Probability prediction system for machine learning of liver cancer microvessel invasion based on radiomics and clinical features
By utilizing a machine learning system based on radiomics and clinical features, and employing radiomics feature extraction and various machine learning algorithms, the problem of preoperative non-invasive prediction of microvascular invasion (MVI) in liver cancer has been solved, achieving high-precision and reliable MVI prediction.
Patent Information
- Application Number
- CN202511345222.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-19
- Publication Date
- 2025-12-30
AI Technical Summary
Current technologies for diagnosing microvascular invasion (MVI) in liver cancer rely on invasive pathological examinations and highly subjective imaging assessments, lacking objective, accurate, and reliable preoperative prediction methods.
A machine learning system based on radiomics and clinical features is used to collect and process CT data, and a predictive model is constructed using radiomics feature extraction and various machine learning algorithms to achieve preoperative probability prediction of MVI.
It enables non-invasive and accurate MVI prediction, reduces reliance on invasive examinations, improves the objectivity and reliability of prediction, and provides quantitative support for clinical decision-making.
Smart Images
Figure CN121237409A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of medical image analysis, and particularly relates to a liver cancer microvascular invasion (MVI) probability prediction system based on imageomics and clinical features and machine learning of medical imageomics features. BACKGROUND
[0002] Microvascular invasion is an important pathological feature of malignant tumors such as hepatocellular carcinoma (HCC), and its existence significantly affects the prognosis and recurrence risk of patients. Studies have shown that the 5-year survival rate of MVI-positive patients is significantly lower than that of MVI-negative patients, and the postoperative recurrence rate is significantly increased. Therefore, preoperative accurate prediction of MVI status is of great significance for developing individualized treatment plans, assessing surgical risks and prognosis. Radiomics is a technology for high-throughput extraction of quantitative features from medical images. By converting medical images into high-dimensional data that can be mined, combined with machine learning algorithms to establish prediction models. This technology can capture image features that the human eye cannot identify, providing a new approach to disease diagnosis and prognosis prediction.
[0003] Currently, MVI diagnosis relies on postoperative pathology or puncture biopsy with sampling errors, which is lagging and risky. In addition, traditional imaging evaluation is highly dependent on physician experience, is highly subjective and has poor reproducibility. Moreover, most existing prediction models are based on a single imaging modality, with limited information dimension, and performance has reached a bottleneck. Therefore, for the field, there is an urgent need to develop an objective, accurate and reliable non-invasive prediction technology to overcome the problems of the prior art. SUMMARY
[0004] To solve the problem of preoperative prediction of microvascular invasion, the application provides a liver cancer microvascular invasion probability prediction system based on imageomics and clinical features and machine learning.
[0005] The application adopts the following technical solutions: The liver cancer microvascular invasion probability prediction system based on imageomics and clinical features and machine learning includes the following modules: A data collection module for collecting patient clinical CT data and labeled lesion regions, the lesion regions being stored in the form of segmentation masks; the CT data is composed of a series of DICOM header files, and the DICOM header files contain patient information, sequence / scanning information, and geometric information; A data integrity check module for extracting DICOM (Digital Imaging and Communications in Medicine) header file information and performing data integrity checks; Image format conversion module: Converts CT data to NRRD format, uses the SimpleITK library to read DICOM sequences, preserves the spatial and intensity information of the original image (CT data and lesion area), and generates standardized medical image format files; Image preprocessing module: performs data resampling and standardization on standardized medical image format files; Radiomics feature extraction module: Extracts quantitative features from preprocessed images and segmentation masks using the PyRadiomics library; Dataset partitioning module: Performs cross-validation data partitioning on the fully labeled dataset (image + feature + label) into training and test sets; the dataset generated by the imageomics feature extraction module contains image features and labels; Feature selection module: performs feature selection on the training set; Model building and training module: Performs model building and training on the training set; Model performance evaluation module: Evaluates the model built and trained by the model building and training module by calculating evaluation metrics; Probability prediction module: Input the test set data into the evaluated model and output the predicted probability of microvascular invasion.
[0006] In the image preprocessing module: Resampling: Resamples the image to a uniform pixel spacing; Resizing: Adjusts the image to a uniform resolution; Intensity Normalization: Normalizes the image intensity to eliminate differences between devices.
[0007] In the radiomics feature extraction module, feature extraction parameters are set: gray-level histogram bin width, LoG filter parameters, resampling pixel spacing, etc., to extract first-order statistical features: energy, entropy, mean, standard deviation, skewness, kurtosis, etc.; shape features: volume, surface area, sphericity, compactness, etc.; texture features: gray-level co-occurrence matrix, gray-level run matrix, gray-level region size matrix, etc.; and filtering features: LoG filtering, wavelet transform, and other derived features. The patient's clinical data before CT is then combined.
[0008] In the dataset partitioning module, cross-validation data partitioning is performed, specifically including: five-fold cross-validation: the dataset is randomly divided into 5 subsets, including 4 training sets and 1 test set. The 4 training sets are used in turn for training, and the 1 test set is used for testing. Finally, the results of the 5 tests are combined.
[0009] In the feature selection module, features are selected from the training set using algorithms for feature filtering; specifically, this includes: removing highly correlated features (correlation coefficient > 0.8) and calculating Spearman correlation. The specific formula for the Spearman correlation analysis is as follows: in, The Spearman rank correlation coefficient. Let be the rank difference between two variables, and n be the sample size. This formula is used to calculate the nonlinear correlation between features. If the value is greater than 0.8, the relevant features are removed.
[0010] In model building and training, the model is built and trained on the allocated training set, using various machine learning algorithms to construct the prediction model; specifically including: Support Vector Machine (SVM): uses the RBF kernel function; Random Forest: integrates multiple decision trees for classification; Logistic Regression: a linear classification model; Gradient Boosting: an ensemble learning model based on Boosting. The specific formula for the logistic regression model is as follows: in, For the i-th eigenvalue, To predict probabilities, The intercept is... These are the feature weights. The model is trained by minimizing the logarithmic loss function: L represents the logarithmic loss, y represents the true label, and p represents the predicted probability.
[0011] The specific formula for the Support Vector Machine (SVM) model is as follows: Where s is the number of support vectors. For kernel functions, specifically Gaussian radial basis function, where, , It is the weight that controls the distance between samples ( The larger the size, the more complex the model. Here, b represents the support vector coefficients, and b is the bias term of the decision hyperplane. The label of the i-th training sample. The model maximizes the classification margin by solving the following optimization problem: w is the normal vector, and the constraints are: The specific formula for the random forest model is as follows: in, Let be the prediction for the t-th decision tree, where T is the number of trees. The model trains each tree through bootstrap sampling and random subsets of features, and then uses ensemble voting to obtain the final prediction.
[0012] The model performance evaluation module calculates various evaluation metrics, including: calculating AUC value, accuracy, sensitivity, and specificity; plotting ROC curves and confusion matrices; and evaluating the model's generalization ability through cross-validation.
[0013] The probability prediction module outputs the predicted probability of microvascular invasion; specifically, it includes: performing binary classification prediction based on the optimal threshold (Youden index), outputting continuous probability values (between 0 and 1), and providing a reliability assessment of the prediction results.
[0014] This invention proposes a preoperative probability prediction system for microvascular invasion (MVI) in liver cancer based on radiomics and machine learning. The system aims to quantify and predict MVI risk through automated analysis of clinical CT image data, thereby reducing reliance on invasive pathological examinations. This invention utilizes standardized preprocessing and quantitative image feature extraction techniques, combined with training and evaluation of multiple machine learning models, to achieve accurate prediction of MVI incidence probability from medical images. Attached Figure Description
[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the processing steps of a machine learning-based probability prediction system for microvascular invasion in liver cancer based on radiomics and clinical features, according to a preferred embodiment of the present invention. Figure 2 It's an ROC curve. Detailed Implementation
[0016] To make the embodiments, technical solutions, and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.
[0017] To address the challenge of preoperative prediction of microvascular invasion, this invention proposes a machine learning-based probabilistic prediction system for microvascular invasion in liver cancer, based on radiomics and clinical features. The system comprises the following modules: Data collection module: used to collect patients' clinical CT data and lesion areas marked by professional doctors. The lesion areas are stored in the form of segmented masks. The CT data includes multiple DICOM header files, which contain patient information, sequence and scan information, and geometric information. Data integrity check module: Extracts DICOM header file information and performs data integrity checks; Image format conversion module: Converts CT data to NRRD format, uses the SimpleITK library to read DICOM sequences, preserves the spatial and intensity information of the original images, and generates standardized medical image format files; Figure 1 In this context, image data preprocessing includes the process of image format conversion.
[0018] Image preprocessing module: performs data resampling and standardization, including: resampling: resampling the image to a uniform pixel spacing, resizing: adjusting the image to a uniform resolution, and intensity standardization: standardizing the image intensity to eliminate differences between devices; The radiomics feature extraction module uses the PyRadiomics library to extract quantitative features from preprocessed images and segmentation masks (which can be annotated by doctors). Specifically, this includes: setting feature extraction parameters such as gray-level histogram binning width, LoG filter parameters, and resampling pixel spacing; extracting first-order statistical features such as energy, entropy, mean, standard deviation, skewness, and kurtosis; extracting shape features such as volume, surface area, sphericity, and compactness; extracting texture features such as gray-level co-occurrence matrix, gray-level run-length matrix, and gray-level region size matrix; and extracting filtering features such as LoG filtering and wavelet transform. Finally, the module combines clinical data. Dataset partitioning module: performs cross-validation data partitioning on the dataset; the dataset generated by the radiomics feature extraction module contains image features and doctor-annotated labels; specifically including: five-fold cross-validation: randomly divides the dataset into 5 subsets, including 4 training sets and 1 test set, and uses 4 of the training sets for training and 1 test set for testing in turn, and finally combines the results of 5 tests; Feature selection module: Selects features from the training set using algorithms; specifically, it includes: removing highly correlated features (correlation coefficient > 0.8) and calculating Spearman correlation.
[0019] The specific formula for the Spearman correlation analysis is as follows: in, The Spearman rank correlation coefficient. Let be the rank difference between two variables, and n be the sample size. This formula is used to calculate the nonlinear correlation between features. If the value is greater than 0.8, the relevant features are removed.
[0020] Model building and training module: This module builds and trains models on the allocated training set, using various machine learning algorithms to construct predictive models; specifically including: Support Vector Machine (SVM): uses the RBF kernel function; Random Forest: integrates multiple decision trees for classification; Logistic Regression: a linear classification model; Gradient Boosting: an ensemble learning model based on Boosting. The specific formula for the logistic regression model is as follows: in, To predict probabilities, The intercept is... These are the feature weights. The model is trained by minimizing the logarithmic loss function: .
[0021] The specific formula for the Support Vector Machine (SVM) model is as follows: in, Here, the kernel function is specifically the Gaussian radial basis function, where... , It is the weight that controls the distance between samples ( The larger the size, the more complex the model. Here, b represents the support vector coefficients, and b represents the bias term of the decision hyperplane. Let be the label of the i-th training sample. The model maximizes the classification margin by solving the following optimization problem: The constraints are: .
[0022] The specific formula for the random forest model is as follows: in, Let be the prediction for the t-th decision tree, where T is the number of trees. The model trains each tree through bootstrap sampling and a random subset of features (which is dynamically generated when splitting at each node of each decision tree), and then integrates votes to obtain the final prediction.
[0023] Model performance evaluation module: Calculates various evaluation metrics, including: calculating AUC, accuracy, sensitivity, and specificity; plotting ROC curves and confusion matrices; and evaluating the model's generalization ability through cross-validation. Probability prediction module: outputs the predicted probability of microvascular invasion; specifically, it includes: performing binary classification prediction based on the optimal threshold (Youden index), outputting continuous probability values (between 0 and 1), and providing a reliability assessment of the prediction results.
[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0025] See attached document Figure 1, Figure 1 This is a flowchart of a specific embodiment of the present invention; to solve the problem of preoperative prediction of microvascular invasion, this embodiment proposes a machine learning-based probability prediction system for microvascular invasion in liver cancer based on radiomics and clinical features, including the following modules: Data collection module: used to collect patients' clinical CT data and labeled lesion areas, which are stored in the form of segmented masks; Data integrity check module: Extracts DICOM header file information and performs data integrity checks; Image format conversion module: Converts DICOM sequences to NRRD format, uses the SimpleITK library to read DICOM sequences, preserves the spatial and intensity information of the original image, and generates standardized medical image format files; Image preprocessing module: performs data resampling and standardization, including: resampling: resampling the image to a uniform pixel spacing, resizing: adjusting the image to a uniform resolution, and intensity standardization: standardizing the image intensity to eliminate differences between devices; Radiomics Feature Extraction Module: This module uses the PyRadiomics library to extract quantitative features. Specifically, it includes: setting feature extraction parameters such as gray-level histogram bin width, LoG filter parameters, and resampling pixel spacing; extracting first-order statistical features such as energy, entropy, mean, standard deviation, skewness, and kurtosis; extracting shape features such as volume, surface area, sphericity, and compactness; extracting texture features such as gray-level co-occurrence matrix, gray-level run matrix, and gray-level region size matrix; and extracting filtering features such as LoG filtering and wavelet transform. Finally, it combines clinical data. Dataset partitioning module: performs cross-validation data partitioning; specifically includes: five-fold cross-validation: randomly divides the dataset into 5 subsets, uses 4 subsets for training and 1 subset for testing in turn, and finally combines the results of the 5 tests; Feature selection module: performs feature selection on the training set, using algorithms to filter features; specifically including: removing highly correlated features (correlation coefficient > 0.8), and using Spearman correlation calculation; The specific formula for the Spearman correlation analysis is as follows: in, The Spearman rank correlation coefficient. Let be the rank difference between two variables, and n be the sample size. This formula is used to calculate the nonlinear correlation between features. If the value is greater than 0.8, the relevant features are removed.
[0026] Model building and training module: This module builds and trains models on the allocated training set, using various machine learning algorithms to construct predictive models; specifically including: Support Vector Machine (SVM): uses the RBF kernel function; Random Forest: integrates multiple decision trees for classification; Logistic Regression: a linear classification model; Gradient Boosting: an ensemble learning model based on Boosting. The specific formula for the logistic regression model is as follows: in, To predict probabilities, The intercept is... These are the feature weights. The model is trained by minimizing the logarithmic loss function: .
[0027] The specific formula for the Support Vector Machine (SVM) model is as follows: in, Here, the kernel function is specifically the Gaussian radial basis function, where... , It is the weight that controls the distance between samples ( The larger the size, the more complex the model. Here, b represents the support vector coefficients, and b is the bias term of the decision hyperplane. The label of the i-th training sample. The model maximizes the classification margin by solving the following optimization problem: The constraints are: .
[0028] The specific formula for the random forest model is as follows: in, Let be the prediction for the t-th decision tree, where T is the number of trees. The model trains each tree through bootstrap sampling and random subsets of features, and then uses ensemble voting to obtain the final prediction.
[0029] Model performance evaluation module: Calculates various evaluation metrics, including: calculating AUC, accuracy, sensitivity, and specificity; plotting ROC curves and confusion matrices; and evaluating the model's generalization ability through cross-validation. The formula for calculating accuracy and its meaning are as follows: Wherein, TP represents the number of true positive samples (correct prediction for the positive class), FP represents the number of false positive samples (incorrect prediction for the negative class), TN represents the number of true negative samples (correct prediction for the negative class), and FN represents the number of false negative samples (incorrect prediction for the positive class).
[0030] AUC (Area Under ROC Curve) calculation formula and meaning: in: TPR (True Positive Rate / Recall Rate) is defined as: FPR (False Positive Rate) is defined as: Wherein, TP represents the number of true positive samples (correct prediction for the positive class), FP represents the number of false positive samples (incorrect prediction for the negative class), TN represents the number of true negative samples (correct prediction for the negative class), and FN represents the number of false negative samples (incorrect prediction for the positive class).
[0031] Probability prediction module: outputs the predicted probability of microvascular invasion; specifically, it includes: binary classification prediction based on the optimal threshold (Youden index), outputting continuous probability values (between 0 and 1), and providing a confidence assessment of the prediction results (AUC and accuracy).
[0032] The optimal decision threshold for predicting microvascular invasion was determined by maximizing the Youden index. in, is the candidate threshold, and TP, TN, FP, and FN are the number of true positive, true negative, false positive, and false negative samples, respectively.
[0033] This invention compares multiple algorithms to accurately verify the generalization ability of the model, successfully selects the optimal model, effectively eliminates models with high overfitting risk, and ensures that the prediction engine has both high accuracy and high reliability, providing solid support for practical applications.
[0034] Figure 2 This is an ROC curve, used to evaluate the performance of the constructed SVC (Support Vector Machine) model in predicting the probability of microvascular invasion (MVI) in liver cancer.
[0035] Figure 2 The ROC curve clearly demonstrates the superior performance and excellent generalization ability of the model constructed in this invention. The AUC on the training set reaches 0.938, indicating that the model has excellent discrimination and fitting effect on known data. The model also achieves an AUC of 0.708 on the independent test set, proving that the model has effectively applied the learned patterns to unknown data, and its prediction results are stable, reliable, and possess clinical auxiliary diagnostic value.
[0036] In summary, this invention relates to a machine learning-based preoperative probability prediction system for microvascular invasion (MVI) in liver cancer, based on radiomics and clinical features. The aim is to quantify and predict MVI risk through automated analysis of clinical CT image data, thereby reducing reliance on invasive pathological examinations. First, patient clinical CT data and physician-annotated lesion segmentation masks are collected, and data standardization is achieved through DICOM to NRRD format conversion. Second, image resampling, size adjustment, and intensity standardization preprocessing are performed. Radiomics features, including shape, texture, and filtering features, are extracted using PyRadiomics. The dataset is partitioned using five-fold cross-validation, and feature selection is performed using Spearman correlation analysis. A prediction model is constructed using support vector machines, random forests, logistic regression, and gradient boosting algorithms. Finally, the predicted MVI probability value is output, and model performance is evaluated using metrics such as AUC and sensitivity. This invention achieves non-invasive and accurate preoperative prediction, providing quantitative support for clinical decision-making.
[0037] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.
Claims
1. A machine learning system for predicting probability of microvascular invasion of hepatocellular carcinoma based on radiomics and clinical features, characterized in that, The application comprises the following modules: A data collection module is configured to collect patient clinical CT data and labeled lesion regions; the CT data comprises a plurality of DICOM header files, and the DICOM header files contain patient information, sequence and scanning information, and geometric information; A data integrity checking module is configured to extract DICOM header file information and perform data integrity checking; An image format conversion module is configured to convert the CT data into an NRRD format, read the DICOM sequence by using a SimpleITK library, maintain the spatial information and intensity information of the original image, and generate a standardized medical image format file; An image preprocessing module is configured to perform data resampling and intensity standardization on the standardized medical image format file; An imageomics feature extraction module is configured to extract quantitative features from the preprocessed image and a segmentation mask by using a PyRadiomics library; A data set division module is configured to divide the complete labeled data set into cross-validation data, and divide the cross-validation data into a training set and a test set; A feature selection module is configured to select features from the training set; A model construction and training module is configured to construct and train a model on the training set; A model performance evaluation module is configured to evaluate the model constructed and trained by the model construction and training module by calculating evaluation indexes; A probability prediction module is configured to input the test set data into the evaluated model, and output a prediction probability of microvessel invasion. 2.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, In the image preprocessing module, resampling is specifically as follows: the image is resampled to a uniform pixel pitch and a uniform resolution; and intensity standardization is specifically as follows: the image intensity is standardized. 3.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, In the imageomics feature extraction module, the feature extraction parameters are set as follows: a gray level histogram bin width, LoG filter parameters, and resampling pixel pitch; first-order statistical features are extracted, including energy, entropy, mean, standard deviation, skewness, and kurtosis; shape features are extracted, including volume, surface area, sphericity, and compactness; texture features are extracted, including a gray level co-occurrence matrix, a gray level run-length matrix, and a gray level zone size matrix; filter features are extracted, including LoG filtering and wavelet transform features; and the clinical data are combined. 4.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, In the data set division module, five-fold cross-validation is performed: the data set is randomly divided into five subsets, including four training sets and one test set; four training sets are used for training in turn, and one test set is used for testing; and finally, the test results of the five times are integrated. 5.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, In the feature selection module, Spearman correlation is calculated. 6.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 5, wherein, The specific formula of the Spearman correlation calculation is as follows: wherein, In the model construction and training module, the constructed model is specifically as follows: is the Spearman rank correlation coefficient, d i is the rank difference of two variables, n is the sample size; if A support vector machine SVM uses an RBF kernel function; |> 0.8, the relevant features are removed. 7.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, A random forest integrates a plurality of decision trees for classification; A logistic regression is a linear classification model; A gradient boosting is an ensemble learning based on Boosting; The specific formula of the logistic regression model is as follows: The specific formula of the support vector machine SVM model is as follows: The specific formula of the random forest model is as follows: where x i is the ith feature value, p(y = 1 | x) is the predicted probability, β0is the intercept, β i is the feature weight, i = 1, 2, …, n; the model is trained on the training set by minimizing the log loss function: ; L is the log loss, y is the true label, and p is the predicted probability. In the model performance evaluation module, the AUC value, accuracy, sensitivity, and specificity are calculated, the ROC curve and the confusion matrix are drawn, and the model generalization ability is evaluated by cross-validation. where s is the number of support vectors, K(x i ,x) is a kernel function, where , the weight that controls the distance between samples, is a support vector coefficient, b is a bias term of the hyperplane, y i the label of the i-th training set sample; the model maximizes the margin between classes by solving the following optimization problem: w is a normal vector, with the constraint that where h t (x) is the prediction of the tth decision tree, T is the number of trees; this model trains each tree by Bootstrap sampling and feature random subset, and the final prediction is obtained by ensemble voting. 8.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of imaging genomics and clinical characteristics according to claim 1, wherein, 9.The system for predicting probability of microvascular invasion of liver cancer based on machine learning of radiomics and clinical features according to any one of claims 1-8, wherein, In the probability prediction module, binary classification prediction is performed based on the optimal threshold value, and a continuous probability value is output to provide a confidence evaluation of the prediction result.