Cerebral stroke prognosis early warning system and method based on CTP radiomics and machine learning

Abstract

The invention discloses a cerebral apoplexy prognosis early warning system based on CTP image omics and machine learning, and the system comprises a data collection module which is used for obtaining and standardizing the CTP image data and clinical data of a patient; the radiomics feature extraction module is used for extracting a high-dimensional feature set containing morphological features, textural features and hemodynamic functional features; the machine learning prognosis model module is used for fusing the high-dimensional feature set and the clinical data into a training feature matrix; and a prognosis early warning output module. According to the method, a multi-dimensional and high-dimensional feature set including form, texture and functional parameters is automatically extracted from a CTP image through a radiomics technology, focus microcosmic heterogeneity and hemodynamic information which cannot be recognized by human eyes are deeply excavated, deep features and clinical data are fused, learning is performed by using a machine learning model, and the accuracy and accuracy of the fusion of the deep features and the clinical data are improved. Therefore, the dependence of doctors on personal experience is effectively reduced, and the prognosis early warning result is more objective.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a stroke prognosis early warning system and method based on CTP radiomics and machine learning. Background Technology

[0002] Stroke is an acute cerebrovascular disease with high incidence, disability, and mortality rates, with acute ischemic stroke accounting for the majority. Accurate assessment of patient prognosis is crucial for developing individualized treatment plans, optimizing the allocation of medical resources, and improving patients' long-term quality of life.

[0003] Currently, the assessment of stroke prognosis in clinical practice mainly relies on the following methods:

[0004] Clinical experience-based judgment: Doctors make empirical predictions based on basic clinical information such as the patient's age, neurological deficit score at admission, and past medical history. This method is highly subjective, relies heavily on the doctor's personal experience, lacks objective and quantifiable standards, and leads to inconsistent prediction results and limited accuracy.

[0005] Traditional imaging assessment: Computed tomography (CTP) perfusion imaging has been widely used in the diagnosis of acute stroke. Clinicians typically assess the ischemic penumbra and infarct core by visually interpreting or simply measuring individual hemodynamic parameters generated by CTP, such as cerebral blood flow and cerebral blood volume. However, this method only utilizes surface information from the imaging data and cannot effectively extract and utilize the deeper functional and textural features that reflect tissue heterogeneity and microstructure, resulting in a significant waste of information value.

[0006] Static prediction models: Some existing prediction models based on statistics or early machine learning methods often have the following limitations: First, the model features have limited dimensions and fail to deeply integrate deep radiomics features with clinical data; second, the model is fixed after training and lacks an effective mechanism for continuous learning and updating. As medical practice and population change, the model performance will gradually decline, i.e., the "model aging" problem, making it difficult to guarantee the timeliness and accuracy of long-term applications.

[0007] Therefore, we propose to design a stroke prognostic early warning system and method based on CTP radiomics and machine learning. Summary of the Invention

[0008] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0009] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0010] A stroke prognostic early warning system based on CTP radiomics and machine learning, comprising:

[0011] The data acquisition module is used to acquire and standardize the patient's CTP imaging data and clinical data;

[0012] The radiomics feature extraction module is used to process the CTP image data and extract a high-dimensional feature set including morphological features, texture features and hemodynamic functional features.

[0013] The machine learning prognostic model module is used to fuse the high-dimensional feature set with the clinical data into a training feature matrix, and train a machine learning model based on the matrix to output prognostic prediction results; it includes a model optimization unit for performing automatic hyperparameter tuning and feature selection based on feature importance, which facilitates the improvement of model generalization performance.

[0014] The prognosis warning output module is used to receive data from the patients to be tested and to generate and display stroke prognosis warning information using the trained machine learning model.

[0015] It includes a model performance evaluation unit, which is used to calculate and output one or more metrics among accuracy, sensitivity, specificity and AUC value on the test set to quantify the model's predictive performance.

[0016] Preferably, the specific configuration of the data acquisition module includes: directly acquiring raw image data from the CTP imaging device via the DICOM standard protocol; integrating the patient's clinical data through the data interface of the hospital information system; performing data cleaning and standardization preprocessing on the integrated data; and storing it in a database with hierarchical access permissions.

[0017] Preferably, in the radiomics feature extraction module, the morphological features include volume, surface area, maximum diameter, and sphericity calculated based on the lesion region; the texture features are calculated using gray-level co-occurrence matrix and gray-level run-length matrix algorithms, and include at least contrast, entropy, energy, and run-length non-uniformity; the hemodynamic functional features are obtained by analyzing the contrast agent time-concentration curve, and include at least cerebral blood flow, cerebral blood volume, mean transit time, and time to peak concentration.

[0018] Preferably, in the machine learning prognostic model module, the machine learning algorithm used is one or more of neural networks, random forests, or support vector machines. The model training process includes: dividing the dataset into training and testing sets, using the patient's actual prognostic status as the supervision label, and minimizing the prediction loss function through an iterative optimization algorithm until the model converges.

[0019] Preferably, the specific configuration of the model optimization unit includes: using grid search or Bayesian optimization algorithms to optimize the hidden layer structure, activation function, and learning rate of the neural network, or the maximum depth and minimum number of samples per node of the tree model; using k-fold cross-validation to evaluate the performance of different hyperparameter combinations during training; selecting the AUC value or F1 score as the core performance indicator; and eliminating redundant features and constructing a simplified feature subset by calculating the correlation between features and labels and ranking the importance of features in the model.

[0020] Preferably, in the prognosis warning output module, the warning information includes classification results and probability values, wherein the classification results are at least divided into two levels: "good prognosis" and "poor prognosis", and the probability values ​​are the predicted probability that the patient will reach a specific level of functional recovery.

[0021] A stroke prognostic prediction method based on CTP radiomics and machine learning includes the following steps:

[0022] S1: Collect the patient's CTP imaging data and related clinical data, and perform standardized preprocessing;

[0023] S2: Perform radiomics analysis on the preprocessed CTP images to extract multi-dimensional features including morphological, textural, and functional parameters;

[0024] S3: Integrate the aforementioned multi-dimensional features with clinical data, use machine learning algorithms to train a stroke prognosis prediction model, and optimize and validate the model.

[0025] S4: Input the characteristic data of the patient to be evaluated into the optimized prediction model to generate and output an individualized prognostic warning report.

[0026] Preferably, in step S1, the standardized preprocessing includes: spatial registration, noise filtering, and pixel value normalization of the CTP image data; imputing missing values ​​in the clinical data using the mean, median, or mode according to the data type; and identifying and correcting outliers. For missing values ​​in the clinical data, if the data is numerical, the mean or median is used for imputation; if the data is categorical, the most frequently occurring category is used.

[0027] Preferably, in step S3, the model optimization and validation includes: evaluating model stability within the training set using cross-validation and validating its generalization ability with an external independent dataset; triggering a model update process when the accumulated new data reaches a preset threshold, and retraining the model using the full dataset to ensure the continuity and timeliness of prediction performance. By triggering the model update process through a preset threshold and retraining using the full dataset, such as a combination of historical and new data, the model can continuously learn new patterns and features.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] This invention utilizes radiomics technology to automatically extract multi-dimensional high-dimensional feature sets, including morphological, texture, and functional parameters, from CTP images. It deeply mines the microscopic heterogeneity and hemodynamic information of lesions that are invisible to the human eye. After fusing these deep features with clinical data, machine learning models are used for learning, thereby effectively reducing doctors' reliance on personal experience and making the prognostic warning results more objective. Detailed Implementation

[0030] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be further described in detail below.

[0031] This invention provides a stroke prognostic early warning system based on CTP radiomics and machine learning, comprising:

[0032] The data acquisition module is used to acquire and standardize the patient's CTP imaging data and clinical data.

[0033] The specific configuration of the data acquisition module includes directly acquiring raw image data from CTP imaging equipment via the DICOM standard protocol; integrating patients' clinical data through the data interface of the hospital information system; performing data cleaning and standardization preprocessing on the integrated data; and storing it in a database with hierarchical access permissions.

[0034] The radiomics feature extraction module is used to process CTP image data and extract a high-dimensional feature set including morphological features, texture features, and hemodynamic functional features.

[0035] In the radiomics feature extraction module, morphological features include volume, surface area, maximum diameter, and sphericity calculated based on the lesion region; texture features are calculated using gray-level co-occurrence matrix and gray-level run matrix algorithms, and include at least contrast, entropy, energy, and run length non-uniformity; hemodynamic functional features are obtained by analyzing contrast agent time-concentration curves, and include at least cerebral blood flow, cerebral blood volume, mean transit time, and time to peak.

[0036] This invention constructs a fusion feature matrix, enabling machine learning models to learn the complex, non-linear relationships between radiomics features and clinical indicators. This multimodal fusion analysis can more comprehensively reflect the patient's pathophysiological state, thus providing more reliable and accurate data compared to a single data source.

[0037] The machine learning prognostic model module is used to fuse high-dimensional feature sets with clinical data into a training feature matrix, and to train a machine learning model based on this matrix to output prognostic prediction results. It includes a model optimization unit for performing automatic hyperparameter tuning and feature selection based on feature importance, which facilitates the improvement of model generalization performance.

[0038] The specific configuration of the model optimization unit includes: using grid search or Bayesian optimization algorithms to optimize the hidden layer structure, activation function, and learning rate of the neural network, or the maximum depth and minimum number of samples per node of the tree model; using k-fold cross-validation to evaluate the performance of different hyperparameter combinations during training; and using AUC value or F1 score as the core performance indicator for selection; and eliminating redundant features and constructing a simplified feature subset by calculating the correlation between features and labels and ranking the importance of features in the model.

[0039] This invention incorporates a model optimization unit that employs advanced techniques such as grid search and Bayesian optimization for automatic hyperparameter tuning, combined with k-fold cross-validation and validation on external independent datasets. This combined strategy effectively prevents model overfitting and ensures stable performance on unknown data.

[0040] In the machine learning prognostic model module, the machine learning algorithm used is one or more of neural networks, random forests, or support vector machines. The model training process includes: dividing the dataset into training and testing sets, using the patient's actual prognostic status as the supervision label, and minimizing the prediction loss function through an iterative optimization algorithm until the model converges.

[0041] The prognosis warning output module is used to receive data from patients to be tested and to generate and display stroke prognosis warning information using a trained machine learning model.

[0042] It includes a model performance evaluation unit, which is used to calculate and output one or more metrics among accuracy, sensitivity, specificity and AUC value on the test set to quantify the model's predictive performance.

[0043] In the prognosis warning output module, the warning information includes classification results and probability values. The classification results are divided into at least two levels: "good prognosis" and "poor prognosis". The probability values ​​are the predicted probability that the patient will reach a specific level of functional recovery.

[0044] A stroke prognostic prediction method based on CTP radiomics and machine learning includes the following steps:

[0045] S1: Collect the patient's CTP imaging data and related clinical data, and perform standardized preprocessing.

[0046] Standardized preprocessing includes: spatial registration, noise filtering, and pixel value normalization of CTP image data, normalizing pixel values ​​to the range [0,1], using the formula: Where X is the original pixel value, X min and X max These represent the minimum and maximum values ​​in the original data. For missing values ​​in the clinical data, they are filled using the mean, median, or mode, depending on the data type. Outliers are also identified and corrected.

[0047] For missing values ​​in clinical data, if the data is numerical, the mean or median is used for filling; if the data is categorical, the most frequent category is used for filling. For outliers, box plot analysis is used to identify data below Q1-1.5IQR and above Q3+1.5IQR as outliers, where Q1 is the lower quartile, Q3 is the upper quartile, and IQR = Q3-Q1. These outliers are then corrected or deleted.

[0048] S2: Perform radiomics analysis on the preprocessed CTP images to extract multi-dimensional features including morphological, textural, and functional parameters;

[0049] S3: Integrate multi-dimensional features and clinical data, use machine learning algorithms to train a stroke prognosis prediction model, and optimize and validate the model.

[0050] Model optimization and validation include: using cross-validation to evaluate model stability within the training set and validating its generalization ability with an external independent dataset; when the amount of new data accumulated reaches a preset threshold, triggering the model update process and retraining the model using the full dataset to ensure the continuity and timeliness of prediction performance.

[0051] By triggering the model update process with a preset threshold, the model is retrained using all the data, such as a combination of historical data and new data, so that the model can continuously learn new patterns and features.

[0052] S4: Input the characteristic data of the patient to be evaluated into the optimized prediction model to generate and output an individualized prognostic warning report.

[0053] Working principle:

[0054] Phase 1:

[0055] The data acquisition module communicates directly with the CTP imaging equipment via the standard DICOM protocol, automatically acquiring raw, continuous dynamic perfusion image sequences to ensure the integrity and original accuracy of the image data. The system automatically extracts structured patient data, such as age, gender, NIHSS score at admission, medical history, and vital signs, through a hospital information system interface.

[0056] The integrated data is cleaned to remove duplicate records, correct obvious formatting errors, and handle missing and outlier values.

[0057] Spatial registration of CTP images ensures image alignment at different time points, noise filtering, and pixel value normalization, eliminating interference from equipment differences and random noise on subsequent analysis and providing a standardized data foundation for feature extraction.

[0058] The processed data is categorized and stored in a secure database with strict hierarchical access permissions to ensure patient privacy and data security.

[0059] Phase Two:

[0060] The radiomics feature extraction module first uses automatic or semi-automatic algorithms to accurately delineate the core region and penumbra region of the cerebral ischemia lesion on the preprocessed CTP image.

[0061] Based on the segmented lesions, geometric parameters such as volume, surface area, maximum diameter, and sphericity are calculated to quantify the macroscopic spatial occupancy effect of the lesions.

[0062] Algorithms such as gray-level co-occurrence matrix and gray-level run-length matrix are used to calculate indicators such as contrast, entropy, energy, and run-length non-uniformity. These features can quantify the microscopic heterogeneity and structural complexity within lesions, reflecting the pathological state of tissues and providing deep information that cannot be obtained through manual interpretation.

[0063] By analyzing the contrast agent concentration curve of each pixel over time, parameters such as cerebral blood flow, cerebral blood volume, mean transit time, and time to peak concentration were quantitatively calculated. These parameters directly reflect the perfusion status and ischemia degree of brain tissue.

[0064] Ultimately, these features are combined into a high-dimensional feature vector representing each patient, namely the radiomics feature set.

[0065] Phase Three:

[0066] The machine learning prognostic model module integrates the high-dimensional radiomics feature set extracted in the second stage with the clinical data compiled in the first stage to form a comprehensive feature matrix. Each patient is represented by a row in this matrix, and their actual prognostic status, such as the mRS score after 90 days, serves as the label for supervised learning.

[0067] The system can choose from a variety of algorithms such as neural networks, random forests, and support vector machines for modeling.

[0068] The model optimization unit is activated, employing algorithms such as grid search or Bayesian optimization to automatically find the optimal configuration of the model, including the number of neural network layers, learning rate, and tree depth in the random forest. Evaluation uses k-fold cross-validation, with AUC or F1 score as the core metric to ensure that the optimal solution is found and to avoid overfitting.

[0069] Meanwhile, by calculating the correlation between features and prognosis and the importance of features evaluated by the model itself, redundant features are eliminated, and a more concise and complete feature subset is constructed to improve model efficiency and generalization ability.

[0070] The trained model must not only be evaluated on a reserved test set, but also validated on a completely independent external dataset to demonstrate its true clinical applicability.

[0071] After the system is put into clinical use, it will continuously monitor the accumulation of new data. When the amount of data reaches a preset threshold, the system will automatically trigger the model update process, merging the new data with historical data and re-executing the entire training and optimization process.

[0072] This invention innovatively designs a trigger-based update mechanism based on a data accumulation threshold. When new data reaches a certain scale, the system automatically initiates retraining, updating the model using the full dataset. This enables the model to continuously learn new clinical patterns and characteristics, dynamically adapting to changes in medical practice. It solves the "model aging" problem, ensuring the long-term effectiveness and timeliness of the system.

[0073] Phase Four:

[0074] The prognostic warning output module repeats the first and second stages for a newly admitted patient, obtains its standardized feature data, and inputs it into the trained and optimized final model.

[0075] The model outputs two forms of prediction results, directly providing intuitive level judgments such as "good prognosis" or "poor prognosis," which facilitates doctors' quick decision-making.

[0076] It outputs the specific probability value of a patient recovering to a certain functional level at a specific time point. For example, "the probability of this patient having mRS≤2 at 90 days is 78%". This probability output provides a richer and more detailed risk assessment, which is convenient for doctors to refer to.

[0077] The model performance evaluation unit will periodically or after each update calculate metrics such as accuracy, sensitivity, specificity, and AUC on the test set and present them to the administrator in the form of a report to ensure that the reliability of the system is always known and controllable.

Claims

1. A stroke prognostic early warning system based on CTP radiomics and machine learning, characterized in that, include: The data acquisition module is used to acquire and standardize the patient's CTP imaging data and clinical data; The radiomics feature extraction module is used to process the CTP image data and extract a high-dimensional feature set including morphological features, texture features and hemodynamic functional features. The machine learning prognostic model module is used to fuse the high-dimensional feature set with the clinical data into a training feature matrix, and train a machine learning model based on the matrix to output prognostic prediction results. Includes a model optimization unit for performing automatic hyperparameter tuning and feature selection based on feature importance; The prognosis warning output module is used to receive data from the patients to be tested and to generate and display stroke prognosis warning information using the trained machine learning model. It includes a model performance evaluation unit, which is used to calculate and output one or more metrics among accuracy, sensitivity, specificity and AUC value on the test set to quantify the model's predictive performance.

2. The stroke prognostic early warning system based on CTP radiomics and machine learning according to claim 1, characterized in that, The specific configuration of the data acquisition module includes directly acquiring raw image data from the CTP imaging device via the DICOM standard protocol; integrating patient clinical data through the data interface of the hospital information system; performing data cleaning and standardization preprocessing on the integrated data; and storing it in a database with hierarchical access permissions.

3. The stroke prognostic early warning system based on CTP radiomics and machine learning according to claim 1, characterized in that, In the radiomics feature extraction module, the morphological features include volume, surface area, maximum diameter, and sphericity calculated based on the lesion region; the texture features are calculated using gray-level co-occurrence matrix and gray-level run-length matrix algorithms, and include at least contrast, entropy, energy, and run-length non-uniformity; the hemodynamic functional features are obtained by analyzing the contrast agent time-concentration curve, and include at least cerebral blood flow, cerebral blood volume, mean transit time, and time to peak.

4. The stroke prognostic early warning system based on CTP radiomics and machine learning according to claim 1, characterized in that, The machine learning prognostic model module uses one or more of the following machine learning algorithms: neural network, random forest, or support vector machine. The model training process includes: dividing the dataset into training set and test set, using the patient's actual prognostic status as the supervision label, and minimizing the prediction loss function through an iterative optimization algorithm until the model converges.

5. A stroke prognostic early warning system based on CTP radiomics and machine learning according to claim 1, characterized in that, The specific configuration of the model optimization unit includes: using grid search or Bayesian optimization algorithms to optimize the hidden layer structure, activation function, and learning rate of the neural network, or the maximum depth and minimum number of samples per node of the tree model; using k-fold cross-validation to evaluate the performance of different hyperparameter combinations during training; and using AUC value or F1 score as the core performance indicator for selection; and eliminating redundant features and constructing a simplified feature subset by calculating the correlation between features and labels and ranking the importance of features in the model.

6. A stroke prognostic early warning system based on CTP radiomics and machine learning according to claim 1, characterized in that, In the pre-early warning output module, the warning information includes classification results and probability values.

7. A stroke prognostic early warning method based on CTP radiomics and machine learning according to any one of claims 1-6, characterized in that, Includes the following steps: S1: Collect the patient's CTP imaging data and related clinical data, and perform standardized preprocessing; S2: Perform radiomics analysis on the preprocessed CTP images to extract multi-dimensional features including morphological, textural, and functional parameters; S3: Integrate the aforementioned multi-dimensional features with clinical data, use machine learning algorithms to train a stroke prognosis prediction model, and optimize and validate the model. S4: Input the characteristic data of the patient to be evaluated into the optimized prediction model to generate and output an individualized prognostic warning report.

8. A stroke prognostic early warning method based on CTP radiomics and machine learning according to claim 7, characterized in that, In S1, the standardized preprocessing includes: spatial registration, noise filtering, and pixel value normalization of CTP image data; imputation of missing values ​​in clinical data using the mean, median, or mode according to the data type; and identification and correction of outliers.

9. A stroke prognostic early warning method based on CTP radiomics and machine learning according to claim 7, characterized in that, In S3, the model optimization verification includes: using cross-validation to evaluate the model's stability within the training set and verifying its generalization ability with an external independent dataset; when the amount of new data accumulated reaches a preset threshold, triggering the model update process and retraining the model using the full amount of data to ensure the continuity and timeliness of the prediction performance.