resulting in the final cerebellar mutism syndrome risk assessment model. The performance metrics included any one or any combination of the following: area under the ROC curve (AUC), F1 score, precision, recall, and accuracy. (See [link to relevant documentation]). Figure 6 In this embodiment, the final risk assessment model for cerebellar mutism syndrome is a model trained using random forest.

[0082] The following two specific embodiments describe the construction method of an AI-based automated risk assessment model for postoperative cerebellar mutism provided by the present invention (the experimental design of the embodiments has passed the ethical approval standards of Beijing Children's Hospital affiliated to Capital Medical University).

[0083] Example 1: Construction of a Human-Machine Collaborative Annotation and Automatic Tumor Segmentation Model

[0084] Data were obtained from preoperative T1-weighted MRI images (e.g., using a 3.0T or 1.5T GE / Philips scanner, 12-channel head coil, TR=500ms, TE=12ms, flip angle 70°, slice thickness 5mm) of 105 pediatric patients with posterior fossa tumors diagnosed at Beijing Children's Hospital between January 2015 and December 2020 (including patients aged 0-18 years who underwent surgical treatment, had a clear diagnosis or exclusion of CMS, and were followed up at least once, excluding cases with arachnoid cysts, lack of preoperative MRI, or only biopsy). Preprocessing included intensity normalization with a fixed width of 5 bins, spline interpolation, and resampling to 1x1x1 mm. 3 ).

[0085] 1. Two-stage human-machine collaborative annotation process:

[0086] Phase 1: Initial model building and noise label generation.

[0087] A pediatric neurosurgeon with over 3 years of experience used the semi-automatic annotation tool ITK-SNAP 3.8.0 to perform preliminary tumor region delineation on 52 T1-weighted MRI images, with annotation time per case controlled within 30-45 minutes. Labels generated at this stage may contain discontinuities, insufficient accuracy, and random noise.

[0088] The NVIDIA Auto3DSeg framework was adopted, which integrates three different deep learning segmentation architectures: SegResNet, DiNTS, and SwinUNetR. Using this dataset of 52 noisy labeled examples, 5-fold cross-validation training was performed (generating a total of 3x5=15 candidate segmentation models).

[0089] From 15 candidate segmentation models, the top 5 models with a Dice coefficient of at least 0.75 (or the best overall performance) on the validation set were selected (e.g., 2 SwinUNetR, 2 SegResNet, and 1 DiNTS). A preliminary segmentation mask for all 105 images was generated by probabilistically averaging the predictions of these 5 models. This ensemble approach leverages the noise tolerance of deep neural networks to improve robustness to labeled artifacts.

[0090] Phase Two: Iterative Optimization and Construction of the Gold Standard Dataset.

[0091] A senior radiologist (with more than 3 years of experience) reviewed and manually corrected each of the 105 segmentation masks generated in Phase 1 using 3D Slicer 5.2.2 software. The corrections focused on the tumor-brainstem junction (with an error tolerance of ±1.2 mm), cystic necrosis areas (requiring careful differentiation of non-enhanced areas), and other inaccurately segmented areas.

[0092] The corrected labels are used as new training data to retrain the model in the Auto3DSeg framework.

[0093] This "correction-retraining" process is performed in two iterations. After each iteration, the annotation quality is evaluated, for example, by calculating the Dice coefficient and the average surface distance. After two iterations, the final segmentation dataset achieves a Dice coefficient of 0.91 ± 0.03 and an average surface distance of 0.7 ± 0.3 mm.

[0094] Training Completion Conditions and Quality Control: For the segmentation model training, during the iterative optimization phase, training is considered complete when the performance indicators (such as the Dice coefficient) do not significantly improve after two consecutive iterations, or when the preset target is reached (such as Dice > 0.90), and the segmentation quality is confirmed by a senior physician to meet clinical requirements. For 10 particularly difficult cases, two senior radiologists independently optimized the annotations. If the Dice coefficients of both were higher than 0.9, the annotations of the first physician were adopted; otherwise, the annotations were done by a third senior physician, and the average result was taken as the gold standard. In this embodiment, after two rounds of iterative optimization, the annotation quality has reached a high level.

[0095] Training and selection of automated tumor segmentation models:

[0096] Using the aforementioned 105 high-quality labeled datasets, they were divided into a training / validation set (e.g., 88 examples) and an independent test set (e.g., 17 examples). Five-fold cross-validation was employed on the training / validation sets to train and evaluate various state-of-the-art 3D segmentation models, including nnU-Net (v2, ResEncM configuration), SegResNet, SwinUNetR, DiNTS, and the Auto3DSeg ensemble model. The training and evaluation environment was a system configured with Ubuntu 22.04.5LTS, an Intel Xeon Silver 4316 CPU, 256GB RAM, and four NVIDIA A40 GPUs, using PyTorch (v2.3.1), nnU-Net (v2.6.0), and MONAI (v1.4). The models were trained using their default configurations and standard preprocessing and post-processing were applied (e.g., retaining only the largest connected components).

[0097] The performance of each model was evaluated on an independent test set. For example, nnU-Net achieved a Dice coefficient of 0.922 and an HD95 distance of 1.151 mm, while the Auto3DSeg ensemble model achieved a Dice coefficient of 0.887 and an HD95 distance of 1.047 mm. Based on the actual application requirements (such as a balance between accuracy, speed, and robustness), the optimal segmentation model was selected for the subsequent online service system.

[0098] Table 3 Performance results of different segmentation models

[0099] Segmentation method Dice NN dist mm Hd95 dist mm nnUnet 0.922(0.192) 0.656(1.163) 1.151(2.197) Segresnet 0.876(0.242) 1.517(2.624) 1.173(3.589) DiNTs 0.896(0.227) 1.574(2.233) 1.031(3.730) swinunetr 0.903(0.227) 1.103(2.375) 1.079(2.005) Auto3dseg Aggregate Model 0.887(0.263) 0.888(2.014) 1.047(2.228)

[0100] Example 2: Construction and Validation of the CMS Risk Assessment Model

[0101] We used the 105 patient dataset constructed in Example 1 (which includes high-quality tumor segmentation ROI and clinical information).

[0102] The 105-case dataset was stratified and divided into a training set (e.g., 75 cases) and a test set (e.g., 30 cases) using a 7:3 ratio. On the training set, 5-fold cross-validation was used to optimize hyperparameters and evaluate the performance of five machine learning classifiers: Logistic Regression, Random Forest, Support Vector Machine, XGBoost, and AdaBoost. Hyperparameter optimization was performed using a grid search. The optimal model was selected based on its average performance metrics in cross-validation (e.g., AUC, F1 score, precision, recall, accuracy) and its performance on the independent test set.

[0103] Taking this example, preoperative MRI images and clinical data of 105 pediatric patients with posterior fossa tumors from Beijing Children's Hospital were obtained (the clinical data used in this embodiment includes age, gender, surgical approach, and pathological type). The CMS risk prediction model was trained according to the scheme provided in this embodiment. After comparison, the random forest model showed superior overall performance (or, depending on the specific application scenario, such as an emphasis on the F1 score). Its optimal parameter combination is: 120 estimators (n_estimators) and a maximum depth of 5 layers (max_depth). The model was validated on an independent test set, achieving an area under the ROC curve (AUC) of 0.82 and an F1 score of 0.80 (this F1 value is for example; actual validation should be considered, as the original validation set F1 score was 0.88).

[0104] Feature importance analysis shows that features that contribute significantly to CMS risk prediction may include original first-order skewness, original shape elongation, original shape flatness, surgical age, and original shape sphericity.

[0105] The radiomics features (first-order statistical features and morphological features) extracted in this embodiment can capture the heterogeneity and morphological specificity of tumor regions at the microscopic level. These are potential biological markers that are difficult to detect with the naked eye but may be related to the occurrence of CMS. Clinical features (such as age, surgical approach, and pathological type) provide macroscopic patient background information and tumor biological behavior information. The occurrence of CMS is a complex multifactorial process. The selected random forest algorithm (or other high-performing machine learning algorithms such as SVM) is good at handling high-dimensional data and can effectively learn the nonlinear relationships and complex interactions between these multimodal features, thereby constructing a more accurate predictive model than traditional methods. This directly meets the urgent clinical need for objective and quantitative preoperative assessment of CMS risk to guide personalized treatment and improve patient prognosis.

[0106] This invention provides a method and system for constructing an AI-based automated risk assessment model for postoperative cerebellar mutism. Based on MRI image data of the target population, combined with tumor region label data obtained from a semi-automatic annotation tool and manually corrected tumor region label data, the annotation time is significantly shortened, reducing over-reliance on experienced physicians. Through iterative model optimization, a segmentation model with high accuracy and noise robustness is obtained, improving the consistency and accuracy of annotation and laying a solid data foundation for subsequent CMS model training. High-quality tumor region label data of the target population is obtained using the segmentation model, and then combined with MRI image data, clinical data, and label data indicating whether or not the patient has cerebellar mutism. The model is trained to learn the clinical characteristics and tumor region features on MRI images of patients with posterior fossa tumors and those without cerebellar mutism, resulting in a cerebellar mutism risk assessment model with strong generalization ability. On the test set, it achieves an AUC value of 0.82, accurately predicting CMS risk and providing objective and quantitative decision support for clinicians. This facilitates preoperative risk stratification, personalized surgical planning, and early postoperative intervention, thereby improving patient prognosis.

[0107] The postoperative cerebellar mutism risk assessment system provided by this invention will be described below. The postoperative cerebellar mutism risk assessment system described below can be referred to in correspondence with the construction method of the AI-based automated risk assessment model for postoperative cerebellar mutism described above.

[0108] See Figure 7 The present invention provides a postoperative cerebellar mutism risk assessment system, which may include:

[0109] The test data receiving module is used to: receive MRI image data and clinical data of the test subject from at least one terminal, wherein the test subject is a patient with a posterior fossa tumor;

[0110] The postoperative cerebellar mutism risk assessment module is used to: obtain the postoperative cerebellar mutism risk assessment result of the test subject by using the final postoperative cerebellar mutism risk assessment model obtained by the AI-based automated risk assessment model for postoperative cerebellar mutism constructed by any of the above-mentioned methods based on the MRI image data and clinical data of the test subject.

[0111] The postoperative cerebellar mutism risk assessment result output module is used to send the postoperative cerebellar mutism risk assessment result of the subject to at least one terminal.

[0112] The obtained postoperative cerebellar mutism risk assessment model can be used to deploy an online prediction service system.

[0113] System Architecture:

[0114] Frontend: Developed using the Vue 3.0 framework and ECharts 5.0 charting library. Key features include: a user-friendly drag-and-drop upload interface for DICOM files (or other standard medical image formats), supporting files up to 2GB; and 3D tumor rendering and interaction based on Three.js, supporting linked display of axial, sagittal, and coronal planes for easy and intuitive viewing of tumor morphology and segmentation results (e.g., ...). Figure 4 (as shown); the generation and display of CMS risk assessment reports, including risk probability values, risk levels (such as low / medium / high), and visualization of key contributing characteristics (such as radar charts).

[0115] Backend: A RESTful API service cluster is built using Python Flask 2.0 (or other frameworks such as Django).

[0116] Database: An appropriate database (such as PostgreSQL, MongoDB) can be selected to store user information, uploaded records, prediction results, etc. (if persistent storage is required).

[0117] Deployment plan:

[0118] We employ Docker containerization technology to encapsulate and deploy front-end and back-end services. We utilize tools such as Docker Compose or Kubernetes for container orchestration and management, facilitating rapid service deployment, horizontal scaling, version control, and cross-platform migration.

[0119] Performance metrics (examples):

[0120] Single tumor segmentation time (using A40 GPU): approximately 32 ± 1.3 seconds.

[0121] Radiomics feature extraction time: approximately 5 ± 0.4 seconds.

[0122] API average response latency (P95 latency): less than 3 seconds.

[0123] By integrating the above functional modules, an online CMS risk assessment system can be constructed. Users (such as clinicians) can upload preoperative MRI images of patients (e.g., DICOM format files) via a web interface. The system automatically performs tumor segmentation, feature extraction, and risk prediction, and returns a comprehensive report that includes CMS risk probability (classified as low, medium, and high risk), contribution of key features (e.g., TOP5 feature radar chart), and 3D visualization of the tumor (e.g., Three.js rendering supporting axial, sagittal, and coronal planes). This module can use frameworks such as Flask to build the backend RESTful API, Vue.js to build the frontend interactive interface, and containerization technologies such as Docker for deployment to ensure service stability and scalability. The provided online service system automates and simplifies the complex model prediction process. The systematic process and objective evaluation indicators help promote the standardization and reproducibility of CMS risk prediction research.

[0124] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the following steps:

[0125] The MRI image data and clinical data of the subject are received from at least one terminal, wherein the subject is a patient with a posterior fossa tumor;

[0126] Based on the MRI imaging data and clinical data of the test subject, the final postoperative cerebellar mutism risk assessment model is obtained by constructing the AI-based automated risk assessment model for postoperative cerebellar mutism as described above, and the postoperative cerebellar mutism risk assessment result of the test subject is obtained.

[0127] The postoperative cerebellar mutism risk assessment results of the test subjects are sent to at least one terminal.

[0128] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps:

[0130] The MRI image data and clinical data of the subject are received from at least one terminal, wherein the subject is a patient with a posterior fossa tumor;

[0131] Based on the MRI imaging data and clinical data of the test subject, the final postoperative cerebellar mutism risk assessment model is obtained by constructing the AI-based automated risk assessment model for postoperative cerebellar mutism as described above, and the postoperative cerebellar mutism risk assessment result of the test subject is obtained.

[0132] The postoperative cerebellar mutism risk assessment results of the test subjects are sent to at least one terminal.

[0133] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0134] The MRI image data and clinical data of the subject are received from at least one terminal, wherein the subject is a patient with a posterior fossa tumor;

[0135] Based on the MRI imaging data and clinical data of the test subject, the final postoperative cerebellar mutism risk assessment model is obtained by constructing the AI-based automated risk assessment model for postoperative cerebellar mutism as described above, and the postoperative cerebellar mutism risk assessment result of the test subject is obtained.

[0136] The postoperative cerebellar mutism risk assessment results of the test subjects are sent to at least one terminal.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0138] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing an AI-based postoperative cerebellar mutism automatic risk assessment model, characterized in that, The method comprises the following steps: obtaining MRI image data, clinical data and label data of whether the target group has cerebellar mutism, wherein the target group is a postoperative posterior fossa tumor patient group after posterior fossa tumor surgery; training a segmentation model according to the MRI image data of the target group, the tumor region label data obtained by using a semi-automatic labeling tool and the tumor region label data manually corrected, and obtaining final tumor region label data by using the segmentation model; training a model to learn the features of the tumor region on the MRI image and the clinical data of the posterior fossa tumor patients with cerebellar mutism and the posterior fossa tumor patients without cerebellar mutism, and obtaining a cerebellar mutism risk assessment model according to the MRI image data of the target group, the final tumor region label data, the clinical data and the label data of whether the target group has cerebellar mutism. According to the MRI image data of the target group, the tumor region label data obtained by using a semi-automatic labeling tool and the tumor region label data manually corrected, the segmentation model is trained, and the final tumor region label data is obtained by using the segmentation model. The method comprises the following steps: using a semi-automatic labeling tool to outline the tumor region of the MRI image data of the target group to obtain first tumor region label data; training a plurality of segmentation models according to the MRI image data of the target group and the first tumor region label data, and selecting a segmentation model meeting a first preset condition from the plurality of segmentation models, and obtaining second tumor region label data by using the segmentation model meeting the first preset condition according to the MRI image data of the target group; receiving the second tumor region label data manually corrected; retraining the plurality of segmentation models according to the MRI image data of the target group and the second tumor region label data manually corrected, and selecting a segmentation model meeting a second preset condition from the plurality of segmentation models; obtaining final tumor region label data by using the segmentation model meeting the second preset condition according to the MRI image data of the target group. 2.The method of claim 1, wherein, The plurality of segmentation models comprise any one or any combination of the following: an nnU-Net deep learning segmentation model, a SegResNet deep learning segmentation model, a SwinUNetR deep learning segmentation model, a DiNTS deep learning segmentation model, an Auto3DSeg deep learning segmentation model, the first preset condition comprises that the Dice coefficient on a validation set is not less than a first preset threshold, and the second preset condition comprises that the Dice coefficient on the validation set is greater than a second preset threshold and the average surface distance is greater than a third preset threshold. When the number of segmentation models meeting the first preset condition is more than one, the probability map average of the segmentation results of all segmentation models meeting the first preset condition is performed to obtain the second tumor region label data. 3.The method of claim 1 or 2, wherein, The model is trained according to the MRI image data of the target group, the final tumor region label data, the clinical data, and the label data of whether the patient has cerebellar mutism, to learn the features of the tumor region on the MRI image corresponding to the clinical data of the patient with cerebellar mutism and the patient without cerebellar mutism, and a cerebellar mutism risk assessment model is obtained, including: Obtaining radiomics feature data from the MRI image data and the final tumor region label data of the target group; Normalizing the radiomics feature data; Encoding the clinical data; According to the normalized radiomics feature data and the encoded clinical data, a plurality of machine learning algorithms are used to train the model to learn the features of the tumor region on the MRI image corresponding to the clinical data of the patient with cerebellar mutism and the patient without cerebellar mutism, and a plurality of candidate cerebellar mutism risk assessment models are obtained; Performance evaluation is performed on the plurality of candidate cerebellar mutism risk assessment models to obtain a final cerebellar mutism risk assessment model. 4.The method of claim 3, wherein the method further comprises: determining a risk of postoperative cerebellar mutism of the patient based on the AI-based model. The radiomics feature data includes any one or any combination of the following: first-order statistical features, two-dimensional morphological features, and three-dimensional morphological features, wherein the first-order statistical features include any one or any combination of the following: energy, entropy, mean, median, minimum value, maximum value, range, interquartile range, standard deviation, variance, skewness, kurtosis, root mean square, and 10th percentile; the two-dimensional morphological features include any one or any combination of the following: perimeter, grid surface, pixel surface, maximum diameter, perimeter-to-surface ratio; the three-dimensional morphological features include any one or any combination of the following: grid volume, surface area, sphericity, maximum three-dimensional diameter, long axis length, short axis length, minimum axis length, surface area-to-volume ratio, compactness, and elongation; and / or The clinical data includes any one or any combination of the following: age, gender, surgical approach, tumor histopathology subtype, and pathological type. 5.The method of claim 3, wherein the method further comprises: determining a risk of postoperative cerebellar mutism of the patient based on the AI-based model. The normalization of the radiomics feature data includes: Standardizing the radiomics feature data using a first expression, wherein the first expression is: wherein wherein denotes the data mean, denotes the data standard deviation, denotes the data to be normalized, denotes the normalized data; Performing minimum-maximum intensity normalization on the standardized radiomics feature data to the range [0, 1] to obtain the normalized radiomics feature data; and / or The encoding of the clinical data includes: Encoding non-categorical variables in the clinical data according to a preset rule; Performing one-hot encoding on categorical variables in the clinical data. 6.The method of claim 3, wherein the method further comprises: determining a risk of postoperative cerebellar mutism of the patient based on the AI-based model. The plurality of machine learning algorithms includes any one or any combination of the following: logistic regression, random forest, support vector machine, extreme gradient boosting, and adaptive boosting, and the final cerebellar mutism risk assessment model is a model trained using random forest.

7. A postoperative cerebellar mutism risk assessment system, characterized by, Including: A to-be-tested data receiving module configured to receive MRI image data and clinical data of a to-be-tested person from at least one terminal, wherein the to-be-tested person is a patient with a posterior fossa tumor; A postoperative cerebellar mutism risk assessment module is configured to: obtain a postoperative cerebellar mutism risk assessment result of a to-be-tested person according to MRI image data and clinical data of the to-be-tested person and a final postoperative cerebellar mutism risk assessment model obtained by the method for constructing an AI-based postoperative cerebellar mutism automatic risk assessment model according to any one of claims 1-6; and A postoperative cerebellar mutism risk assessment result output module is configured to: send the postoperative cerebellar mutism risk assessment result of the to-be-tested person to at least one terminal.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the following steps when executing the program: receiving, from at least one terminal, MRI image data and clinical data of a to-be-tested person, wherein the to-be-tested person is a posterior fossa tumor patient; obtaining a postoperative cerebellar mutism risk assessment result of the to-be-tested person according to the MRI image data and the clinical data of the to-be-tested person and a final postoperative cerebellar mutism risk assessment model obtained by the method for constructing an AI-based postoperative cerebellar mutism automatic risk assessment model according to any one of claims 1-6; and sending the postoperative cerebellar mutism risk assessment result of the to-be-tested person to at least one terminal. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the following steps: receiving, from at least one terminal, MRI image data and clinical data of a to-be-tested person, wherein the to-be-tested person is a posterior fossa tumor patient; obtaining a postoperative cerebellar mutism risk assessment result of the to-be-tested person according to the MRI image data and the clinical data of the to-be-tested person and a final postoperative cerebellar mutism risk assessment model obtained by the method for constructing an AI-based postoperative cerebellar mutism automatic risk assessment model according to any one of claims 1-6; and sending the postoperative cerebellar mutism risk assessment result of the to-be-tested person to at least one terminal. The computer program is executed by the processor to implement the following steps: receiving, from at least one terminal, MRI image data and clinical data of a to-be-tested person, wherein the to-be-tested person is a posterior fossa tumor patient; obtaining a postoperative cerebellar mutism risk assessment result of the to-be-tested person according to the MRI image data and the clinical data of the to-be-tested person and a final postoperative cerebellar mutism risk assessment model obtained by the method for constructing an AI-based postoperative cerebellar mutism automatic risk assessment model according to any one of claims 1-6; and sending the postoperative cerebellar mutism risk assessment result of the to-be-tested person to at least one terminal.

Images