Multi-center research method for evaluating activity of Crohn disease

Abstract

The invention relates to the field of Crohn's disease analysis, and discloses a multi-center research method for evaluating Crohn's disease activity, which comprises the following steps: firstly, carrying out radiomics model multi-region focus segmentation, feature extraction, feature selection, model establishment, multi-modal clinical construction, hybrid model construction, model evaluation and comparison and a statistical method; according to the method, a clinical prediction model is constructed by adopting a machine learning algorithm based on demographic characteristics, clinical manifestation, laboratory examination and image examination data of a CD patient; image omics features are extracted from intestinal wall and mesenteric fat areas of a CT small intestine angiography CTE image, clinical and image omics features are fused, and a plurality of machine learning models are constructed to evaluate the disease activity of a CD patient; the problems that in the prior art, a treatment scheme cannot be systematically formulated for a patient suffering from the Crohn disease, the disease progress cannot be monitored, the application vacancy of medical imaging and artificial intelligence in the field of inflammatory intestinal diseases exists, and the application effect is not ideal are solved.

Description

Technical Field

[0001] This invention relates to the field of Crohn's disease analysis, and more particularly to a multicenter study method for assessing Crohn's disease activity. Background Technology

[0002] Crohn's disease (CD) is a chronic, relapsing inflammatory bowel disease (IBD) that can affect any part of the digestive tract and is highly concerning due to its recurrent flare-ups and difficulty in treatment. Activity assessment of Crohn's disease is a crucial aspect of clinical management, as it directly impacts treatment strategy development and prognosis. Increasing evidence suggests that accurate, frequent, and objective assessment can guide treatment decisions, leading to better outcomes and preventing further bowel damage, hospitalization, or surgery. However, activity assessment still faces challenges and difficulties in clinical practice. Although various assessment methods exist, including clinical scoring systems, imaging studies, and colonoscopy, each method has its limitations.

[0003] Imaging plays a crucial role in the early diagnosis, activity assessment, and treatment efficacy monitoring of Crohn's disease (CD). Currently, three main imaging assessment methods are commonly used: ultrasound, CT enterography, and MR enterography. Ultrasound is safe and inexpensive, but its clinical application is not widespread due to its reliance on operator skill and susceptibility to interference from intestinal gas. MR enterography offers advantages such as no ionizing radiation and high soft tissue resolution, but appointment and examination times are long, requiring high patient cooperation, and it is susceptible to image artifacts caused by abdominal breathing. Compared to MR enterography, CT enterography (CTE) has become the primary means of assessing intestinal lesions due to its shorter imaging time and higher spatial and density resolution. CTE can observe transmural inflammation, mesenteric adipose tissue, and complications of CD such as abscesses and fistulas, offering advantages that cannot be replaced by endoscopy or laboratory tests. However, traditional imaging assessment relies on the experience of radiologists, lacks standardized quantitative indicators, and the observation results are subjective.

[0004] In summary, current methods for assessing CD activity face challenges in terms of accuracy, objectivity, and accessibility. Therefore, exploring non-invasive assessment methods based on radiomics and machine learning to improve the accuracy and reproducibility of CD activity assessment is of significant clinical importance.

[0005] Radiomics combines medical imaging and computer science to construct imaging biomarkers oriented towards clinical goals by mining multimodal and multiregional imaging features. This comprehensively and systematically reveals the intrinsic information of diseases, improving the accuracy of diagnosis or prognosis, facilitating lesion characterization, and enhancing patient monitoring and management. In recent years, radiomics research on gastrointestinal tumors such as colorectal cancer and stromal tumors has emerged in large numbers and is gradually expanding to inflammatory diseases. The application of radiomics in Crohn's disease (CD) research has also increased in recent years. The team led by Xuehua Li at the First Affiliated Hospital of Sun Yat-sen University developed and validated a radiomics model based on CT small bowel contrast imaging to assess intestinal fibrosis in Crohn's disease through a retrospective multicenter study. By comparing the radiomics model with the diagnostic results of two radiologists, they concluded that radiomics can more accurately describe intestinal fibrosis in Crohn's disease patients. Chen et al. developed a new radiomics nomogram based on CTE, providing a predictive tool for assessing the loss of secondary response after infliximab treatment in Crohn's disease patients. Liu et al. used a novel multiparameter model based on CT small bowel imaging scores to assess endoscopic activity and surgical risks in Crohn's disease patients. However, existing research mainly focuses on intestinal wall structural features, with less attention paid to the impact of mesenteric adipose tissue (MAT) on disease activity, while MAT plays an important role in the inflammatory response of CD.

[0006] Meanwhile, the application of machine learning in radiomics analysis is rapidly developing. Through feature selection and pattern recognition, machine learning can optimize the extraction of image biomarkers and improve the generalization ability of models. In the field of Crohn's disease activity assessment, machine learning algorithms can automatically extract high-dimensional image features (such as intestinal wall thickness, enhancement level, texture heterogeneity, etc.), combine them with clinical data and biomarkers, and construct multi-dimensional predictive models. For example, supervised learning algorithms (such as random forests and support vector machines) have been used to identify radiomics features highly correlated with endoscopic activity from CTE images and optimize the model's generalization ability through feature selection. Wasnik et al. constructed an automatic detection model for Crohn's disease lesions in CTE (AUC 0.94) by integrating radiomics and deep learning features. Its consistency with radiologist assessment was as high as Kappa=0.82, and the analysis efficiency was improved by more than 10 times, demonstrating the feasibility of machine learning in standardized image assessment. Integrated learning frameworks, by fusing radiomics features, laboratory indicators, and clinical scores, can establish more comprehensive assessment systems. For example, jointly modeling CTE imaging features with dynamic changes in FCP has been shown to significantly improve the predictive accuracy of mucosal healing status. However, existing studies are mostly limited to single-modality data analysis and have not fully integrated clinical information, biomarkers, and radiomics data, affecting the clinical usability of the models. Furthermore, the lack of systematic modeling of intestinal wall-mesenteric fat interactions may lead to the omission of key pathological mechanisms.

[0007] This study aims to develop a multi-parameter radiomics model based on CTE machine learning to assess Crohn's disease activity. By integrating intestinal wall morphology features, mesenteric fat heterogeneity indicators, and clinical biochemical data from CT small bowel imaging of Crohn's disease patients, the model can non-invasively and accurately assess disease activity. The established multimodal and multiregional omics model will be validated in multiple centers, providing more reliable support for treatment planning and disease progression monitoring. Summary of the Invention

[0008] This invention aims to provide a multicenter research method for assessing Crohn's disease activity, enabling the development of treatment plans for Crohn's disease patients and monitoring disease progression, providing new treatment ideas and methods, and promoting the application of medical imaging and artificial intelligence in the field of inflammatory bowel disease.

[0009] To achieve the above objectives, the present invention provides the following method:

[0010] This invention provides a multicenter study method for assessing Crohn's disease activity:

[0011] S1: Multi-region lesion segmentation in radiomics model: The training set and validation set were drawn using the Draw function of the segment editor in the open-source medical imaging software 3D-Slicer. The most severely affected intestinal segment of each patient was selected as the region of interest (VOI), avoiding the intestinal lumen and mesentery.

[0012] The Hollow function is used to delineate the mesenteric tissue within a 5mm radius around the intestinal wall, and adjacent intestinal segments, lumens, or organs within the VOI are manually erased as the feature extraction region of interest (VOI) for establishing the radiomics model.

[0013] Two radiologists used the same tools and settings to use multiple patients suspected of having Crohn's disease as a training set. Then, 30 patients randomly selected from the training set were used for additional region of interest (VOI) segmentation to select candidate radiomics features.

[0014] S2: Feature extraction: Radiomic features of the intestinal wall and mesentery were extracted from CT images using the PyRadiomics open-source Python package to standardize the image spatial resolution;

[0015] Nonlinear intensity transformation is performed on image voxels; Gaussian Laplacian operators are used for filtering with sigma values ​​of 1, 2, and 3 mm; and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) are applied to first-order statistical and texture features. The fixed bin width is set to 5 HU to extract radiomics features of the original CT image and segmented lesions based on discrete voxel intensity values, including shape-based features, first-order features, and texture features, to describe the morphological features and internal and surface texture features of the region of interest (VOI).

[0016] S3: Feature selection: Use the Z-score method to standardize the features and calculate the mean and standard deviation of each feature vector;

[0017] S4: Model Building: A radiomics risk model was built using ten machine learning algorithms, including Support Vector Machine, Logistic Regression, Random Forest, Naive Bayes, Extreme Random Tree, Extreme Gradient Boosting, Lightweight Gradient Boosting Machine, Gradient Boosting, Adaptive Boosting, and Multilayer Perceptron.

[0018] S5: Multimodal Clinical Construction: Feature selection was performed using recursive feature elimination, ANOVA, Kruskal-Wallis test, and Relief algorithm, retaining 15 features strongly associated with the outcome; logistic regression with minimum absolute contraction and selection operator constraints was used as the classifier; 10-fold cross-validation was performed on the training dataset, hyperparameters were set based on the model's performance on the validation dataset, and the model with better performance on the test set was selected as the clinical model;

[0019] S6: Hybrid Model Construction: A clinical-radiomics feature model is constructed using the same machine learning algorithm and the aforementioned radiomics features. The ten-fold cross-validation method is used to select the model with the best performance on the cross-validation set as the clinical-radiomics feature model, and a nomogram is plotted.

[0020] S7: Model Evaluation and Comparison: The diagnostic efficacy of the clinical-radiomics model in differentiating active and remission CD in the training and testing cohorts was evaluated using the area under the receiver operating characteristic curve (AUC). Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated at the critical value for maximizing the Youden index.

[0021] S8: Statistical methods: All data analysis was performed using MedCalc and Python 3.7.12 on the OnekeyAI platform version 4.9.1. Statistical evaluation was performed using statsmodels, and radiomics features were extracted using PyRadiomics. Machine learning implementations, including support vector machines, were performed using Scikit-learn.

[0022] Preferably, the delineation training set and validation set include a complete diseased intestinal segment from the beginning to the end of the lesion. If a patient in remission has no obvious diseased intestinal segment, the ileocecal region is selected as the delineation object.

[0023] Preferably, the steps of selecting candidate radiomics features by having the two radiologists use the same tools and settings to use multiple patients suspected of having Crohn's disease as a training set, and then performing additional region of interest (VOI) segmentation on 30 randomly selected patients from the training set, include: the two radiologists using the same tools and settings to use multiple patients suspected of having Crohn's disease as a training set, and then performing additional region of interest (VOI) segmentation on 30 randomly selected patients from the training set; the time interval between two image readings by the two radiologists being more than 6 months; evaluating the consistency of the delineation results by the two radiologists using intra-group and inter-group correlation coefficients (ICCs); and selecting features whose intra-group and inter-group correlation coefficients (ICCs) are greater than 0.75 as candidate radiomics features.

[0024] Preferably, the step of using the PyRadiomics open-source Python package to extract radiomics features of the intestinal wall and mesentery from CT images to standardize the image spatial resolution includes: automatically extracting 1834 Radiomics features from CT images using the PyRadiomics open-source Python package, and resampling the voxel size of the CTE image and its corresponding 3D-VOI to 1×1×1mm³ (x,y,z) isotropic voxels using a linear interpolation algorithm to standardize the image spatial resolution.

[0025] Preferably, the texture features include: a gray-level co-occurrence matrix, a gray-level dependency matrix, a gray-level running length matrix, a gray-level size region matrix, and a neighboring gray-level difference matrix.

[0026] Preferably, the step of standardizing the features using the Z-score method and calculating the mean and standard deviation of each feature vector includes: subtracting the mean from each feature vector and dividing by the standard deviation to perform standardization; each vector has a zero mean and a unit standard deviation; considering the high-dimensionality of the feature space, an association matrix is ​​constructed by calculating the similarity of each pair of features; if the Pearson correlation coefficient (PCC) of a feature pair is greater than 0.9, one of the features is deleted; the maximum-minimum redundancy algorithm is used to select features, and then the minimum absolute shrinkage and selection operator algorithm is used for feature selection.

[0027] Preferably, after constructing the radiomics risk model, the method further includes: calculating the radiomics score for each patient by linearly combining and weighting the selected features; and determining the hyperparameters of the radiomics risk model based on its performance on the validation dataset.

[0028] Preferably, after calculating accuracy, sensitivity, specificity, positive predictive value, and negative predictive value at the critical value where the Youden index value is maximized, the method further includes: estimating a 95% confidence interval using a bootstrap method with 1000 samples; evaluating the consistency between the predicted probability of the clinical-radiomics model and the actual results using a calibration curve, and verifying the reliability of the clinical-radiomics feature model using the Hosmer-Lemeshow goodness-of-fit test; evaluating the calibration capability of the clinical-radiomics feature model using a calibration curve; and comparing the AUC values ​​of different clinical-radiomics feature models using the DeLong test.

[0029] Preferably, after the step of implementing machine learning, including support vector machines, using Scikit-learn, the method further includes: performing a normality test on clinical characteristics using the Shapiro-Wilk test; representing continuous variables as mean and variance, and performing component comparisons using a t-test or Mann-Whitney U test based on their distribution characteristics; presenting categorical variables as frequencies and percentages, and performing inter-group comparisons using a chi-square test or Fisher's exact test.

[0030] Preferably, after presenting the categorical variables in the form of frequencies and percentages, and performing inter-group comparisons using the chi-square test or Fisher's exact test, the method further includes: performing univariate analysis on radiological characteristics with good consistency within and between observers, where the intra- and inter-group correlation coefficients (ICCs) are >0.75, and a two-sided P <0.05 is considered statistically significant; and using receiver operating characteristic (ROC) curves to evaluate the diagnostic efficacy of the model, calculating the area under the curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value to quantify the diagnostic efficacy.

[0031] The beneficial effects of this invention are as follows: This invention constructs a clinical prediction model using machine learning algorithms based on the demographic characteristics, clinical manifestations, laboratory tests, and imaging data of CD patients. Subsequently, radiomics features are extracted from the intestinal wall and mesenteric fat region of CT enterography (CTE) images, and clinical and radiomics features are fused to construct multiple machine learning models to assess the disease activity of CD patients. The results show that the fusion model achieved an AUC of 0.928 on the test set, demonstrating high diagnostic efficacy. Furthermore, this study followed strict variable selection and cross-validation strategies during feature selection and modeling to ensure the accuracy and generalization ability of the model.

[0032] In terms of imaging findings, statistically significant differences were observed in lymphadenopathy, abdominal abscess, intestinal wall thickness, target sign, comb-like sign, blurred fat spaces, number of affected intestinal segments, enterocutaneous fistula, cellulitis, ileocecal involvement, arterial phase CT values, portal venous phase CT values, and ascites. These imaging findings provide a wealth of information, and when combined with laboratory and clinical indicators, they can provide a more comprehensive assessment of Crohn's disease activity.

[0033] In the selection of features and the construction of clinical models, the preprocessing of the feature space is crucial. This study did not standardize or normalize the feature matrix because the measurement units and distribution characteristics of each feature may differ, and standardization or normalization may affect the interpretability of the model in some cases. To reduce redundant features, the Pearson correlation coefficient (PCC) was first calculated for each pair of features. When the PCC value is greater than 0.9, it indicates a high correlation between these features; therefore, one feature was removed, reducing the dimensionality of the feature space and avoiding bias caused by feature redundancy. High correlation between features can lead to overfitting or excessive reliance on certain features, thus affecting the model's stability and performance. Therefore, this preprocessing method ensures that the final selected features have good independence and lays the foundation for subsequent model training. Through recursive feature elimination (RFE) and LASSO regularization, nine important features were selected from the original features, and the final model was constructed. These features include one clinical and demographic characteristic, three laboratory indicators, and five imaging characteristics: diarrhea, qualitative CRP, qualitative ESR, qualitative albumin, intestinal wall thickness, target sign, comb sign, blurred fat spaces, and cellulitis. The selection of these features is of significant clinical importance and improves the accuracy of disease activity assessment.

[0034] The radiomics model incorporated 11 features, 6 of which originated from the intestinal wall itself and 5 from the mesenteric adipose tissue. Features from the intestinal wall primarily manifested as parameters related to the complexity, gray-level uniformity, and contrast of the intestinal tissue, reflecting structural changes, regional complexity, and lesion diversity. Zone Entropy reflects the complexity of the image's gray-level distribution. In active CD, the disordered tissue structure, inflammatory cell infiltration, and fibrosis in the lesion area lead to uneven image gray-level distribution, resulting in increased Zone Entropy.

[0035] On the training dataset of the radiomics model, the hyperparameters of the model were tuned using 10-fold cross-validation to ensure the model's generalization ability. Cross-validation involves dividing the dataset into 10 subsets, using 9 subsets for training in turn, and using the remaining subset for testing. This method effectively reduces bias caused by imbalanced training data or sample partitioning issues. After cross-validation tuning, a model containing 11 features was finally selected. The core value of the machine learning framework used in this study lies in its ability to handle high-dimensional heterogeneous data. Crohn's disease activity is driven by multiple factors, including local intestinal wall inflammation, mesenteric fat remodeling, and systemic immune responses. Traditional statistical models struggle to effectively integrate such complex interactions. In contrast, machine learning algorithms, through nonlinear modeling (such as SVM kernel tricks and decision boundaries of tree models) and feature importance ranking, can capture potential association patterns between radiomics features and clinical variables. For example, Extreme Gradient Boosting (XGBoost) and Random Forest (RF) significantly reduce the overfitting risk of a single decision tree by integrating multiple weak classification trees, while L1 regularization of LASSO regression achieves sparse modeling in a high-dimensional feature space, which is highly compatible with the characteristics of radiomics data—high dimensionality and limited sample size. On the training set, the model achieved a high AUC (0.979) and accuracy (0.938), indicating high diagnostic efficacy in assessing Crohn's disease activity. The AUC and accuracy on the test dataset were 0.926 and 0.857, respectively, slightly lower than the training set performance, but still demonstrating good stability. The model's consistent performance across different datasets demonstrates its strong generalization ability and adaptability, enabling its application in diverse clinical settings.

[0036] The hybrid model, which fused the radiomics model and the clinical model, achieved an AUC of 0.928 and an accuracy of 0.875. The Delong test showed a statistically significant difference between the hybrid model and the clinical model. This study constructed a radiomics model from intestinal wall and mesenteric adipose tissue. By combining clinical data, laboratory tests, and radiomics data, it simulated real-world clinical scenarios as closely as possible. The fused model achieved an AUC of 0.928 on the external validation set, demonstrating its good generalization ability.

[0037] In the research design and model construction process, the present invention adopted the following three standardization strategies: (1) Image spatial standardization: all images were resampled into isotropic voxels (1×1×1mm³) using linear interpolation algorithm before feature extraction to eliminate the spatial resolution inconsistency caused by layer thickness differences in different centers; (2) Intensity normalization processing: the image signal intensity was normalized to a uniform gray range (±5HU) to reduce the difference in the distribution of scanning signal intensity between different devices; (3) Feature stability screening: all extracted radiomics features were screened by ICC (intra-group / inter-group consistency coefficient), and only high repeatability features with ICC>0.75 were retained to ensure that the modeling variables have cross-center stability. Attached Figure Description

[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0039] Figure 1 A flowchart illustrating a multicenter study method for assessing Crohn's disease activity, provided as an embodiment of the present invention;

[0040] Figure 2 A flowchart illustrating the inclusion and exclusion criteria for the population provided in this embodiment of the invention;

[0041] Figure 3 This is a schematic diagram of segmenting Crohn's disease lesions on a CT image provided by an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the clinical-radiomics feature model construction process provided in an embodiment of the present invention;

[0043] Figure 5 ROC curves for the clinical model training and testing sets provided in this embodiment of the invention;

[0044] Figure 6 A schematic diagram of the correlation matrix of the features of each clinical model provided in the embodiments of the present invention;

[0045] Figure 7 The following is a screening of radiomic features of diseased intestinal wall and mesenteric fat using regression analysis with minimum absolute contraction and selection operator provided in this embodiment of the invention: (a) binomial deviation curve, with the regularization parameter (λ) selected as 0.0168 using ten-fold cross-validation with minimum criterion; (b) LASSO coefficient path diagram;

[0046] Figure 8The feature coefficient histogram provided in this embodiment of the invention;

[0047] Figure 9 ROC curves of ten machine learning models in the training set (a) and validation set (b) provided for embodiments of the present invention. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0050] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0051] The application of machine learning in radiomics analysis is rapidly developing. Through feature selection and pattern recognition, machine learning can optimize the extraction of image biomarkers and improve the generalization ability of models. In the field of Crohn's disease activity assessment, machine learning algorithms can automatically extract high-dimensional image features (such as intestinal wall thickness, enhancement level, texture heterogeneity, etc.) and combine them with clinical data and biomarkers to construct multi-dimensional predictive models. For example, supervised learning algorithms (such as random forests and support vector machines) have been used to identify radiomics features highly correlated with endoscopic activity from CTE images and optimize model generalization ability through feature selection. Wasnik et al. constructed an automatic detection model for Crohn's disease lesions in CTE (AUC 0.94) by integrating radiomics and deep learning features. Its consistency with radiologist assessment was as high as Kappa=0.82, and the analysis efficiency was improved by more than 10 times, demonstrating the feasibility of machine learning in standardized image assessment. Integrated learning frameworks, by fusing radiomics features, laboratory indicators, and clinical scores, can establish more comprehensive assessment systems. For example, jointly modeling CTE imaging features with dynamic changes in FCP has been shown to significantly improve the predictive accuracy of mucosal healing status. However, existing studies are mostly limited to single-modality data analysis and have not fully integrated clinical information, biomarkers, and radiomics data, affecting the clinical usability of the models. Furthermore, the lack of systematic modeling of intestinal wall-mesenteric fat interactions may lead to the omission of key pathological mechanisms.

[0052] This study aims to develop a multi-parameter radiomics model based on CTE machine learning to assess Crohn's disease activity. By integrating intestinal wall morphology features, mesenteric fat heterogeneity indicators, and clinical biochemical data from CT small bowel imaging of Crohn's disease patients, the model can non-invasively and accurately assess disease activity. The established multimodal and multiregional omics model will be validated in multiple centers, providing more reliable support for treatment planning and disease progression monitoring.

[0053] This invention aims to provide a multicenter research method for assessing Crohn's disease activity, enabling the development of treatment plans and monitoring of disease progression in Crohn's disease patients, providing new treatment ideas and methods, and contributing to the application of medical imaging and artificial intelligence in the field of inflammatory bowel disease.

[0054] like Figure 1 As shown in the figure, a specific embodiment of the present invention provides a multicenter study method for assessing Crohn's disease activity, comprising the following steps:

[0055] S1: Multi-region lesion segmentation in radiomics model: The training and validation sets were drawn using the Draw function of the segment editor in the open-source medical imaging software 3D-Slicer. The most severely affected intestinal segment in each patient was selected as the region of interest (VOI), avoiding the intestinal lumen and mesentery. The Hollow function was used to draw the mesenteric tissue within a 5mm radius around the intestinal wall, and adjacent intestinal segments, lumens, or organs within the VOI were manually erased as features for the VOI of the radiomics model. Two radiologists used the same tools and settings to use multiple patients suspected of having Crohn's disease as the training set. Additional VOI segmentation was performed on 30 patients randomly selected from the training set to select candidate radiomics features.

[0056] In this embodiment of the invention, the delineation of the training and validation sets includes the complete diseased intestinal segment from the beginning to the end of the lesion. If a patient in remission has no obvious diseased intestinal segment, the ileocecal region is selected as the delineation object. Two radiologists use the same tools and settings to use multiple patients suspected of Crohn's disease as the training set. An additional region of interest (VOI) segmentation is performed on 30 randomly selected patients from the training set. The steps for selecting candidate radiomics features include: two radiologists using the same tools and settings to use multiple patients suspected of Crohn's disease as the training set; an additional region of interest (VOI) segmentation is performed on 30 randomly selected patients from the training set; the time interval between two radiologists' image readings exceeds 6 months; the consistency of the delineation results between the two radiologists is evaluated by the intra- and inter-group correlation coefficients (ICCs); features with intra- and inter-group correlation coefficients (ICCs) greater than 0.75 are selected as candidate radiomics features.

[0057] S2: Feature Extraction: Radiomic features of the intestinal wall and mesentery were extracted from CT images using the PyRadiomics open-source Python package to normalize the image spatial resolution; nonlinear intensity transformations were performed on image voxels; Gaussian Laplacian operators were used for filtering with sigma values ​​of 1, 2, and 3 mm; and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) were applied to first-order statistical and texture features; the fixed bin width was set to 5 HU to extract radiomic features of the original CT images and segmented lesions based on discrete voxel intensity values, including shape-based features, first-order features, and texture features, to describe the morphological features and internal and surface texture features of the region of interest (VOI).

[0058] In this embodiment of the invention, the step of using the PyRadiomics open-source Python package to extract radiomics features of the intestinal wall and mesentery from CT images to standardize the image spatial resolution includes: automatically extracting 1834 Radiomics features from CT images using the PyRadiomics open-source Python package; and resampling the voxel size of the CTE image and its corresponding 3D-VOI to 1×1×1 mm³ (x,y,z) isotropic voxels using a linear interpolation algorithm to standardize the image spatial resolution: gray-level co-occurrence matrix, gray-level dependency matrix, gray-level run length matrix, gray-level size region matrix, and neighboring gray-level difference matrix.

[0059] S3: Feature selection: Use the Z-score method to standardize the features and calculate the mean and standard deviation of each feature vector;

[0060] In this embodiment of the invention, each feature vector is standardized by subtracting its mean and dividing by its standard deviation. Each vector has a zero mean and a unit standard deviation. To take advantage of the high-dimensionality of the feature space, an association matrix is ​​constructed by calculating the similarity of each pair of features. If the Pearson correlation coefficient (PCC) of a feature pair is greater than 0.9, one of the features is deleted. The maximum-minimum redundancy algorithm is used to select features, and then the minimum absolute shrinkage and selection operator algorithm is used for feature selection.

[0061] S4: Model Building: A radiomics risk model was built using ten machine learning algorithms, including Support Vector Machine, Logistic Regression, Random Forest, Naive Bayes, Extreme Random Tree, Extreme Gradient Boosting, Lightweight Gradient Boosting Machine, Gradient Boosting, Adaptive Boosting, and Multilayer Perceptron.

[0062] In this embodiment of the invention, after constructing the radiomics risk model, the method further includes: calculating the radiomics score for each patient by linearly combining and weighting the selected features; and determining the hyperparameters of the radiomics risk model based on its performance on the validation dataset.

[0063] S5: Multimodal Clinical Construction: Feature selection was performed using recursive feature elimination, ANOVA, Kruskal-Wallis test, and Relief algorithm, retaining 15 features strongly associated with the outcome; logistic regression with minimum absolute contraction and selection operator constraints was used as the classifier; 10-fold cross-validation was performed on the training dataset, hyperparameters were set based on the model's performance on the validation dataset, and the model with better performance on the test set was selected as the clinical model.

[0064] S6: Hybrid Model Construction: A clinical-radiomics feature model is constructed using the same machine learning algorithm and radiomics features. The 10-fold cross-validation method is used to select the model with the best performance on the cross-validation set as the clinical-radiomics feature model, and a nomogram is plotted.

[0065] S7: Model Evaluation and Comparison: The diagnostic efficacy of the clinical-radiomics model in differentiating active and remission CD in the training and testing cohorts was evaluated using the area under the receiver operating characteristic curve (AUC). Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated at the cutoff value for maximizing the Youden index.

[0066] In this embodiment of the invention, after calculating the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value at the critical value where the Youden index is maximized, the method further includes: estimating a 95% confidence interval using a bootstrap method with 1000 samples; evaluating the consistency between the predicted probability of the clinical-radiomics model and the actual results using a calibration curve, and verifying the reliability of the clinical-radiomics feature model using the Hosmer-Lemeshow goodness-of-fit test; evaluating the calibration capability of the clinical-radiomics feature model using a calibration curve; and comparing the AUC values ​​of different clinical-radiomics feature models using the DeLong test.

[0067] S8: Statistical methods: All data analysis was performed using MedCalc and Python 3.7.12 on the OnekeyAI platform version 4.9.1. Statistical evaluation was performed using statsmodels, and radiomics features were extracted using PyRadiomics. Machine learning implementations, including support vector machines, were performed using Scikit-learn.

[0068] In this embodiment of the invention, after the step of implementing machine learning, including support vector machines, using Scikit-learn, the method further includes: performing a normality test on clinical characteristics using the Shapiro-Wilk test; representing continuous variables as mean and variance, and performing component comparisons using a t-test or Mann-Whitney U test based on their distribution characteristics; presenting categorical variables as frequencies and percentages, and performing inter-group comparisons using a chi-square test or Fisher's exact test; after the step of presenting categorical variables as frequencies and percentages, and performing inter-group comparisons using a chi-square test or Fisher's exact test, the method further includes: performing univariate analysis on radiological characteristics with good consistency within and between observers, where intra- and inter-group correlation coefficients (ICCs) > 0.75, and considering two-sided P < 0.05 as statistically significant; evaluating the diagnostic efficacy of the model using receiver operating characteristic (ROC) curves, and quantifying diagnostic efficacy by calculating the area under the curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. Specific Implementation

[0069] 1. Research Subjects

[0070] This retrospective study has been approved by the Medical Ethics Review Committee of our hospital, exempting it from informed consent. We retrospectively collected data from 252 patients suspected of having Crohn's disease at Ruijin Hospital (Center 1), affiliated with Shanghai Jiao Tong University School of Medicine, from 2017 to 2022; 68 patients suspected of having Crohn's disease at Kunshan Traditional Chinese Medicine Hospital (Center 2), from 2017 to 2023; and 47 patients suspected of having Crohn's disease at the First Affiliated Hospital of Soochow University (Center 3), from 2017 to 2022. Patients from Center 1 were randomly assigned to the training set and internal test set in a 7:3 ratio. Centers 2 and 3 served as the external test set. The patient recruitment process is as follows: Figure 2 .

[0071] The inclusion criteria were as follows: (1) patients diagnosed with Crohn's disease according to guidelines; (2) all patients underwent CTE and colonoscopy, with a time difference of no more than one week; and (3) complete clinical history, laboratory test data, etc. The exclusion criteria were as follows: (1) incomplete data; (2) other intestinal diseases; (3) age less than 18 years; and (4) artifacts in CTE images that affected lesion delineation. Ileocolonic examination and histological identification of characteristic manifestations are the gold standard for the diagnosis of Crohn's disease. Therefore, the simplified endoscopic score (SES-CD) was used as the grouping standard to divide patients into two groups: active phase and remission phase.

[0072] 2. Patient Information

[0073] The following information was collected from the patient: (1) General information: age, gender, height, weight, BMI, clinical symptoms, duration of symptoms, and surgical history; (2) Biochemical information: complete blood count, stool routine, fecal occult blood, fecal calprotectin, C-reactive protein, erythrocyte sedimentation rate, total protein, and albumin; (3) Imaging information: the patient's CT small bowel imaging was evaluated according to existing literature standards, and the intestinal wall thickness, number of affected intestinal segments, enhancement mode, increased mesenteric fat density, cellulitis, abdominal abscess, lymphadenopathy, intestinal fistula, intestinal perforation, intestinal obstruction, and pseudodiverticulum were recorded.

[0074] 3. CT small bowel contrast imaging data acquisition

[0075] All patients underwent standardized bowel preparation before the examination. Patients fasted on the morning of the examination and drank 1500-2000 ml of 2.5% isotonic mannitol solution in divided doses 45-50 minutes before scanning. Ruijin Hospital affiliated with Shanghai Jiao Tong University School of Medicine used Siemens Somatom Drive (No.1), Philips Iqon Spectral (No.2), Toshiba Aquilion ONE-TSX-301C (No.3), and United Imaging uCT760 (No.4) scanners. The First Affiliated Hospital of Soochow University used Siemens second-generation dual-source Somatom Definition Flash CT (No.5), Philips Iqon Spectral (No.6), and GE Revolution CT (No.7) scanners. Kunshan Traditional Chinese Medicine Hospital used Siemens second-generation dual-source Somatom Definition Flash CT (No.8) scanners. Scanning parameters are shown in Table 1. 90-100 ml of non-ionic contrast agent was injected via the antecubital vein at a rate of 3-4 ml / s. The arterial phase was scanned automatically (the region of interest was located in the abdominal aorta, and the threshold was set to 100 HU), and the portal venous phase was scanned after a delay of 28 seconds.

[0076] Table 1 CT scan parameters for each center

[0077] machine No.1 No.2 No. 3 No. 4 No. 5 No. 6 No. 7 No. 8 Tube voltage (Kev) 100 120 120 120 120 120 120 120 Tube current (mA) 183 103 100 155 163 129 355 210 pitch 0.8 0.921 0.992 1.1875 0.6 0.990 0.992 0.6 Layer thickness (mm) 1 1 1 1 1.5 1 1 1 Interlayer spacing (mm) 0.6 1 0.625 1 1.5 1 1 1

[0078] 4. Establishment of radiomics models

[0079] Data from Center 1 was divided into a training set and an internal test set in a 7:3 ratio, resulting in 128 cases (75 / 53 = active phase / remission phase) as the training dataset and 56 cases (33 / 23 = active phase / remission phase) as the test dataset. An additional 38 CD patients recruited from Center 2 and 30 from Center 3 served as the external validation set. Similar to other intestinal radiomics studies, all patients underwent radiomics analysis using portal venous images. The enhancement of the intestinal wall and mesenteric vessels tends to stabilize during the portal venous phase, reducing the uneven contrast agent concentration between intestinal segments caused by hemodynamic differences during the arterial phase. The specific steps are as follows:

[0080] (1) Multi-region lesion segmentation: A radiologist (with 12 years of experience in abdominal imaging diagnosis) used the Draw function of the segment editor in the open-source medical imaging software 3D-Slicer (version 4.11.2; https: / / www.slicer.org / ) to delineate the training and validation sets, selecting the most severely affected intestinal segment for each patient as the volume of interest (VOI), avoiding the intestinal lumen and mesentery. The complete lesion segment from the beginning to the end of the lesion was delineated. If there was no obvious lesion segment in patients in remission, the ileocecal region was selected as the delineation object. The Hollow function was used to delineate the mesenteric tissue within a 5mm radius around the intestinal wall, and adjacent intestinal segments, intestinal lumens, or organs within the VOI were manually erased. Straight small vessels are considered an important imaging feature of Crohn's disease during active phase, so straight small vessels within the VOI were preserved during erasure. That is, for each patient, the most severely affected intestinal segment and the adjacent mesentery were delineated as the feature extraction VOI for establishing the radiomics model. To investigate the inter- and intra-observer reproducibility of radiomics features, two radiologists performed additional VOI segmentation on 30 patients randomly selected from the training set recruited from Center 1, using the same tools and settings. One radiologist had an interval of more than 6 months between two image readings. The consistency of the two radiologists' delineation results was evaluated using intraclass correlation coefficients (ICCs). Only features with ICCs greater than 0.75 were selected as candidate radiomics features. The segmentation process is as follows: Figure 3 As shown.

[0081] like Figure 3 As shown, Figure 3 This is a schematic diagram of segmenting Crohn's disease lesions on CT images provided in an embodiment of the present invention; the most severely affected intestinal segment of the patient is selected for delineation, and a three-dimensional model of the active lesion is obtained by dividing it layer by layer, and the mesenteric fat within a 5mm radius around the intestine is automatically obtained using the Hollow function, while adjacent intestinal segments, major blood vessels, muscles or organs are manually erased.

[0082] (2) Feature Extraction: Radiomics features of the intestinal wall and mesentery were extracted from CT images using the PyRadiomics open-source Python package. This module automatically extracted 1834 Radiomics features. To correct for voxel resolution variations between different CT scanners, a linear interpolation algorithm was used to resample the voxel size of the CTE images and their corresponding 3D-VOIs to 1×1×1 mm³ (x,y,z) isotropic voxels to normalize the image spatial resolution. To obtain high-throughput features, nonlinear intensity transformations (square, square root, logarithmic, and exponential) were applied to the image voxels; the Laplacian Gaussian operator (LoG) was filtered using sigma values ​​of 1, 2, and 3 mm; and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) were applied for first-order statistical and texture features. The fixed bin width was set to 5 HU to discrete voxel intensity values. Then, radiomics features of the original CT images and segmented lesions were extracted, including 1) shape-based features (2D shape-based, 3D shape-based); 2) first-order features; and 3) texture features: gray-level co-occurrence matrix, gray-level dependency matrix, gray-level run length matrix, gray-level size region matrix, and neighboring gray-level difference matrix, describing the morphological features and internal and surface texture features of the VOI. Ultimately, this included 14 morphological features, 18 first-order statistical features, and 73 texture features (22 gray-level co-occurrence matrices (GLCM), 14 gray-level dependency matrices (GLDM), 16 gray-level size region matrices (GLSZM), 16 gray-level run length matrices (GLRLM), and 5 neighboring gray-level difference matrices (NGTDM)).

[0083] (3) Feature Selection: Features were standardized using the Z-score method. The mean and standard deviation of each feature vector were calculated. The mean was subtracted from each feature vector and the standard deviation was divided. After standardization, each vector had a zero mean and a unit standard deviation. To address the high dimensionality of the feature space, an association matrix was constructed by calculating the similarity of each pair of features. If the Pearson correlation coefficient (PCC) of a feature pair was greater than 0.9, one of the features was deleted. After this process, the dimensionality of the feature space was reduced, and each feature became independent of the others.

[0084] Before building the model, the Maximum Relevance Minimum Redundancy (MRMR) algorithm is used to select features. MRMR is a commonly used feature selection method that filters features based on their correlation with the target variable to select features that are highly correlated with the target variable and have low redundancy. Then, the least absolute shrinkage and selection operator (LASSO) algorithm is used for rigorous feature selection. The LASSO feature selection method is a penalty-based method that adds an L1 regularization term to the objective function, compressing some regression coefficients to zero, thereby achieving feature selection.

[0085] (4) Model Building: A radiomics risk model was constructed using ten machine learning algorithms: Support Vector Machine, Logistic Regression, Random Forest, Naive Bayes, Extremely Random Tree, Extreme Gradient Boosting, Lightweight Gradient Boosting Machine, Gradient Boosting, Adaptive Boosting, and Multilayer Perceptron. A radiomics score (Rad-score) for each patient was calculated by linearly combining and weighting selected features. To determine the model's hyperparameters (e.g., the number of features), 10-fold cross-validation was applied to the training dataset. The hyperparameter settings were determined based on the model's performance on the validation dataset.

[0086] 5. Construction of multimodal clinical and hybrid models

[0087] Clinical model construction: The clinical features included in this study were clinical manifestations, laboratory tests, and CTE findings. Since the grouping of the study was based on the Simplified Endoscopic Score (SES-CD), colonoscopy and pathological findings were not included in the model construction to prevent bias.

[0088] No normalization method was used on the feature matrix. Due to the high dimensionality of the feature space, the similarity of each pair of features was compared. If the Pearson correlation coefficient (PCC) of a feature pair was greater than 0.9, one of them was deleted. This process reduced the dimensionality of the feature space and ensured that each feature was independent. Before building the model, recursive feature elimination (RFE), analysis of variance (ANOVA), Kruskal-Wallis (KW) test, and the Relief algorithm were used for feature selection, retaining 15 features that were strongly associated with the outcome. Logistic regression (LR) with least absolute shrinkage and selection operator (LASSO) constraints was used as the classifier. Logistic regression with LASSO constraints is a linear classifier based on logistic regression. The L1 criterion was added to the final loss function, and the weights were constrained to make the features sparse. To determine the model's hyperparameters (such as the number of features), 10-fold cross-validation was performed on the training dataset. The hyperparameters are set based on the model's performance on the validation dataset, and the model that performs better on the test set is selected as the final model.

[0089] Hybrid Model Construction: A clinical-radiomics feature model was constructed using the same machine learning algorithm and the same radiomics features. During training, ten-fold cross-validation was employed, and the model with the best performance on the cross-validation set was selected as the final model. A noctilinear plot was then generated. The model construction process is as follows: Figure 4 ;

[0090] like Figure 4 As shown, Figure 4This is a schematic diagram illustrating the construction process of a clinical-radiomics feature model provided in this embodiment of the invention. The diseased intestinal segment is three-dimensionally segmented on a CT small bowel imaging image. The Hollow function based on 3D-Slicer software is used to obtain mesenteric adipose tissue extending 5 mm outwards from the diseased intestinal wall. Radiomic features of the intestinal wall and mesenteric adipose tissue are extracted separately. Features are screened using LASSO regression, and a predictive radiomics model is built based on these features using machine learning methods. Clinical information such as medical history, laboratory tests, and CTE imaging manifestations is used for feature screening using methods such as RFE. A clinical model is built using LASSO regression. A fusion model is constructed by integrating omics features and clinical features. Receiver operating characteristic (ROC) curves and calibration curves are used to evaluate the model's effectiveness. RFE is recursive feature elimination, ANOVA is multivariate analysis of variance, and LASSO is the minimum absolute contraction and selection operator.

[0091] 6. Model Evaluation and Comparison

[0092] The diagnostic efficacy of the clinical-radiomics model in differentiating active and remission CD in both training and testing cohorts was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). Accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (PPV) were calculated at the cutoff value maximizing the Youden index. A 95% confidence interval was estimated using a bootstrap method with 1000 samples. The consistency between the predicted probabilities of the clinical-radiomics model and actual results was assessed using a calibration curve, and the reliability of the model was validated using the Hosmer-Lemeshow goodness-of-fit test. The calibration capability of the model was assessed using a calibration curve. The DeLong test was used to compare the AUC values ​​of different models.

[0093] 7. Statistical methods

[0094] All data analysis was performed on the OnekeyAI platform version 4.9.1 using MedCalc (version 22.026) and Python 3.7.12. Statistical evaluation was performed using statsmodels (version 0.13.2), and radiomics features were extracted using PyRadiomics (version 3.0.1). Machine learning implementations, including Support Vector Machines (SVM), were performed using Scikit-learn (version 1.0.2).

[0095] Normality of clinical features was tested using the Shapiro-Wilk test. Continuous variables were expressed as mean and variance (SD), and component comparisons were performed using t-tests or Mann-Whitney U tests based on their distribution characteristics. Categorical variables were presented as frequencies and percentages, and inter-group comparisons were performed using chi-square tests or Fisher's exact tests. Univariate analysis was performed on radiological features with good within-observer and between-observer agreement (ICC > 0.75). Two-sided p < 0.05 was considered statistically significant. The diagnostic efficacy of the model was assessed using receiver operating characteristic (ROC) curves, and the area under the curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated to quantify diagnostic efficacy.

[0096] Results: 1. Patient information:

[0097] A total of 252 cases were enrolled, including 128 cases in the training set, 56 cases in the internal validation set, and 68 cases in the external validation set. The demographic and clinical characteristics of the training and validation sets are shown in Table 2. There were 187 males (74.21%) and 65 females (25.79%). The mean age and BMI of all patients were 34.62±11.05 years and 20.26±3.10 kg / m², respectively (normal range: 18.5-23.9 kg / m²). In the training and validation sets, there were 132 patients in the active phase (52.4%) and 120 patients in remission (47.6%). Except for weight and BMI, there were no significant differences in demographic and clinical characteristics between the two cohorts (all P>0.05). Perhaps because the hospital's gastroenterology department has a certain authority in the diagnosis and treatment of Crohn's disease, the proportion of patients with low weight and poor nutritional status was slightly higher. However, since the model construction parameters were not statistically significant, they did not affect the results.

[0098] Table 2. Demographic and Clinical Characteristics

[0099]

[0100] Crohn's disease exhibits high heterogeneity in its clinical presentation and course. To help clinicians better describe and classify Crohn's disease, the Montreal Classification was developed. This classification method categorizes Crohn's disease into different subtypes based on disease course, affected sites, and clinical manifestations, thereby helping physicians predict disease progression, management strategies, and prognosis. It is now routinely used in clinical practice. This classification system is based on three main parameters: 1. Age of onset: A1: <16 years (childhood type), A2: 16-40 years (adolescent type), A3: >40 years (adult type); age group has a certain influence on the clinical presentation and disease progression. Generally, early-onset Crohn's disease (A1) typically has a longer course, more severe symptoms, and is prone to complications such as intestinal stricture and perforation. 2. Lesion Location: L1: Ileal type (ileum only), L2: Colonic type (colon only), L3: Ileocolic type (both ileum and colon are involved), and L4: Upper Gastrointestinal type (involvement of the upper gastrointestinal tract, such as the stomach, duodenum, and esophagus; usually rare). This classification helps clinicians understand the specific areas affected by the disease. The ileal type (L1) and colonic type (L2) are the most common types, while the ileocolic type (L3) may have more complications. 3. Disease Behavior: B1: Inflammatory type (non-strictive, non-perforated), mainly characterized by chronic inflammation, with damage to the intestinal mucosa but without stricture or perforation. B2: Stricture type (with intestinal stricture), with fibrosis and narrowing of the intestine, leading to limited or obstructed intestinal lumen, which may require surgical intervention. B3: Penetrating type (with fistula or abscess), the lesion causes intestinal perforation, which may result in fistula formation or abscess, usually requiring surgical intervention or treatment. This classification helps doctors predict the course of the disease and treatment strategies based on the pathological characteristics and clinical manifestations of Crohn's disease. Generally, patients with inflammatory type (B1) have a better prognosis, while those with stenotic type (B2) and penetrating type (B3) require more extensive treatment intervention. In addition, the Montreal classification has an additional category for assessing perianal lesions: P1 indicates the presence of perianal lesions, manifested as fistulas, abscesses, anal fissures, etc., around the anus; P0 indicates the absence of perianal lesions. The presence of perianal lesions significantly impacts a patient's quality of life and can have a significant influence on disease progression.

[0101] The Montreal classification of patients in this study is shown in Table 3. There were no statistically significant differences between the training set and the validation set in the three Montreal classifications.

[0102] Table 3 Montreal Classification of Crohn's Disease Patients

[0103] 2. Evaluation and comparison of model diagnostic efficacy

[0104] 2.1 Clinical Model

[0105] Univariate logistic regression analysis showed that within the training set, there were statistically significant differences in clinical manifestations and demographics regarding abdominal pain, diarrhea, gastrointestinal bleeding, weight loss, and abdominal mass. (See Table 4)

[0106] Table 4 Clinical Manifestations and Demographics

[0107]

[0108] In terms of laboratory tests, there were statistically significant differences in erythrocyte quantitative and qualitative tests, platelet qualitative and quantitative tests, CRP qualitative tests, erythrocyte sedimentation rate (ESR) qualitative and quantitative tests, total protein quantitative tests, albumin quantitative tests, albumin qualitative tests, and fecal occult blood tests. (See Table 5)

[0109] Table 5 Laboratory Tests

[0110]

[0111] In terms of imaging manifestations, there were statistically significant differences in lymphadenopathy, abdominal abscess, intestinal wall thickness, target sign, comb sign, blurred fat spaces, number of affected intestinal segments, intestinal fistula, cellulitis, ileocecal involvement, arterial phase CT value, portal venous phase CT value, and ascites.

[0112] The study found that the clinical model based on eight features achieved the highest AUC on the validation dataset. The area under the curve (AUC) and accuracy reached 0.855 and 0.786, respectively. The eight features ultimately selected were diarrhea, qualitative CRP, number of affected intestinal segments, blurred fat spaces, intestinal wall thickness, target sign, comb-like sign, and qualitative albumin, including one clinical and demographic feature, two laboratory tests, and five imaging findings. The ROC curves for the training and testing sets are shown below. Figure 5 As shown. The correlation matrix between the features is as follows. Figure 6 The model has three core features (target sign, blurred fat space, intestinal wall thickness, correlation coefficient > 0.68) and three secondary features (CRP qualitative, albumin qualitative, number of affected intestinal segments, correlation coefficient 0.3-0.5).

[0113] Figure 5 ROC curves for the clinical model training and testing sets provided in this embodiment of the invention;

[0114] like Figure 6 As shown, Figure 6The diagram shows the correlation matrix of the clinical model features provided in the embodiments of the present invention, revealing the correlation strength between each variable and Crohn's disease activity (target variable, label), where C = diarrhea, X = CRP qualitative, AD = albumin qualitative, AY = intestinal wall thickness, AZ = target sign, BB = blurred fat space, and BC = number of affected intestinal segments.

[0115] 2.2 Radiomics Model

[0116] Based on imaging data of the intestinal wall and mesenteric fat region for each patient, a total of 3668 standardized radiomics features were extracted. First, radiomics features with ICCs > 0.75 in the training cohort were retained. Then, radiomics features with P < 0.05 were retained through univariate analysis. Next, radiomics features with Pearson correlation coefficients ≥ 0.9 were removed. Finally, LASSO regression analysis was performed... Figure 7 Eleven radiomics features with non-zero coefficients were preserved from the diseased intestinal wall and mesenteric fat of CD patients in the training cohort. Figure 8 This data was used to construct a radiomics model based on random forest. Five features were derived from mesenteric adipose tissue, and six features were derived from the intestinal wall itself. Notably, mesenteric adipose tissue features occupied three of the top five positions in importance, suggesting that the texture heterogeneity of periintestinal fat may be closely related to inflammatory activity in Crohn's disease.

[0117] like Figure 7 As shown, Figure 7 The following is a summary of the radiomics features of the diseased intestinal wall and mesenteric fat screened using minimum absolute contraction and selection operator regression analysis provided in this embodiment of the invention: (a) Binomial deviation curve, with a regularization parameter (λ) of 0.0168 selected using the minimum criterion ten-fold cross-validation. (b) LASSO coefficient path diagram, where 11 radiomics features with non-zero coefficients were retained when the regularization parameter (λ) was 0.0168.

[0118] like Figure 8 As shown, Figure 8 The feature coefficient histogram provided for embodiments of the present invention shows the names and coefficients of 11 radiomics features of the diseased intestinal wall and surrounding mesenteric adipose tissue.

[0119] This study involves ten major machine learning models: LR, NaiveBayes, SVM, RF, ExtraTrees, XGBoost, LightGBM, GradientBoosting, AdaBoost, and MLP. In the training set, the AUC values ​​of the 10 machine learning models were 0.977 (95% CI 0.956-0.997), 0.966 (95% CI 0.936-0.996), 0.970 (95% CI 0.944-0.997), 0.979 (95% CI 0.958-1.000), 0.997 (95% CI 0.994-1.000), 0.997 (95% CI 0.993-1.000), 0.859 (95% CI 0.800-0.919), 0.997 (95% CI 0.992-1.000), 0.999 (95% CI 0.996-1.000), and 0.973 (95% CI 0.950-0.996). In the test set, the AUC values ​​of the 10 machine learning models were 0.913 (95% CI 0.841-0.985), 0.905 (95% CI 0.827-0.983), 0.916 (95% CI 0.845-0.986), 0.926 (95% CI 0.861-0.990), 0.921 (95% CI 0.852-0.990), 0.895 (95% CI 0.815-0.974), 0.783 (95% CI 0.677-0.890), 0.881 (95% CI 0.797-0.966), 0.908 (95% CI 0.834-0.982), and 0.908 (95% CI 0.831-0.984).

[0120] like Figure 9 As shown, Figure 9 ROC curves of ten machine learning models in the training set (a) and validation set (b) provided for embodiments of the present invention.

[0121] 2.3 Clinical-omics hybrid model and nomogram

[0122] By further combining radiomics model scores with clinical models, a combined clinical-radiomics model was established using logistic regression. To improve clinical operability, nomograms were used to present the combined model. ROC curves, calibration curves, and Delong tests were performed for the clinical, radiomics, and combined models. There was no significant difference in AUC between the radiomics and clinical models (p=0.211). The combined model significantly improved the AUC compared to the clinical model (p=0.044). There was no significant difference between the radiomics and combined models (p=0.946). This suggests that using radiomics features may be more valuable than using clinical features alone, and the combined radiomics + clinical model statistically significantly improves upon using clinical features alone.

[0123] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A multicenter study method for assessing Crohn's disease activity, characterized in that, The method includes: S1: Multi-region lesion segmentation in radiomics model: The training set and validation set were drawn using the Draw function of segmenteditor in the open-source medical imaging software 3D-Slicer. The most severely affected intestinal segment of each patient was selected as the region of interest (VOI), avoiding the intestinal lumen and mesentery. The Hollow function is used to delineate the mesenteric tissue within a 5mm radius around the intestinal wall, and adjacent intestinal segments, lumens, or organs within the VOI are manually erased as the feature extraction region of interest (VOI) for establishing the radiomics model. Two radiologists used the same tools and settings to use multiple patients suspected of having Crohn's disease as a training set. Then, 30 patients randomly selected from the training set were used for additional region of interest (VOI) segmentation to select candidate radiomics features. S2: Feature extraction: Radiomic features of the intestinal wall and mesentery were extracted from CT images using the PyRadiomics open-source Python package to standardize the image spatial resolution; Nonlinear intensity transformation is performed on image voxels; Gaussian Laplacian operators are used for filtering with sigma values ​​of 1, 2, and 3 mm; and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) are applied to first-order statistical and texture features. The fixed bin width is set to 5 HU to extract radiomics features of the original CT image and segmented lesions based on discrete voxel intensity values, including shape-based features, first-order features, and texture features, to describe the morphological features and internal and surface texture features of the region of interest (VOI). S3: Feature selection: Use the Z-score method to standardize the features and calculate the mean and standard deviation of each feature vector; S4: Model Building: A radiomics risk model was built using ten machine learning algorithms, including Support Vector Machine, Logistic Regression, Random Forest, Naive Bayes, Extreme Random Tree, Extreme Gradient Boosting, Lightweight Gradient Boosting Machine, Gradient Boosting, Adaptive Boosting, and Multilayer Perceptron. S5: Multimodal Clinical Construction: Feature selection was performed using recursive feature elimination, ANOVA, Kruskal-Wallis test, and Relief algorithm, retaining 15 features strongly associated with the outcome; logistic regression with minimum absolute contraction and selection operator constraints was used as the classifier; 10-fold cross-validation was performed on the training dataset, hyperparameters were set based on the model's performance on the validation dataset, and the model with better performance on the test set was selected as the clinical model; S6: Hybrid Model Construction: A clinical-radiomics feature model is constructed using the same machine learning algorithm and the aforementioned radiomics features. The ten-fold cross-validation method is used to select the model with the best performance on the cross-validation set as the clinical-radiomics feature model, and a nomogram is plotted. S7: Model Evaluation and Comparison: The diagnostic efficacy of the clinical-radiomics model in differentiating active and remission CD in the training and testing cohorts was evaluated using the area under the receiver operating characteristic curve (AUC). Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated at the critical value for maximizing the Youden index. S8: Statistical methods: All data analysis was performed using MedCalc and Python 3.7.12 on the OnekeyAI platform version 4.9.

1. Statistical evaluation was performed using statsmodels, and radiomics features were extracted using PyRadiomics. Machine learning implementations, including support vector machines, were performed using Scikit-learn.

2. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that: The training and validation sets for delineation include complete diseased intestinal segments from the beginning to the end of the lesion. If a patient in remission has no obvious diseased intestinal segments, the ileocecal region is selected as the delineation object.

3. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, The two radiologists used the same tools and settings to perform additional region of interest (VOI) segmentation on 30 randomly selected patients from the training set, and the steps for selecting candidate radiomics features included: Two radiologists used the same tools and settings to perform additional region of interest (VOI) segmentation on 30 patients randomly selected from the training set, using multiple patients suspected of having Crohn's disease as a training set. The time interval between the two radiologists' two image readings exceeded 6 months; the consistency of the two radiologists' delineation results was evaluated by the intra- and inter-group correlation coefficients (ICCs); features with intra- and inter-group correlation coefficients (ICCs) greater than 0.75 were selected as candidate radiomics features.

4. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, The step of using the PyRadiomics open-source Python package to extract radiomics features of the intestinal wall and mesentery from CT images to standardize image spatial resolution includes: The PyRadiomics open-source Python package was used to automatically extract 1834 Radiomics features from CT images. A linear interpolation algorithm was then used to resample the voxel size of the CTE images and their corresponding 3D-VOIs to 1×1×1 mm³ (x,y,z) isotropic voxels to normalize the image spatial resolution.

5. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that: The texture features include: gray-level co-occurrence matrix, gray-level dependency matrix, gray-level running length matrix, gray-level size region matrix, and neighboring gray-level difference matrix.

6. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, The steps of standardizing features using the Z-score method and calculating the mean and standard deviation of each feature vector include: Each feature vector is standardized by subtracting its mean and dividing it by its standard deviation. Each vector has zero mean and unit standard deviation. To take advantage of the high-dimensionality of the feature space, an association matrix is ​​constructed by calculating the similarity of each pair of features. If the Pearson correlation coefficient (PCC) of a feature pair is greater than 0.9, one of the features is deleted. The maximum-minimum redundancy algorithm is used to select features, and then the minimum absolute shrinkage and selection operator algorithm is used to perform feature selection.

7. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, Following the construction of the radiomics risk model, the following is also included: The radiomics score for each patient is calculated by linearly combining and weighting the selected features; the hyperparameters of the radiomics risk model are determined based on the performance of the radiomics risk model on the validation dataset.

8. The multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, After calculating accuracy, sensitivity, specificity, positive predictive value, and negative predictive value at the critical value where the Youden index value is maximized, the method further includes: Estimate 95% confidence intervals using a bootstrap method with 1000 samples; The consistency between the predicted probabilities of the clinical-radiomics model and the actual results was evaluated using calibration curves, and the reliability of the clinical-radiomics feature model was verified by the Hosmer-Lemeshow goodness-of-fit test. The calibration capability of the clinical-radiomics characterization model was evaluated using calibration curves. The DeLong test was used to compare the AUC values ​​of different clinical-radiomics feature models.

9. A multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, After the steps of implementing machine learning, including support vector machines, using Scikit-learn, the method further includes: The Shapiro-Wilk test was used to test the normality of clinical characteristics; continuous variables were expressed as mean and variance, and component comparisons were performed using t-tests or Mann-Whitney U tests based on their distribution characteristics. Categorical variables are presented as frequencies and percentages, and intergroup comparisons are performed using the chi-square test or Fisher's exact test.

10. A multicenter study method for assessing Crohn's disease activity according to claim 1, characterized in that, After the categorical variables are presented in the form of frequencies and percentages, and between-group comparisons are performed using a chi-square test or Fisher's exact test, the following steps are also included: Univariate analysis was performed on radiological characteristics with good consistency within and between observers and with correlation coefficients (ICCs) > 0.

75. Two-sided P < 0.05 was considered statistically significant. The diagnostic efficacy of the model was evaluated using receiver operating characteristic (ROC) curves, and the area under the curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated to quantify the diagnostic efficacy.

Images