Enhanced MRI radiomics model evaluation method for predicting EGF

Abstract

The invention discloses an enhanced MRI (Magnetic Resonance Imaging) imageomics model evaluation method for predicting EGF (Epidermal Growth Factor), and relates to the technical field of medical image processing and tumor diagnos.The method comprises the following steps: acquiring an enhanced MRI image of a high-grade glioma patient, and performing image preprocessing and tumor area sketching to obtain an enhanced MRI image of the high-grade glioma patient; first-order statistical features, shape features, texture features and wavelet transform features are extracted, a maximum correlation minimum redundancy algorithm and a recursive feature elimination algorithm are adopted for feature screening, key radiomics features are obtained, then a gradient enhancement algorithm is adopted for constructing an EGF expression level prediction model, a radiomics score is calculated, and the EGF expression level is predicted. The non-invasive evaluation of the expression of the high-grade glioma EGF is realized, the constructed radiomics score is obviously related to the prognosis of a patient, the clinical formulation of an individualized treatment scheme can be assisted, and the method has an important clinical application value.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and tumor diagnosis technology, specifically to an evaluation method for enhanced MRI radiomics models used to predict EGF. Background Technology

[0002] Gliomas are the most common primary malignant tumors of the central nervous system. High-grade gliomas (HGGs), including World Health Organization (WHO) grade III and IV gliomas, are characterized by highly invasive growth, high recurrence rate, and poor prognosis. Although current comprehensive treatment methods include surgical resection, radiotherapy, chemotherapy, and targeted therapy, the 5-year survival rate of patients with high-grade gliomas remains low, and the disability and mortality rates remain high. Epidermal growth factor (EGF) and its receptor (EGFR) play important roles in tumorigenesis and development. EGFR is a transmembrane receptor tyrosine kinase that is overexpressed, amplified, and mutated in various malignant tumors. Studies have shown that EGFR is one of the most common gene alterations in glioblastoma. Abnormal activation of its signaling pathway can promote tumor cell proliferation, invasion, and angiogenesis, and is closely related to poor patient prognosis. As a ligand of EGFR, the expression level of EGF is also related to the activation state of the EGFR signaling pathway, playing an important regulatory role in the tumor microenvironment. Accurate assessment of EGF expression levels in high-grade gliomas is crucial for prognosis, treatment selection, and efficacy monitoring. Traditional methods for EGF expression detection primarily rely on pathological examination following tissue biopsy, including immunohistochemical staining, RNA sequencing, and fluorescence in situ hybridization. However, these methods are all invasive and have the following limitations: (1) Tumor tissue needs to be obtained through surgery or puncture biopsy, which carries surgical risks and complications; (2) Biopsy sampling may have sampling errors and cannot fully reflect the spatial heterogeneity of tumors; (3) For tumors located deep within functional areas, biopsy carries a higher risk and clinical procedures are limited; (4) Dynamic monitoring is not possible, making it difficult to assess changes in EGF expression during treatment; (5) The pathological testing cycle is relatively long, which affects the timeliness of clinical decision-making; Radiomics is a novel image analysis technology that has emerged in recent years. It extracts quantitative features from medical images through high-throughput processing, transforms image data into high-dimensional feature data that can be mined, and then builds predictive models to achieve non-invasive assessment of tumor biological characteristics, gene expression, and prognosis. Radiomics has the advantages of being non-invasive, repeatable, and dynamically monitorable, and shows broad application prospects in precision diagnosis and treatment of tumors. Magnetic resonance imaging (MRI), especially contrast-enhanced T1-weighted imaging (T1CE), can clearly display the morphological features, enhancement patterns, and surrounding edema of high-grade gliomas. It is an important imaging examination tool for the diagnosis and treatment evaluation of gliomas. Radiomics analysis based on enhanced MRI can extract multi-dimensional features reflecting tumor density, texture, shape, and heterogeneity. These features have potential associations with the molecular phenotype and gene expression of tumors. However, there is currently a lack of systematic methods for non-invasively predicting EGF expression levels in high-grade gliomas based on enhanced MRI radiomics technology. Existing radiomics research mainly focuses on tumor grading, subtyping, and prognostic prediction, with relatively few studies on predicting specific gene expression. Furthermore, there are shortcomings in feature selection strategies, model building algorithms, and clinical validation. Therefore, there is an urgent need to develop an EGF expression prediction method based on enhanced MRI radiomics to achieve non-invasive, accurate, and reproducible EGF expression assessment, providing a new technical means for the precision diagnosis and treatment of high-grade gliomas. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide an evaluation method for enhanced MRI radiomics models to predict EGF. By extracting multi-dimensional omics features from enhanced MRI images, a machine learning algorithm is used to construct an EGF expression prediction model, enabling non-invasive and accurate assessment of EGF expression levels in high-grade gliomas, and providing technical support for precision clinical diagnosis and treatment. Technical solution

[0004] To achieve the above-mentioned objectives, the technical solution provided by this invention is as follows: An evaluation method for an enhanced MRI radiomics model for predicting EGF, comprising the following steps: Step S1: Acquire enhanced MRI image data Enhanced MRI images of patients with high-grade gliomas were acquired, including contrast-enhanced T1-weighted sequences (T1CE). The acquisition of the images followed standard MRI scanning protocols to ensure that the image quality met diagnostic requirements.

[0005] Step S2: Image Preprocessing The acquired enhanced MRI image data undergoes preprocessing, specifically including: (1) N4 bias correction: eliminates image intensity deviation caused by magnetic field inhomogeneity and improves image quality; (2) Intensity normalization: The z-score normalization method is used to normalize the image intensity to eliminate the intensity differences caused by different scanners and different scanning parameters; (3) Isotropic resampling: The image is resampled to a uniform voxel size, preferably 1mm×1mm×1mm, to ensure the consistency of feature extraction.

[0006] Step S3: Outline the area of ​​interest Experienced radiologists manually delineate tumor regions on contrast-enhanced T1-weighted images as regions of interest (ROIs). These ROIs include both enhanced and non-enhanced tumor areas, encompassing the entire three-dimensional structure of the tumor. The delineation process utilizes specialized medical image analysis software to ensure the accuracy and repeatability of the ROIs.

[0007] Step S4: Extract radiomics features Imagemic features are extracted from the region of interest, the imagemic features including: (1) First-order statistical characteristics: statistical characteristics reflecting the distribution of voxel intensity within the ROI, including mean, standard deviation, skewness, kurtosis, energy, and entropy; (2) Shape characteristics: Reflecting the three-dimensional morphological characteristics of the tumor, including volume, surface area, sphericity, compressibility, and elongation; (3) Texture features: Reflecting the spatial distribution pattern and heterogeneity of the image, including: Gray Level Co-occurrence Matrix (GLCM) features: describe the co-occurrence relationship of different gray values ​​in space, including contrast, correlation, energy, entropy, uniformity, difference, information correlation 1, and information correlation 2. Gray Level Run Length Matrix (GLLM) characteristics: describes the length distribution of consecutive occurrences of specific gray values, including short run emphasis, long run emphasis, gray level non-uniformity, run length non-uniformity, and run percentage; Gray Level Size Zone Matrix (GLSZM) features: describe the size distribution of connected regions, including small region emphasis, large region emphasis, gray level non-uniformity, region size non-uniformity, and region percentage; Neighboring Gray Tone Difference Matrix (NGTDM) features: describe the gray level difference between the central voxel and its neighboring voxels, including roughness, contrast, busyness, complexity, and intensity; (4) Wavelet transform features: The image is decomposed into different frequency components by wavelet decomposition, and the first-order statistical features and texture features mentioned above are extracted from the decomposed sub-band images to capture the multi-scale information of the image.

[0008] The extraction of radiomics features employs standardized feature extraction algorithms, with PyRadiomics software being the preferred choice to ensure the repeatability and standardization of feature extraction.

[0009] Step S5: Perform preliminary feature selection using the maximum correlation minimum redundancy algorithm. Because the number of extracted radiomics features is large (usually between 1,000 and 1,500), and there is strong correlation and redundancy among the features, feature filtering is required to reduce dimensionality and improve the generalization ability of the model.

[0010] The Maximum Relevance Minimum Redundancy (mRMR) algorithm was used to initially screen radiomics features. The core idea of ​​this algorithm is to ensure that the features are highly correlated with the target variable (EGF expression level) while minimizing the redundancy between features.

[0011] Specifically, the mRMR algorithm is based on mutual information measurement. It calculates the mutual information between each feature and the EGF expression level (reflecting correlation) as well as the mutual information between features (reflecting redundancy). It iteratively selects features by maximizing the average mutual information between the selected features and the EGF expression level and minimizing the average mutual information between the selected features, forming a feature subset.

[0012] In a preferred embodiment of the present invention, the top 30 features are selected using the mRMR algorithm to form a feature subset, which retains important prediction information while significantly reducing the feature dimensionality.

[0013] Step S6: Further screening is performed using a recursive feature elimination algorithm. The feature subset obtained by the mRMR algorithm is further filtered using the Recursive Feature Elimination (RFE) algorithm to obtain key features with strong predictive power.

[0014] The working principle of the RFE algorithm is as follows: (1) Train a machine learning model (such as a support vector machine or random forest) using all feature subsets. (2) Based on the feature importance scores given by the model, remove several features with low importance; (3) Retrain the model using the remaining features; (4) Repeat the above process until the preset number of features is reached.

[0015] The RFE algorithm can dynamically evaluate the contribution of features during model training, gradually eliminate features that contribute little to the model's predictive ability, and finally obtain the key features.

[0016] In a preferred embodiment of the present invention, the RFE algorithm ultimately selects two key features: the information correlation 2 feature of the gray-level co-occurrence matrix (GLCM Imc2) and the correlation feature of the gray-level co-occurrence matrix (GLCM Correlation). These two features reflect the spatial correlation and complexity of gray values ​​within the tumor region and are closely related to the tumor's microstructure and heterogeneity.

[0017] Step S7: Construct an EGF expression prediction model using the gradient enhancement algorithm. Based on the key features obtained through screening, an EGF expression level prediction model was constructed using the Gradient Boosting Machine (GBM) algorithm.

[0018] Gradient boosting is an ensemble learning method that trains multiple weak learners (usually decision trees) sequentially. Each new learner focuses on fitting the residuals (i.e., prediction errors) of the previous learner. Finally, the predictions of all weak learners are weighted and combined to obtain a strong learner. Gradient boosting has strong prediction ability and generalization performance, and can effectively handle nonlinear relationships and feature interactions.

[0019] In this invention, the specific implementation steps of the gradient enhancement algorithm are as follows: (1) Data set partitioning: The collected patient data is randomly divided into a training set and a validation set, with the preferred partitioning ratio being 7:3; (2) Model training: Using training set data, a decision tree is used as the base learner, and the model is trained iteratively through gradient boosting; (3) Hyperparameter tuning: The 10-fold cross-validation method is used to perform grid search and optimization on the training set for the key hyperparameters of the model (such as learning rate, tree depth, number of trees, and minimum number of leaf node samples) to select the hyperparameter combination that minimizes the cross-validation error. (4) Model determination: The final prediction model is trained on the complete training set using the optimized hyperparameters.

[0020] The trained prediction model can output the predicted probability of EGF expression level based on the input key radiomics features.

[0021] Step S8: Calculate radiomics score Using the trained prediction model, the patient's key radiomics features are input to calculate the radiomics score (Rad-score).

[0022] The radiomics score is a continuous numerical value output by the model, reflecting the probability of EGF expression level predicted based on image features. This score integrates information from key features, transforming multidimensional image features into a single index that is easy to understand and apply.

[0023] The process of calculating the radiomics score is as follows: the patient's key feature values ​​are input into the trained gradient enhancement model, and the model outputs the probability value of high EGF expression through the ensemble prediction of multiple decision trees. This probability value is the radiomics score.

[0024] Step S9: Predict EGF expression levels Based on radiomics scores, the patient's EGF expression level can be predicted, enabling non-invasive assessment of EGF expression in high-grade gliomas.

[0025] Specifically, radiomics scores are divided into high-expression and low-expression groups based on a preset threshold. The preset threshold is determined in the training dataset using survival analysis methods, and is usually selected to make the high-expression and low-expression groups significantly different in terms of survival outcomes.

[0026] In one specific embodiment of the present invention, by analyzing the survival data of patients in the training dataset, the Kaplan-Meier method and logarithmic test are used to determine the threshold of the radiomics score. When the patient's radiomics score is higher than the threshold, it is determined to be EGF high expression; when it is lower than the threshold, it is determined to be EGF low expression.

[0027] Through the above steps, non-invasive prediction of EGF expression levels based on enhanced MRI radiomics was achieved.

[0028] Step S10: Model Evaluation To evaluate the performance of the prediction model, multiple evaluation metrics were used to assess the model's performance on the validation set, including: (1) Receiver Operating Characteristic (ROC) curve: The curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis to intuitively show the classification performance of the model at different thresholds; (2) Area Under the Curve (AUC): The area under the ROC curve reflects the overall discrimination ability of the model. The AUC value is between 0.5 and 1. The closer it is to 1, the better the model performance. (3) Accuracy: The proportion of correctly predicted samples out of the total number of samples; (4) Sensitivity: The proportion of true positive samples that are correctly predicted as positive, reflecting the model's ability to identify patients with high expression; (5) Specificity: The proportion of true negative samples that are correctly predicted as negative, reflecting the model's ability to identify patients with low expression; (6) Positive Predictive Value (PPV): The proportion of samples predicted to be positive that are actually positive; (7) Negative Predictive Value (NPV): The proportion of samples that are predicted to be negative but are actually negative.

[0029] In addition, calibration curves can be plotted to assess the consistency between predicted probabilities and actual incidence rates, and decision curve analysis (DCA) can be plotted to assess the clinical applicability of the model.

[0030] In one specific embodiment of the present invention, the constructed prediction model achieves an AUC of 0.809 on the validation set and an AUC of 0.734 on the 10-fold cross-validation set, demonstrating good prediction performance and stability.

[0031] Step S11: Prognostic Assessment Survival analysis based on radiomics scores to assess patient prognosis involves the following steps: (1) Collect patient follow-up data, including overall survival (OS) and survival status; (2) Patients were divided into high-scoring and low-scoring groups based on their radiomics scores; (3) The Kaplan-Meier method was used to plot survival curves, and the survival differences between the two groups of patients were compared. The log-rank test was used for statistical analysis. (4) Univariate and multivariate analyses were performed using the Cox Proportional Hazards Model to assess whether radiomics scores were independent prognostic factors and to calculate the hazard ratio (HR) and its 95% confidence interval.

[0032] Survival analysis revealed that radiomics scores were significantly associated with patient prognosis, with patients in the high-scoring group having significantly shorter survival times than those in the low-scoring group. Radiomics scores were an independent prognostic risk factor, indicating that radiomics-based EGF expression prediction not only reflects molecular phenotype but also has important prognostic value, which can assist clinicians in developing individualized treatment plans.

[0033] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. This invention is based on conventional enhanced MRI images, eliminating the need for invasive biopsies and avoiding the risks and complications associated with surgery or puncture. It has high patient acceptance and a wide range of applications.

[0034] 2. Biopsy sampling can only represent local tumor tissue, while this invention is based on the imaging characteristics of the entire tumor region, which can comprehensively reflect the spatial heterogeneity of the tumor and provide a more accurate assessment of EGF expression.

[0035] 3. Imaging examinations can be repeated. The method of this invention can be used for dynamic monitoring during treatment, assessing changes in EGF expression, and adjusting the treatment plan in a timely manner.

[0036] 4. Through scientific feature selection strategies and advanced machine learning algorithms, the prediction model constructed in this invention achieves an AUC of 0.809 on the validation set, demonstrating high prediction accuracy and stability.

[0037] 5. This invention employs a two-stage feature selection strategy of mRMR and RFE, effectively reducing feature dimensionality, avoiding overfitting, and improving the model's generalization ability. Accurate prediction can be achieved with only two key features, simplifying model complexity and facilitating clinical application.

[0038] 6. The radiomics score of this invention can not only predict EGF expression level, but also is significantly associated with the patient's overall survival, and is an independent prognostic factor, providing a new tool for clinical prognostic assessment.

[0039] 7. The EGF / EGFR signaling pathway is an important therapeutic target. Accurately assessing EGF expression levels can help screen patients who may benefit from EGFR-targeted therapy and achieve precision treatment.

[0040] 8. This invention adopts a standardized image preprocessing process and feature extraction method, which has good repeatability and is easy to promote and apply in different medical institutions.

[0041] 9. The method of the present invention has a clear process, is easy to operate, and has high computational efficiency. It can be easily integrated into a clinical image analysis system to provide clinicians with fast and accurate decision support.

[0042] 10. The method of the present invention has the potential to be validated and optimized on multi-center, large-sample data. With the accumulation of data and model iteration, the prediction performance is expected to be further improved.

[0043] In summary, this invention provides an innovative, practical, and efficient method for predicting EGF expression, offering crucial technical support for the precise diagnosis and treatment of high-grade gliomas. It has significant clinical application value and broad prospects for wider application. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall technical process of the method of the present invention; Figure 2 This is a schematic diagram of the image preprocessing workflow; Figure 3 This is a schematic diagram of feature extraction and classification. Figure 4 This is a schematic diagram of the feature filtering process; Figure 5 This is a schematic diagram of the model building and evaluation process; Figure 6 A schematic diagram of the predictive application process; Figure 7 A schematic diagram of the ROI is shown in the specific embodiment; Figure 8 This is a schematic diagram of the model evaluation results in a specific embodiment. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0046] Example 1: Prediction of EGF expression in high-grade gliomas based on TCGA and TCIA databases I. Data Sources and Patient Selection The research data in this embodiment comes from the publicly available Cancer Imaging Archive (TCIA) database and the Cancer Genome Atlas (TCGA) database.

[0047] Enhanced MRI images of patients with high-grade gliomas were obtained from the TCIA database, and RNA sequencing data (including the mRNA expression level of the EGF gene) and clinical information of the corresponding patients were obtained from the TCGA database.

[0048] Patient inclusion criteria were: (1) WHO grade III or IV high-grade glioma confirmed by pathology; (2) Preoperative contrast-enhanced T1-weighted MRI images; (3) The corresponding EGF gene mRNA expression data are available in the TCGA database; (4) Complete clinical data and survival information.

[0049] The patient exclusion criteria are: (1) poor image quality, making accurate feature extraction impossible; (2) Key clinical data are missing.

[0050] After screening, 182 patients were ultimately included in the analysis of this example. EGF expression levels were divided into high expression group and low expression group based on the median mRNA expression value. The clinical characteristics of the patients included: The age range was 18 to 85 years, with a median age of 60 years. There were 98 male cases and 84 female cases; WHO Class III: 58 cases; Class IV: 124 cases.

[0051] II. Image Data Acquisition and Preprocessing (Steps S1-S2) 1. Acquisition of MRI Image Data MRI image data of the included patients were downloaded from the TCIA database, including contrast-enhanced T1-weighted sequences (T1CE). The image data was stored in DICOM format and contained complete scan parameter information.

[0052] 2. Image preprocessing All MRI images underwent standardized preprocessing, with the following specific steps: (1) Format conversion: Convert DICOM format image data to NIfTI format for easier subsequent processing and analysis; (2) N4 Bias Field Correction: The N4BiasFieldCorrection algorithm in the ITK (Insight Segmentation and Registration Toolkit) software package is used to correct the bias field of the image and eliminate the image intensity deviation caused by magnetic field inhomogeneity. This algorithm improves the image quality by iteratively estimating and correcting the bias field. (3) Intensity normalization: The z-score normalization method is used to normalize the image intensity. For each patient's image, the average intensity value and standard deviation of the whole brain region are calculated. The intensity value of each voxel is subtracted from the average value and then divided by the standard deviation to make the image intensity distribution of different patients comparable. (4) Isotropic resampling: The image is resampled to a uniform voxel size of 1mm×1mm×1mm using linear interpolation. The slice thickness and interslice spacing of the original MRI image may be inconsistent. Resampling can ensure the consistency and comparability of feature extraction.

[0053] The image quality is significantly improved after preprocessing, laying the foundation for accurate feature extraction in the subsequent process. The image preprocessing workflow is as follows: Figure 2 As shown.

[0054] III. Delineating the Tumor Area (Step S3) Two neuroradiologists with over 10 years of experience independently delineated the tumor region on contrast-enhanced T1-weighted images without knowing the patient's EGF expression data or clinical outcome.

[0055] The outlined area includes: (1) Enhanced tumor area: The solid portion of the tumor that shows significant enhancement on T1CE images; (2) Non-enhancing tumor areas: Necrotic areas and cystic areas inside the tumor.

[0056] The delineation tool uses ITK-SNAP version 3.8 software, which provides a user-friendly 3D visualization interface and precise delineation tools. Physicians delineate the tumor boundary layer by layer on axial images and check and correct it by combining coronal and sagittal images to ensure the 3D integrity and accuracy of the ROI.

[0057] For cases where two physicians' ROI delineation results are inconsistent, a consensus should be reached through consultation. The ROI delineation diagram is shown below. Figure 7 As shown.

[0058] To assess the reproducibility of ROI delineation, 20 patients were randomly selected and delineated again independently by a second physician. Radiomics features were extracted from the two delineated ROIs, and the intraclass correlation coefficient (ICC) was calculated. The results showed that the median ICC for all features was 0.963, indicating that ROI delineation had good reproducibility.

[0059] IV. Radiomics Feature Extraction (Step S4) Radiomics features were extracted from the ROI of each patient using the PyRadiomics package (version 3.0.1). PyRadiomics is an open-source Python package that follows the ImageBiomarker Standardization Initiative (IBSI) standards to ensure the standardization and reproducibility of feature extraction.

[0060] The parameter settings for feature extraction are as follows: Image grayscale discretization: A fixed-width method was used to discretize the grayscale values ​​into 25 grayscale levels; Resampling: The image has been resampled to 1mm×1mm×1mm before feature extraction; Texture feature calculation: A 3D method is used, considering 26 connected neighborhoods.

[0061] The extracted radiomics features include the following categories: 1. First-order statistical features (18): Calculate the statistical distribution characteristics of voxel intensity within the ROI, including: mean, median, standard deviation, variance, skewness, kurtosis, energy, total energy, entropy, minimum, 10th percentile, 90th percentile, range, mean absolute deviation, robust mean absolute deviation, root mean square, consistency, and coefficient of variation.

[0062] 2. Shape characteristics (14): Describe the three-dimensional morphological characteristics of the tumor, including: volume, surface area, surface-to-volume ratio, sphericity, compressibility 1, compressibility 2, sphericity deviation, elongation, flatness, area-to-density ratio, axial length, minor axis length, major axis length, and convex hull volume.

[0063] 3. Gray-level co-occurrence matrix features (24): extracted based on GLCM, including: autocorrelation, clustering tendency, clustering shadow, clustering prominence, contrast, correlation, difference mean, difference variance, difference entropy, summation mean, summation variance, summation entropy, joint mean, joint variance, joint entropy, information correlation 1, information correlation 2, inverse moment normalization, inverse moment, inverse variance, maximum correlation coefficient, difference entropy normalization, entropy, and energy.

[0064] 4. Gray-scale run matrix features (16): extracted based on GLRLM, including: short run emphasis, long run emphasis, gray-scale non-uniformity, gray-scale non-uniformity normalization, run length non-uniformity, run length non-uniformity normalization, run percentage, gray-scale variance, run variance, run entropy, low gray-scale run emphasis, high gray-scale run emphasis, short run low gray-scale emphasis, short run high gray-scale emphasis, long run low gray-scale emphasis, and long run high gray-scale emphasis.

[0065] 5. Gray-scale region size matrix features (16): Extracted based on GLSZM, including: small region emphasis, large region emphasis, gray-scale non-uniformity, gray-scale non-uniformity normalization, region size non-uniformity, region size non-uniformity normalization, region percentage, gray-scale variance, region variance, region entropy, low gray-scale region emphasis, high gray-scale region emphasis, small region low gray-scale emphasis, small region high gray-scale emphasis, large region low gray-scale emphasis, and large region high gray-scale emphasis.

[0066] 6. Neighborhood gray-level difference matrix features (5): extracted based on NGTDM, including: roughness, contrast, busyness, complexity, and intensity.

[0067] 7. Gray-level dependency matrix features (14): extracted based on GLDM, including: small dependency emphasis, large dependency emphasis, gray-level non-uniformity, dependency non-uniformity, dependency non-uniformity normalization, gray-level variance, dependency variance, dependency entropy, low gray-level dependency emphasis, high gray-level dependency emphasis, small dependency low gray-level emphasis, small dependency high gray-level emphasis, large dependency low gray-level emphasis, and large dependency high gray-level emphasis.

[0068] 8. Wavelet transform features (744): The original image is decomposed into 8 sub-band images (LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH). The first-order statistical features and texture features mentioned above are extracted from each sub-band image, for a total of 744 features.

[0069] In summary, a total of 1218 radiomics features were extracted from the ROI of each patient. The feature extraction classification is as follows: Figure 3 As shown.

[0070] V. Feature Filtering (Steps S5-S6) Faced with 1218 high-dimensional features, directly using them for modeling can easily lead to overfitting and computational burden. Therefore, feature screening is required to select key features that are highly correlated with EGF expression levels and have low redundancy.

[0071] 1. Feature standardization Before feature selection, all features are z-score standardized to make features of different dimensions comparable.

[0072] 2. mRMR Feature Filtering (Step S5) The maximum correlation minimum redundancy (mRMR) algorithm is used for preliminary feature screening. The goal of the mRMR algorithm is to select a subset of features that are highly correlated with EGF expression levels and have low redundancy between features.

[0073] In this embodiment, the mRMRe software package in R is used to implement the mRMR algorithm. Based on the patients' EGF gene mRNA expression levels, the median value is used as the boundary to divide the patients into high expression group and low expression group, which serve as classification labels.

[0074] 3. The specific steps of the mRMR algorithm are as follows: (1) Calculate the mutual information between each feature and the classification label to reflect the correlation between the feature and the EGF expression level; (2) Calculate the mutual information between features to reflect the redundancy between features; (3) Define the mRMR evaluation index: the difference between correlation and redundancy; (4) Iteratively select features that result in a larger mRMR evaluation index and gradually construct a feature subset; (5) Select the top 30 features to form a feature subset.

[0075] By using mRMR filtering, the number of feature dimensions was reduced from 1218 to 30.

[0076] RFE Feature Selection (Step S6) The 30 features obtained from mRMR screening were further filtered using the Recursive Feature Elimination (RFE) algorithm to obtain key features.

[0077] In this embodiment, the RFE algorithm is implemented using the caret package in R language. The RFE algorithm is combined with a random forest classifier to evaluate feature importance.

[0078] The specific steps of the RFE algorithm are as follows: (1) Train a random forest classifier using all 30 features; (2) Based on the feature importance score given by the classifier, remove several features with low importance (5 features are removed each time in this embodiment); (3) Retrain the classifier using the remaining features and calculate the cross-validation accuracy; (4) Repeat the above process and record the model performance under different numbers of features; (5) Select a subset of features that have a higher cross-validation accuracy.

[0079] The results of the RFE algorithm show that when the number of features is 2, the model's cross-validation accuracy reaches a high level. Further increasing the number of features does not significantly improve performance. Therefore, two key features were ultimately selected: Feature 1: Information relevance feature of the gray-level co-occurrence matrix (original_glcm_Imc2); Feature 2: Correlation features of the gray-level co-occurrence matrix (original_glcm_Correlation).

[0080] Both of these features are texture features, reflecting the spatial correlation and complexity of gray values ​​within the tumor region. The information correlation feature describes the information entropy of the gray-level co-occurrence matrix, reflecting the complexity of the image. Correlation features describe the degree of linear correlation between gray values ​​of adjacent pixels, reflecting the texture regularity of the image. These features are closely related to the microstructure, cell density, and blood vessel distribution of tumors, and thus have a potential association with the expression level of EGF.

[0081] Feature filtering process as follows Figure 4 As shown.

[0082] VI. Model Construction (Step S7) Based on the two key features obtained from screening, a gradient enhancement algorithm was used to construct an EGF expression level prediction model.

[0083] 1. Dataset partitioning The 182 patients were randomly divided into a training set (127 patients, approximately 70%) and a validation set (55 patients, approximately 30%). The training set was used for model training and hyperparameter optimization, while the validation set was used for model performance evaluation.

[0084] 2. Model Training Using the caret package in R, the gbm algorithm is called for model training. The gradient boosting algorithm uses decision trees as base learners and iteratively trains multiple decision trees through gradient boosting. Each new tree fits the residual of the previous round.

[0085] 3. Hyperparameter Tuning A 10-fold cross-validation method was used to perform grid search and optimization on the training set for the model's key hyperparameters. The adjusted hyperparameters included: Number of trees (n.trees): Candidate values ​​are 50, 100, 150, and 200; Interaction depth (interaction.depth), which is the depth of the tree: candidate values ​​are 1, 2, 3, and 4; Shrinkage, or learning rate: candidate values ​​are 0.01, 0.05, and 0.1. Minimum number of leaf node samples (n.minobsinnode): Candidate values ​​are 5, 10, and 15.

[0086] For each hyperparameter combination, calculate the average accuracy of 10-fold cross-validation. Then, through grid search, select the hyperparameter combination that results in the highest cross-validation accuracy. n.trees=150,interaction.depth=3,shrinkage=0.05,n.minobsinnode=10.

[0087] 4. Final Model Training Using optimized hyperparameters, the final prediction model is trained on the complete training set. After training, the model can output the prediction probability of high EGF expression based on the two key radiomics features input.

[0088] Model building process as follows Figure 5 As shown.

[0089] VII. Model Evaluation (Step S10) The model was evaluated on an independent validation set (55 cases), and various performance metrics were calculated.

[0090] 1. ROC curve and AUC Using the model's predicted high EGF expression probability as the predictor variable and the actual EGF expression level (high expression / low expression) as the true label, an ROC curve was plotted. The results show that the model's AUC is 0.809, indicating that the model has good discriminative ability.

[0091] Furthermore, 10-fold cross-validation was performed on the training set, and the average AUC of the cross-validation was 0.734, indicating that the model has good stability and generalization ability.

[0092] 2. Accuracy, sensitivity, and specificity Choosing a threshold with a larger Youden index as the classification cutoff, the accuracy, sensitivity, and specificity of the model are calculated: Accuracy: 78.2%; Sensitivity: 80.6%; Specificity: 75.0%; Positive predictive value: 76.5%; Negative predictive value: 79.3%.

[0093] The above indicators show that the model has high accuracy in identifying patients with both high and low EGF expression.

[0094] 3. Calibration curve A calibration curve was plotted to assess the consistency between the model's predicted probability and the actual occurrence rate. The calibration curve was close to the diagonal, and the p-value of the Hosmer-Lemeshow goodness-of-fit test was 0.68 (greater than 0.05), indicating that the model was well calibrated and the predicted probability was consistent with the actual situation.

[0095] 4. Decision Curve Analysis Decision curves were plotted to evaluate the clinical applicability of the model. The results showed that, within a reasonable threshold probability range, using the model for prediction had a higher net benefit than the "all treatment" or "no treatment" strategies, indicating that the model has good clinical application value.

[0096] Model evaluation results are as follows Figure 8 As shown.

[0097] VIII. Radiomics Score Calculation and EGF Expression Prediction (Steps S8-S9) For each patient, two key radiomics features are input into a pre-trained gradient enhancement model. The model outputs the predicted probability of high EGF expression, which is the radiomics score (Rad-score).

[0098] To facilitate clinical application, patients were divided into high-expression and low-expression groups based on Rad-score. Kaplan-Meier analysis was performed on the survival data of patients in the training set, and the surv_cutpoint function in the survminer package was used to determine the cutoff point that made the survival difference between the two groups significant. In this embodiment, the cutoff value of Rad-score was 0.613.

[0099] When a patient's Rad-score is greater than 0.613, it is considered that EGF is highly expressed; A value less than 0.613 is considered low EGF expression.

[0100] On the validation set, Rad-score showed a significant difference between the high and low EGF expression groups (p < 0.001), indicating that radiomics scoring can effectively reflect EGF expression levels.

[0101] IX. Prognostic Assessment (Step S11) Survival analysis based on radiomics scores was conducted to assess their prognostic predictive value.

[0102] 1. Kaplan-Meier survival analysis Based on the Rad-score, 182 patients were divided into a high group (Rad-score greater than 0.613, 74 patients) and a low group (Rad-score less than 0.613, 108 patients). Follow-up data of the patients were collected, including overall survival (OS) and survival status.

[0103] The Kaplan-Meier survival curves were plotted, and the results showed: The median survival in the high-risk group was 17.03 months; The median survival in the low-risk group was 24.5 months.

[0104] The p-value of the logarithmic test was less than 0.001, indicating that the survival difference between the two groups of patients was statistically significant, and the prognosis of patients in the high-risk group was significantly worse.

[0105] 2. Cox proportional hazards regression analysis We used a Cox proportional hazards regression model to conduct univariate and multivariate analyses to assess whether the Rad-score was an independent prognostic factor.

[0106] Univariate analysis showed that Rad-score was a significant risk factor for overall survival (HR=1.74, 95%, CI: 1.211-2.498, p=0.003).

[0107] In the multivariate analysis, covariates such as age, sex, tumor grade, IDH mutation status, MGMT methylation status, chemotherapy, and radiotherapy were included. After adjusting for these confounding factors, Rad-score remained an independent prognostic risk factor (HR=1.451, 95%, CI: 1.002-2.102, p=0.049).

[0108] 3. Subgroup analysis The prognostic value of Rad-score was assessed in different clinical subgroups, and the results showed: In female patients, a high Rad-score was a significant risk factor (HR=2.173, p=0.009). In male patients, a high Rad-score was also a significant risk factor (HR=1.609, p=0.036). In patients under 60 years of age, a high Rad-score significantly predicted a poor prognosis (p=0.034). In patients receiving chemotherapy, a high Rad-score significantly increased the risk of death (p=0.009). A high Rad-score also had prognostic value in patients receiving radiotherapy (p=0.048).

[0109] The above results indicate that radiomics scores have stable prognostic predictive ability across different clinical subgroups.

[0110] X. Summary of Results Based on enhanced MRI images and gene expression data from 182 patients with high-grade gliomas, this embodiment successfully constructed a radiomics model for predicting EGF expression levels. This model requires only two key texture features to accurately predict EGF expression, achieving an AUC of 0.809. The radiomics score is significantly correlated with patient prognosis and is an independent risk factor.

[0111] This embodiment fully verifies the feasibility and effectiveness of the method of the present invention, and provides an important technical means for clinical non-invasive assessment of EGF expression level in high-grade gliomas, guidance of individualized treatment, and assessment of patient prognosis.

[0112] Example 2: Clinical application process of the model This embodiment describes the specific application process of the method of the present invention in clinical practice.

[0113] I. Patient's visit A 58-year-old male patient presented to the neurosurgery department with headache and epileptic seizures, and clinical symptoms suggested an intracranial space-occupying lesion.

[0114] II. MRI Examination The patient underwent a head MRI, including standard sequences and contrast-enhanced T1-weighted sequences, and the MRI scans followed standard protocols. Scanner: 3.0T MRI scanner; Sequence parameters: Contrast-enhanced T1-weighted sequence, TR=500ms, TE=15ms, slice thickness=5mm, inter-slice spacing=1mm; Contrast agent: Gadolinium contrast agent was administered intravenously at a dose of 0.1 mmol / kg body weight.

[0115] MRI results showed a space-occupying lesion in the left frontal lobe with indistinct borders, central necrosis, significant peripheral enhancement, and surrounding edema. The preliminary diagnosis was high-grade glioma.

[0116] III. Image Data Import and Preprocessing The patient's MRI image data (DICOM format) is imported into the radiomics analysis system, where it is automatically preprocessed. (1) Convert the format to NIfTI format; (2) N4 field correction to eliminate image intensity deviation; (3) Intensity normalization, using z-score standardization; (4) Isotropic resampling to 1mm×1mm×1mm.

[0117] The preprocessed image is of good quality and suitable for feature extraction.

[0118] IV. Tumor Area Delineation The tumor region was delineated on contrast-enhanced T1-weighted images by radiologists. The delineation range included the enhanced tumor area and the central necrotic area. The ITK-SNAP software was used to delineate the region layer by layer on axial, coronal and sagittal images to ensure the three-dimensional integrity of the ROI.

[0119] After the ROI is drawn, the system automatically saves the coordinate information of the ROI in preparation for feature extraction.

[0120] V. Feature Extraction and Prediction The system invokes a pre-trained radiomics model and automatically completes the following steps: (1) 1218 radiomics features were extracted from the patient's ROI; (2) Feature standardization: z-score transformation is performed using the mean and standard deviation of the training set; (3) Select two key features required for the model: the information relevance feature of the gray-level co-occurrence matrix and the relevance feature; (4) Input the key features into the trained gradient boosting model; (5) Model output radiomics score (Rad-score): 0.68.

[0121] VI. EGF Expression Prediction and Clinical Interpretation Based on Rad-score = 0.68, which is greater than the preset threshold of 0.613, the patient was determined to have high EGF expression.

[0122] The system generates a clinical report, which includes: Radiomics score: 0.68; EGF expression prediction: High expression; Prognostic prediction: According to the survival analysis model, this patient belongs to the high-risk group, with a median survival of approximately 17 months; Treatment recommendations: EGFR-targeted therapy, such as EGFR inhibitors or EGFR antibody-drug conjugates (EGFR-ADCs), is recommended, which may yield better treatment results.

[0123] VII. Multidisciplinary Discussion and Treatment Decision Neurosurgeons, medical oncologists, radiation oncologists, and radiologists collaborate in multidisciplinary team (MDT) discussions, combining the patient's imaging findings, radiomics scores, predicted EGF expression levels, and prognostic information to develop an individualized treatment plan. (1) Surgical treatment: to remove the tumor as safely and extensively as possible; (2) Postoperative pathological biopsy: to confirm the pathological type and molecular markers, including immunohistochemical detection of EGFR expression; (3) Postoperative adjuvant therapy: concurrent chemoradiotherapy combined with EGFR targeted therapy; (4) Regular follow-up: MRI is performed every 3 months, and radiomics models are used to monitor treatment response and disease progression.

[0124] VIII. Postoperative Verification and Follow-up The patient underwent surgery, and postoperative pathology confirmed glioblastoma (WHO grade IV). Immunohistochemical staining showed high expression of EGFR protein, consistent with the prediction of the radiomics model, thus verifying the accuracy of the model.

[0125] The patient received adjuvant therapy after surgery according to the established treatment plan, including radiotherapy, temozolomide chemotherapy and EGFR targeted therapy. MRI was performed every 3 months, and radiomics models were used to monitor tumor changes and treatment response.

[0126] During the 12-month follow-up after surgery, the patient's condition remained stable with no significant progression. The radiomics score decreased from 0.65 at 3 months post-surgery to 0.58 at 12 months post-surgery, indicating that the treatment was effective and the tumor's biological behavior was under control to some extent.

[0127] IX. Summary of Clinical Application This embodiment demonstrates the complete application process of the method of the present invention in clinical practice. Through non-invasive radiomics analysis, the patient's EGF expression level and prognosis can be predicted before surgery, providing an important basis for developing individualized treatment plans. The prediction results of the radiomics model are consistent with the postoperative pathological examination, verifying the clinical applicability of the model.

[0128] Furthermore, radiomics scoring can be used for dynamic monitoring during treatment, assessing treatment response and disease progression, and adjusting treatment strategies in a timely manner, demonstrating the application value of radiomics in precision medicine. It also marks the first time that non-invasive prediction of EGF based on conventional enhanced MRI has been achieved, which can provide radiological biomarkers for EGFR-targeted therapy.

[0129] Predictive application process such as Figure 6 As shown.

[0130] Example 3: Model Optimization and External Validation To further improve the model's performance and generalization ability, this embodiment describes the process of optimizing the model and performing external validation.

[0131] I. Data Expansion Over time, more data on high-grade glioma patients were collected from the TCIA and TCGA databases and collaborating medical institutions, increasing the sample size to 300 cases. The increased sample size was used for model retraining and optimization.

[0132] II. Feature Engineering Optimization On a larger dataset, feature selection was re-performed. In addition to the original mRMR and RFE algorithms, other feature selection methods such as LASSO regression and elastic network regression were also tried and compared.

[0133] The results showed that, in 300 samples, the number of key features screened by the mRMR+RFE method increased to 5, including the original 2 texture features and 3 newly added wavelet transform features. These features comprehensively reflect the texture heterogeneity and multi-scale information of tumors.

[0134] III. Model Algorithm Comparison In addition to gradient boosting, other machine learning algorithms were also tried, including: Support Vector Machine (SVM) Random Forest (RF); Logistic Regression; Deep Neural Network (DNN).

[0135] The performance of different algorithms was compared on the same training and validation sets, and the results show: Gradient boosting algorithm: AUC=0.823; Support Vector Machine: AUC = 0.805; Random Forest: AUC=0.798; Logistic regression: AUC = 0.772; Deep Neural Network: AUC=0.815.

[0136] The gradient boosting algorithm still performs well and has moderate model complexity, making it suitable for clinical applications. Therefore, the gradient boosting algorithm was chosen as the final modeling method.

[0137] IV. External Validation To evaluate the model's generalization ability, MRI images and EGF expression data of 80 patients with high-grade gliomas were collected from two independent medical institutions as an external validation set. These data, which came from different centers and different scanners than the training data, better reflect the model's performance in real clinical settings.

[0138] On the external validation set, the model's performance metrics are: AUC: 0.785; Accuracy: 74.3%; Sensitivity: 76.9%; Specificity: 71.4%.

[0139] External validation results show that the model has good generalization ability and can maintain high prediction accuracy on data with different centers. Although the AUC is slightly lower than that of the internal validation set, it is still within an acceptable range, indicating that the model has a certain degree of cross-center applicability.

[0140] V. Model Iteration and Continuous Improvement Based on feedback from external validation, the cases in which the model predicted errors were analyzed, and it was found that some misjudgments were related to the following factors: (1) Image quality differences: Image quality differences caused by different scanners and scanning parameters affect feature extraction; (2) Differences in ROI delineation: Different physicians' delineation habits may affect the stability of features; (3) Tumor heterogeneity: Some tumors exhibit significant spatial heterogeneity, which cannot be fully reflected by a single imaging feature.

[0141] To address the above issues, the following improvement measures are proposed: (1) Standardized scanning protocol: Develop unified recommendations for MRI scanning parameters to reduce inter-center differences; (2) Develop automatic segmentation algorithm: Introduce deep learning methods to automatically segment tumors and reduce human delineation errors; (3) Multi-region feature fusion: Features are extracted from different regions of the tumor (such as enhancement area, necrosis area, edema area) and fused to form a model, which more comprehensively reflects the heterogeneity of the tumor.

[0142] With continuous data accumulation and model iteration, the performance of radiomics models is expected to be further improved.

[0143] VI. Multicenter prospective study To further validate the clinical value of the model, a multicenter prospective study was launched, planning to enroll 200 patients with high-grade gliomas in 5 medical institutions. Using a unified MRI scanning protocol and radiomics analysis process, the study will prospectively evaluate the model's predictive ability for EGF expression and its prognostic value for patients.

[0144] The study has been registered with the Chinese Clinical Trial Registry, and the results will further promote the application and translation of radiomics technology in clinical practice.

[0145] In summary, this embodiment demonstrates the complete process of model optimization and validation through data expansion, algorithm comparison, external validation, and continuous improvement, providing a reference path for the clinical translation of radiomics technologies.

[0146] The specific embodiments of the present invention have been described in detail above. It should be understood that the scope of protection of the present invention is not limited to the specific embodiments described above. All equivalent transformations or substitutions based on the technical solutions of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating enhanced MRI radiomics models for predicting EGF, comprising an evaluation method, characterized in that: The evaluation method includes the following steps: Step S1: Acquire enhanced MRI image data of patients with high-grade gliomas, the enhanced MRI image data including contrast-enhanced T1-weighted sequences; Step S2: Preprocess the enhanced MRI image data, including N4 field correction, intensity normalization, and isotropic resampling; Step S3: Delineate the tumor region as the region of interest on the contrast-enhanced T1-weighted sequence image; Step S4: Extract radiomics features from the region of interest (ROI), including first-order statistical features, shape features, texture features, and wavelet transform features; Step S5: The maximum correlation minimum redundancy algorithm is used to initially screen the image omics features. Based on mutual information measurement, the features are selected to form a feature subset by maximizing the average mutual information between the feature and the EGF expression level and minimizing the average mutual information between features. Step S6: The recursive feature elimination algorithm is used to further filter the feature subset. By iteratively training the model and successively eliminating features with lower importance, key features are obtained. Step S7: Based on the key features, construct an EGF expression level prediction model using a gradient enhancement algorithm; Step S8: Calculate the patient's radiomics score using the prediction model; Step S9: Predict the patient's EGF expression level based on the radiomics score to achieve non-invasive assessment of EGF expression in high-grade gliomas.

2. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The texture features mentioned in step S4 include gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level region size matrix features, and neighborhood gray-level difference matrix features.

3. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 2, characterized in that: The key features obtained in step S6 include the information relevance feature of the gray-level co-occurrence matrix and the relevance feature of the gray-level co-occurrence matrix.

4. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The voxel size for isotropic resampling in step S2 is 1mm × 1mm × 1mm.

5. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The gradient boosting algorithm described in step S7 uses a decision tree as the base learner and adjusts the hyperparameters using a 10-fold cross-validation method.

6. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The radiomics score mentioned in step S8 is calculated by inputting key features into the trained prediction model, and reflects the probability value of EGF expression level.

7. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: In step S9, the radiomics scores are divided into high expression group and low expression group according to a preset threshold, which is determined in the training dataset by survival analysis.

8. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The number of radiomics features extracted in step S4 is between 1,000 and 1,500.

9. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The evaluation method further includes step S10: evaluating the prediction model, with evaluation indicators including the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

10. The method for evaluating enhanced MRI radiomics models for predicting EGF according to claim 1, characterized in that: The assessment method further includes step S11: performing survival analysis based on the radiomics score, assessing the patient's overall survival using Kaplan-Meier survival curves and Cox proportional hazards regression models, and achieving prognostic assessment.

Images