Construction method and system of MRI image-based machine learning model for preoperative auxiliary diagnosis of fatty hepatitis type hepatocellular carcinoma

Abstract

The application belongs to the technical field of biological medicine, and particularly relates to a construction method and system of an MRI image-based machine learning model for preoperative auxiliary diagnosis of fatty hepatitis type hepatocellular carcinoma. The application extracts 2,264 high-dimensional space features, and through strict ICC-LASSO-Logisti single factor regression cascade screening and filtering, 37 golden features highly related to SH-HCC are locked. The constructed diagnostic model (such as SGD, Logistic, BaggingDecisionTree, etc.) exhibits excellent performance, and the AUC of the SGD model test set is as high as 0.898. The calibration curve and the decision curve further prove that the application can bring high credibility net benefits to patients in actual clinical application.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, specifically relating to a method and system for constructing an MRI radiomics machine learning model for preoperative auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type. Background Technology

[0002] Accurate identification of steatohepatitis-associated hepatocellular carcinoma (SH-HCC) is crucial for clinical treatment decisions. Although SH-HCC is pathologically rich in fat droplets and inflammatory changes, its steatosis is often confused with the focal steatosis and heterogeneous necrosis found in conventional hepatocellular carcinoma in traditional gross imaging examinations (such as routine CT and conventional MRI), making it difficult to form a quantitative and standardized preoperative accurate differential diagnosis.

[0003] In recent years, the rise of radiomics technology has made it possible to extract thousands of invisible spatial heterogeneity features from traditional medical images in high throughput. Considering the highly complex cellular arrangement and microvascular heterogeneity within SH-HCC lesions, a highly accurate preoperative non-invasive classifier has been developed by combining non-invasive MRI bimodal sequences (in-phase / out-of-phase, enabling precise quantification of water-lipid separation and intracellular micro-fat) with deep radiomics analysis. This classifier has extremely high clinical application and promotion value. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments.

[0005] As one aspect of the present invention, the present invention provides a method for constructing an MRI radiomics machine learning model for preoperative auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type, which includes the following steps: S1: Acquire DICOM images of preoperative non-enhanced T1-weighted in-phase and out-of-phase sequences of magnetic resonance imaging in patients with hepatocellular carcinoma of known pathological outcome; S2: Delineate the region of interest (ROI) of the tumor lesion layer by layer on the DICOM image, and extract the original radiomics features covering first-order statistical features, shape features and high-level texture features based on the ROI. S3: Perform cascaded feature filtering on the original image omics features, specifically including: S3-1) Calculate the intragroup correlation coefficient (ICC) for each feature and remove features with ICC ≤ 0.75; S3-2) For the remaining features, LASSO regression is used to locate the lambda.min parameter with the minimum binomial bias under 10-fold cross-validation, and trajectory compression is performed on the feature operators with non-zero coefficients. S3-3) Univariate logistic regression was used to test clinical significance, and the core difference imaging features with extremely significant differences between groups were finally selected. S4: After Z-score standardization of the core differential image features, the data are input into a pre-set machine learning algorithm for training to construct a machine learning classifier for preoperative auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type.

[0006] As a preferred embodiment of the construction method described in this invention: in step S4, the machine learning classifier is selected from any one or more combinations of stochastic gradient descent (SGD), logistic regression (Logistic), quadratic discriminant analysis (QDA), and bagging decision tree (Bagging Decision Tree).

[0007] As a preferred embodiment of the construction method described in this invention, the specific training parameters of the machine learning classifier in step S4 are set as follows: SGD model: Data preprocessing uses Z-score, class weights are None, learning rate is invscaling, loss function is modified_huber, and penalty term is L1; Logistic regression model: Data preprocessing uses Z-score, regularization parameter C is 0.01, class weight is None, penalty term is l2, and tolerance Tol is 1e-05; The second-order discriminant analysis model: data preprocessing uses Z-score, regularization parameter is 1.0, and tolerance Tol is 1e-05; Bagged decision tree model: Data preprocessing uses Z-score, with a maximum depth of 10, a maximum feature ratio of 0.2, a maximum sample ratio of 0.2, and 100 base learners.

[0008] As a preferred embodiment of the construction method described in this invention: in step S3-2, LASSO regression is performed based on the glmnet package in R language, and the optimal lambda.min parameter corresponding to the minimum binomial bias is found under 10-fold cross-validation in order to compress the trajectory of the feature operator.

[0009] As a preferred embodiment of the construction method described in this invention: in step S4, the sample set is divided into a training set and a test set using five-fold cross-validation, and the performance of the machine learning classifier is evaluated in multiple dimensions on the training set and the test set using the receiver operating characteristic curve, calibration curve and clinical decision curve, respectively.

[0010] As a preferred embodiment of the construction method described in this invention: the high-level texture features extracted in step S2 include any one or more of the following: gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size region matrix (GLSZM), gray-level dependency matrix (GLDM), and neighborhood gray-level difference matrix (NGTDM).

[0011] As a preferred embodiment of the construction method described in this invention: in step S2, the extracted original image omics features are automatically obtained based on standard computational geometry formulas.

[0012] As a preferred embodiment of the construction method described in this invention: a total of 2,264 original image omics features are extracted in step S2; and a total of 37 core differential image features are selected in steps S3-3.

[0013] As a preferred embodiment of the construction method described in this invention: the core differential image features in step S3 are selected from all or part of the following feature list: boxsimagimage_glszm_SmallAreaEmphasis_OP, specklenoise_gldm_LargeDependenceLowGrayLevelEmphasis_OP, log_ngtdm_log-sigma-4-0-mm-3D-Coarseness_IP, wavelet_gldm_wavelet-HLL-SmallDependenceHighGrayLevelEmphasis_OP, wavelet_firstorder_wavelet-HHH-Kurtosis_OP, wavelet_glcm_wavelet-LHH-Correlation_IP, wavelet_firstorder_wavelet-LHL-Mean_IP, wavelet_firstorder_wavelet-HHL-Skewness_OP, log_firstorder_log-sigma-4-0-mm-3D-Skewness_OP, log_fir storder_log-sigma-2-0-mm-3D-Skewness_OP, wavelet_glrlm_wavelet-LLL-ShortRunHighGrayLevelEmphasis_OP, wavelet_glszm_wavelet -HHL-GrayLevelNonUniformity_OP, wavelet_gldm_wavelet-LLL-SmallDependenceHighGrayLevelEmphasis_OP, specklenoise_gldm_LargeDe pendenceLowGrayLevelEmphasis_OP, wavelet_glszm_wavelet-LHH-SmallAreaLowGrayLevelEmphasis_OP, wavelet_glszm_wavelet-HHH-Sma llAreaLowGrayLevelEmphasis_OP, log_glcm_log-sigma-4-0-mm-3D-DifferenceVariance_OP, wavelet_glcm_wavelet-HHL-JointEnergy_OP,wavelet_glcm_wavelet-HHL-JointEntropy_OP, wavelet_glszm_wavelet-HLL-SmallAreaEmphasis_OP, normalize_glrlm_HighGrayLevelRunEmphasis_OP, wavelet_glcm_wavelet-HLH-Imc2_OP, normalize_firstorder_Range_OP, l og_glszm_log-sigma-0-5-mm-3D-SizeZoneNonUniformityNormalized_OP, normalize_gldm_HighGrayLevelEmphasis_IP, log_glszm_log-sigma-0-5-mm-3D-SizeZoneNonUniformityNormalized_IP, wavelet_glcm_wavelet-LLH-Im c1_OP,curvatureflow_glrlm_LowGrayLevelRunEmphasis_OP,wavelet_glcm_wavelet-LLH-Correlation_OP,normalize_firstorder_RootMeanSquared_OP,wavelet_glcm_wavelet-LLH-Imc1_OP,wavelet_glcm_wavelet-HHL-Imc2_ OP, wavelet_glcm_wavelet-HHH-Imc1_IP, normalize_glrlm_ShortRunLowGrayLevelEmphasis_OP, normalize_firstorder_MeanAbsoluteDeviation_OP, specklenoise_glcm_Correlation_OP, wavelet_glcm_wavelet-HHH-Imc2_OP. ,

[0014] The present invention also provides an MRI radiomics-assisted diagnostic system for performing the construction method of claim 1, comprising: The image delineation and feature extraction module is used to receive non-enhanced T1-weighted in-phase and out-of-phase sequence images of preoperative MRI and extract raw radiomics features based on the region of interest (ROI) of the lesion. The cascaded feature filtering module is configured to select core differential image features by filtering through intragroup correlation coefficient (ICC), LASSO regression dimensionality reduction, and univariate logistic regression. The machine learning classification module is equipped with a pre-trained machine learning classifier. It takes the core differential image features as input and outputs the probability of the patient having hepatocellular carcinoma of the steatohepatitis type as an auxiliary diagnostic tool before surgery.

[0015] As a preferred embodiment of the auxiliary diagnostic system described in this invention, the preoperative auxiliary diagnostic probability output by the machine learning classification module shows a net benefit significantly higher than the "full intervention" baseline and the "zero intervention" baseline in the clinical decision curve DCA analysis test.

[0016] The beneficial effects of this invention are as follows: By extracting 2,264 high-dimensional spatial features and rigorously filtering them using an ICC-LASSO-Logisti univariate regression cascade, this invention identified 37 golden features highly correlated with SH-HCC. The constructed diagnostic models (such as SGD, Logistic, Bagging Decision Tree, etc.) exhibit excellent performance, with the SGD model achieving an AUC of 0.898 on the test set. Calibration and decision curves further confirm that this invention can provide patients with a highly reliable net benefit in actual clinical applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, wherein: Figure 1 This is a schematic diagram of the overall process of preoperative non-invasive MRI in-phase / out-of-phase sequence lesion region of interest (ROI) delineation (A) and radiomics feature extraction (B) of the present invention.

[0018] Figure 2 This is a graph showing parameter optimization and image feature coefficient trajectory compression for the LASSO regression model under 10-fold cross-validation. The left side shows the binomial bias curve as a function of Log(lambda), and the right side shows the trajectory compression curve of the feature coefficients as a function of Log(lambda).

[0019] Figure 3 A forest plot of univariate logistic regression analysis for 37 core differential image features that were finally integrated after multi-level screening.

[0020] Figure 4 This is a multidimensional evaluation confirmation plot of the four radiomics machine learning models constructed on the training and test sets, showing the recipient operating characteristic curve (ROC, Figure D), calibration curve (Figure E), and clinical decision benefit curve (DCA, Figure F). Detailed Implementation

[0021] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to specific examples.

[0022] Example 1: Image Acquisition and Interactive Delineation of Region of Interest (ROI) for Lesions This embodiment employs a completely non-invasive preoperative medical imaging approach for data acquisition. Two key sequences—non-enhanced T1-weighted in-phase (IP) and out-of-phase (OP) sequences—obtained preoperatively from 193 patients with HCC under magnetic resonance imaging (MRI) were exported in standard DICOM format and uniformly imported into the Shanghai United Imaging UAI Advanced Research Platform (USIA). Figure 1 A). Among them, there were 94 patients with SH-HCC (hepatocellular carcinoma with steatohepatitis) and 99 patients with cHCC (common hepatocellular carcinoma). A double-blind method was used to precisely delineate the region of interest (ROI / VOI) layer by layer along the tumor edge in three-dimensional voxel space or two-dimensional full-level plane.

[0023] Example 2: High-dimensional radiomics feature extraction and robust cascade screening From the ROI matrix drawn in Example 1, the platform automatically extracted a total of 2,264 original image omics features based on standard computational geometry formulas, including first-order statistics, 2D / 3D shape features, and high-level texture features (including GLCM, GLRLM, GLSZM, GLDM, NGTDM, etc.). Figure 1 B).

[0024] To eliminate human drawing errors and imaging artifacts, this invention employs a multi-tiered, rigorous screening process, the specific steps of which are as follows: 1. Intragroup Correlation Coefficient (ICC) Consistency Filtering: Remove features that are extremely insensitive to the operator or unstable, and retain only stable image variables with ICC > 0.75.

[0025] 2. LASSO Regression Feature Dimensionality Reduction: Based on the glmnet (v4.1.10) package in R v4.2.3, the optimal lambda.min parameter is found when the binomial deviation is minimized under 10-fold cross-validation. Figure 2 Trajectory compression is applied to feature operators with non-zero coefficients, effectively avoiding overfitting caused by multicollinearity.

[0026] 3. Univariate Logistic Regression Screening: Combining clinical significance testing (stats package), 37 key imaging features with highly significant differences between groups were ultimately filtered and identified. The odds ratio (95% CI) and p-value for each feature are attached. Figure 3 The forest plots shown (e.g., wavelet_glcm_wavelet-HHH-Imc2_OP exhibited extremely strong statistical significance, p = 0.00008).

[0027] Example 3: Standardized Machine Learning Modeling and Clinical Efficacy Assessment of Radiomics The 37 key features were standardized using Z-scores and input into the machine learning process. Five-fold cross-validation was used to divide the data into training and test sets (n=154 for the training set and n=39 for the test set). In this stage, this invention constructed and evaluated four classification models: Stochastic Gradient Descent (SGD), Logistic Regression, Quadratic Discriminant Analysis (QDA), and Bagging Decision Tree. Figure 4 ).

[0028] 1. Assessment of diagnostic sensitivity and specificity (ROC curve analysis) In the training set, the Bagging Decision Tree classifier showed the best fit, with a mean area under the curve (Mean AUC) as high as 0.970. Figure 4 D); Both the SGD and Logistic models achieved an excellent AUC score of 0.922; the QDA model's AUC was 0.892. In the test set cross-validation, SGD showed the best prediction generalization ability, with a Mean AUC still as high as 0.898 ( Figure 4 D); the Logistic model followed closely behind, with a Mean AUC of 0.896; QDA (0.876) and BaggingDecisionTree (0.870) both showed excellent preoperative non-invasive diagnostic discriminative power.

[0029] 2. Diagnostic probability reliability assessment (Calibration Curve Analysis) By comparing the consistency between the actual occurrence rate and the model predicted probability ( Figure 4E): During both the training and testing phases, the Brier scores of each model remained within a low ideal threshold range. In the testing set, the Brier score for SGD was as low as 0.140, and QDA was 0.150, indicating that the SH-HCC auxiliary diagnostic probability output by the imaging model of this invention is not a blind guess, but has extremely high pathological consistency and numerical reliability.

[0030] 3. Clinical Real Net Benefit Assessment (Decision Curve DCA Analysis) Clinical decision curve analysis is introduced to evaluate the net benefit of the model under different threshold probabilities. Figure 4 F): The results show that, within a very wide range of clinical selection threshold probabilities (both in the training and test sets), the color benefit curves corresponding to the four models are significantly higher than the two gray-black baselines of "Treat all" and "Treat none". This fully confirms that the preoperative non-invasive MRI radiomics model established in this invention can bring significant non-invasive net benefits to liver cancer patients when actually applied to clinical screening.

[0031] The specific parameter settings for each model are as follows: SGD: Data Preprocessing - Zscore, Class Weight: None, Learning Rate: invscaling, Loss: modified_huber, Penalty: l1; Logistic Model: Data Preprocessing - Z-score, C: 0.01, Class Weight: None, Penalty: l2, Tol: 1e-05; QDA Classifier: Data Preprocessing - Zscore, Reg Param: 1.0, Tol: 1e-05; Bagging Decision Tree: Data Preprocessing - Zscore, Max Depth: 10, Max Features: 0.2, Max Samples: 0.2, N Estimators: 100.

[0032] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing an MRI radiomics machine learning model for preoperative auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type, characterized in that: Includes the following steps: S1: Acquire DICOM images of preoperative non-enhanced T1-weighted in-phase and out-of-phase sequences of magnetic resonance imaging in patients with hepatocellular carcinoma of known pathological outcome; S2: Delineate the region of interest (ROI) of the tumor lesion layer by layer on the DICOM image, and extract the original radiomics features covering first-order statistical features, shape features and high-level texture features based on the ROI. S3: Perform cascaded feature filtering on the original image omics features, specifically including: S3-1) Calculate the intragroup correlation coefficient (ICC) for each feature and remove features with ICC ≤ 0.75; S3-2) For the remaining features, LASSO regression is used to locate the lambda.min parameter with the minimum binomial bias under 10-fold cross-validation, and trajectory compression is performed on the feature operators with non-zero coefficients. S3-3) Univariate logistic regression was used to test clinical significance, and the core difference imaging features with extremely significant differences between groups were finally selected. S4: After Z-score standardization of the core differential image features, the data are input into a pre-set machine learning algorithm for training to construct a machine learning classifier for preoperative auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type.

2. The construction method according to claim 1, characterized in that: In step S4, the machine learning classifier is selected from any one or more combinations of stochastic gradient descent (SGD), logistic regression, quadratic discriminant analysis (QDA), and bagging decision tree.

3. The construction method according to claim 2, characterized in that: In step S4, the specific training parameters for the machine learning classifier are set as follows: SGD model: Data preprocessing uses Z-score, class weights are None, learning rate is invscaling, loss function is modified_huber, and penalty term is L1; Logistic regression model: Data preprocessing uses Z-score, regularization parameter C is 0.01, class weight is None, penalty term is l2, and tolerance Tol is 1e-05; The second-order discriminant analysis model: data preprocessing uses Z-score, regularization parameter is 1.0, and tolerance Tol is 1e-05; Bagged decision tree model: Data preprocessing uses Z-score, with a maximum depth of 10, a maximum feature ratio of 0.2, a maximum sample ratio of 0.2, and 100 base learners.

4. The construction method according to any one of claims 1-3, characterized in that: In step S3-2, LASSO regression is performed using the glmnet package in R language, and the optimal lambda.min parameter corresponding to the minimum binomial bias is found under 10-fold cross-validation in order to compress the trajectory of the feature operator.

5. The construction method according to any one of claims 1-3, characterized in that: In step S4, the sample set is divided into a training set and a test set using five-fold cross-validation, and the performance of the machine learning classifier is evaluated in multiple dimensions on the training set and the test set using receiver operating characteristic curves, calibration curves, and clinical decision curves, respectively.

6. The auxiliary diagnostic system according to claim 5, characterized in that: The high-level texture features extracted in step S2 include any one or more of the following: gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size region matrix (GLSZM), gray-level dependency matrix (GLDM), and neighborhood gray-level difference matrix (NGTDM).

7. The construction method according to any one of claims 1-3, characterized in that: In step S2, the extracted original image omics features are automatically obtained based on standard computational geometry formulas.

8. The construction method according to any one of claims 1-3, characterized in that: A total of 2,264 original image omics features were extracted in step S2; a total of 37 core differential image features were selected in steps S3-3.

9. The construction method according to claim 8, characterized in that: The core difference image features in step S3 are selected from all or a partial combination of the following feature list: boxsimagimage_glszm_SmallAreaEmphasis_OP, specklenoise_gldm_LargeDependenceLowGrayLevelEmphasis_OP, log_ngtdm_log-sigma-4-0-mm-3D-Coarseness_IP, wavelet_gldm_wavelet-HLL-SmallDependenceHighGrayLevelEmphasis_OP, wavelet_firstorder_wavelet-HHH-Kurtosis_OP, wavelet_glcm_wavelet-LHH-Correlation_IP, wavelet_firstorder_wavelet-LHL-Mean_IP, wavelet_firstorder_wavelet-HHL-Skewness_OP, log_firstorder_log-sigma-4-0-mm-3D-Skewness_OP, log_firstorder_log-sigma-2-0-mm-3D-Skewness_OP, wavelet_glrlm_wavelet-LLL-ShortRunHighGrayLevelEmphasis_OP, wavelet_glszm_wavelet-HHL-GrayLevelNonUniformity_OP, wavelet_gldm_wavelet-LLL-SmallDependenceHighGrayLevelEmphasis_OP, specklenoise_gldm_LargeDependenceLowGrayLevelEmphasis_OP, wavelet_glszm_wavelet-LHH-SmallAreaLowGrayLevelEmphasis_OP, wavelet_glszm_wavelet-HHH-SmallAreaLowGrayLevelEmphasis_OP, log_glcm_log-sigma-4-0-mm-3D-DifferenceVariance_OP, wavelet_glcm_wavelet-HHL-JointEnergy_OP, wavelet_glcm_wavelet-HHL-JointEntropy_OP,wavelet_glszm_wavelet-HLL-SmallAreaEmphasis_OP、normalize_glrlm_HighGrayLevelRunEmphasis_OP、wavelet_glcm_wavelet-HLH-Imc2_OP、normalize_firstorder_Range_OP、log_glszm_log-sigma-0-5-mm-3D-SizeZoneNonUniformityNormalized_OP、normalize_gldm_HighGrayLevelEmphasis_IP、log_glszm_log-sigma-0-5-mm-3D-SizeZoneNonUniformityNormalized_IP、wavelet_glcm_wavelet-LLH-Imc1_OP、curvatureflow_glrlm_LowGrayLevelRunEmphasis_OP、wavelet_glcm_wavelet-LLH-Correlation_OP、normalize_firstorder_RootMeanSquared_OP、wavelet_glcm_wavelet-LLH-Imc1_OP、wavelet_glcm_wavelet-HHL-Imc2_OP、wavelet_glcm_wavelet-HHH-Imc1_IP、normalize_glrlm_ShortRunLowGrayLevelEmphasis_OP、normalize_firstorder_MeanAbsoluteDeviation_OP、specklenoise_glcm_Correlation_OP、wavelet_glcm_wavelet-HHH-Imc2_OP。、 10. An MRI radiomics-assisted diagnostic system for performing the construction method of claim 1, characterized in that, include: The image delineation and feature extraction module is used to receive non-enhanced T1-weighted in-phase and out-of-phase sequence images of preoperative MRI and extract raw radiomics features based on the region of interest (ROI) of the lesion. The cascaded feature filtering module is configured to select core differential image features by filtering through intragroup correlation coefficient (ICC), LASSO regression dimensionality reduction, and univariate logistic regression. The machine learning classification module is equipped with a pre-trained machine learning classifier. It takes the core differential image features as input and outputs the probability of the patient having hepatocellular carcinoma of the steatohepatitis type as an auxiliary diagnostic tool before surgery.

Images