An Interpretable MRI-Based Machine Learning Model for Preoperative Identification of the Vascular Dissemination Phenotype in Hepatocellular Carcinoma: A Dual-Center Study

Multiparametric MRI Radiomics-Based Interpretable Machine Learning Model for the Prediction of Lymphovascular Space Invasion in Endometrial Cancer.

RATIONALE AND OBJECTIVES To develop and validate a machine learning model combining multiparametric Magnetic Resonance Imaging (MRI) radiomics and clinical indicators for predicting lymphovascular space invasion (LVSI) in endometrial cancer (EC). MATERIALS AND METHODS This retrospective study enrolled EC patients who underwent preoperative MRI at two centers. Of 567 initially screened patients, 408 were included per inclusion/exclusion criteria, divided into training and validation sets by hospital. Clinical risk factors and intratumoral/peritumoral radiomic features were identified. Six machine learning algorithms were used to build models; the one with the highest validation Area Under the Curve (AUC) was optimal. Five additional models were developed, and performance was evaluated via AUC, calibration curves, and decision curve analysis (DCA). RESULTS Logistic regression identified CA125 and tumor diameter as independent LVSI risk factors. Six machine learning models were built with CA125, tumor diameter, Rad_Score1 and Rad_Score2; the NeuralNetwork performed best (validation AUC=0.803). The combined clinical-intratumoral-peritumoral radiomics model achieved the highest AUC(training AUC = 0.863, validation AUC = 0.803), with good calibration (Hosmer-Lemeshow test, P > 0.05) and favorable net clinical benefit (threshold 0.1-0.7). SHapley Additive exPlanations (SHAP) analysis enhanced model interpretability. CONCLUSION This study systematically compared the predictive performance of six machine learning models for LVSI in EC, identifying the NeuralNetwork model as the most optimal. The combined clinical-intratumoral-peritumoral radiomics model, alongside its SHAP visualization tool, enhanced the accuracy (ACC) of non-invasive preoperative LVSI prediction, demonstrating certain potential for clinical application.

Preoperative prediction of microvascular invasion in hepatocellular carcinoma ≤5 cm based on contrast-enhanced ultrasound features and LI-RADS categorization: a multicenter study

Objectives To investigate the predictive value of contrast-enhanced ultrasound (CEUS) features combined with Liver Imaging Reporting and Data System (LI-RADS) categorization for microvascular invasion (MVI) in hepatocellular carcinoma (HCC) ≤5 cm. Methods This multicenter retrospective study enrolled adult patients with HCC ≤5 cm who underwent CEUS between January 2018 and December 2025. Least absolute shrinkage and selection operator (LASSO) regression and multivariate logistic analysis were employed to screen risk factors and establish the MVI prediction model. Three models were developed: a clinical model, an ultrasound model, and a combined model. The performance of the combined model was evaluated and validated using the area under the receiver operating characteristic (AUC), calibration curves, decision curve analysis (DCA), and the Hosmer-Lemeshow test. Results A total of 261 patients with HCC ≤5 cm were enrolled. Patients were divided into a derivation cohort (n=209) and an external validation cohort (n=52). 85 patients (32.57%) were MVI-positive. LASSO regression and multivariate analysis revealed that AFP, tumor margin, enhanced homogeneity, mosaic, and LI-RADS were significantly associated with MVI. The combined model showed an AUC of 0.880 (95% CI: 0.832–0.929) in the derivation cohort and 0.832 (95% CI: 0.703–0.960) in the external validation cohort. Calibration curves revealed excellent agreement between the model’s predicted probability of MVI and the actual observed outcomes. DCA confirmed the excellent net benefits. Conclusion This model can noninvasive preoperative prediction of MVI risk in patients with HCC ≤5 cm, offering a reliable reference for clinicians in developing personalized treatment strategies.

Machine learning modeling to predict HCC locations in cirrhotic patients undergoing MRI - a proof-of-concept study.

An MRI Radiomics-habitat-clinical Model for Noninvasive Prediction of IDH Mutation in Gliomas With Bioinformatic Correlation: Multicenter Development With External Validation.

RATIONALE AND OBJECTIVES To develop and externally validate a multicenter magnetic resonance imaging (MRI)-based model integrating radiomics, intratumoral habitat features, and clinical variables for noninvasive prediction of isocitrate dehydrogenase (IDH) mutation status in gliomas, and to explore its biological relevance. METHODS This multicenter retrospective study included 1076 patients with histopathologically confirmed gliomas from four sets. Preoperative T2-weighted imaging, contrast-enhanced T1-weighted imaging, and diffusion-weighted imaging images were analyzed. Conventional radiomic features and habitat features derived from voxel-wise clustering of multimodal MRI were extracted. After ComBat harmonization and least absolute shrinkage and selection operator selection, logistic regression models were developed in a training set and validated in three independent external sets. A nomogram was constructed for individualized prediction. Model performance was evaluated using area under the curve (AUC), calibration, and decision curve analysis. Prognostic value was assessed by Kaplan-Meier analysis, and radiogenomic validation was performed using transcriptomic data. RESULTS The combined radiomics-habitat-clinical model achieved high performance, with an AUC of 0.952 in the training set and 0.867 to 0.922 in the external test sets. The habitat model showed comparable discrimination, highlighting the robustness of habitat-derived features. SHapley Additive exPlanations analysis identified the habitat-score as the most influential predictor. The nomogram showed good calibration and clinical utility. Transcriptomic analysis of model-defined groups showed IDH-related differences in proliferation, metabolism, extracellular matrix remodeling, and immune-related pathways, providing indirect biological support for the imaging-derived signatures. CONCLUSION Integrating habitat-based analysis with radiomics enables accurate and biologically interpretable prediction of IDH mutation status in gliomas, supporting its potential for preoperative molecular stratification.

Nomogram for predicting vessels encapsulating tumor clusters in hepatocellular carcinoma using DCE-CT based intratumoral heterogeneity quantification and body fat distribution: A multicenter study.

BACKGROUND Vessels encapsulating tumor clusters (VETC) represent a distinct vascular architecture associated with metastasis in hepatocellular carcinoma (HCC). This study aimed to develop and externally validate a nomogram integrating intratumor heterogeneity (ITH), radiologic, clinical, and body composition features for preoperative VETC prediction. METHODS This two-center retrospective study included 360 patients with pathologically confirmed VETC status between January 2017 and June 2024. Tumors were clustered using k-means with the elbow method to derive an ITH Score quantifying spatial heterogeneity. Based on dynamic contrast-enhanced computed tomography (DCE-CT), five machine learning classifiers constructed four DCE-CT-based models: clinical-radiological-body composition (CRB), ITH Score, Rad Score, and an integrated nomogram. Performance was evaluated using area under the curve (AUC), calibration, and decision curve analysis (DCA). A two-round reader study assessed nomogram assistance. Recurrence-free survival (RFS) was analyzed using Kaplan-Meier methods. RESULTS Independent predictors included intratumor necrosis, non-smooth tumor margin, total adipose tissue index (TATI), and portal venous phase ITH Score. Using a logistic regression (LR) algorithm, the integrated nomogram showed the highest discrimination in the internal and external test sets (AUCs, 0.847 and 0.814, respectively), with numerically higher AUCs than the other models. Calibration and DCA demonstrated good agreement and clinical utility. Nomogram assistance improved accuracy, particularly for the junior radiologist. VETC- patients exhibited significantly longer RFS than VETC + patients (P < 0.0001), while higher nomogram scores correlated with shorter RFS (log-rank P < 0.05). CONCLUSION The integrated nomogram enables preoperative VETC prediction, improves radiologists' performance, and facilitates prognostic stratification in HCC.

Predictive Value of Nomogram-based Multiparametric MRI Combined with Pathological Biomarkers for HIF-1α Expression in Breast Cancer.

RATIONALE AND OBJECTIVES This study aimed to assess the predictive value of clinicopathological characteristics, conventional magnetic resonance imaging (MRI), intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI), and dynamic contrast-enhanced MRI (DCE-MRI) parameters for hypoxia-inducible factor-1α (HIF-1α) expression in breast cancer. MATERIALS AND METHODS We retrospectively enrolled 146 breast cancer patients receiving preoperative multiparametric MRI and surgery from 2019 to 2023, who were randomly assigned into training (n = 103) and validation (n = 43) cohorts at 7:3 ratio. Multivariate logistic regression and receiver operating characteristic (ROC) curve analyses were conducted, and a nomogram was constructed based on independent predictive factors. RESULTS The high-expression group had a higher proportion of axillary lymph node metastasis (ALN_metastasis), advanced histological grades, unclear margin, time intensity curve (TIC)-III type, lower D values, and higher Ktrans and Kep values compared with the low-expression group (P < 0.05). The area under the curves(AUCs) for the pathological, conventional MRI, IVIM-DWI, DCE-MRI, and combined models (ALN_metastasis + TIC type + D + Kep) were 0.765, 0.732, 0.771, 0.804, and 0.958 in the training cohort, respectively. The combined model significantly outperformed individual models (combined model vs. conventional MRI model, Z = 4.890, P < 0.001; combined model vs. pathological model, Z = 4.429, P < 0.001; combined model vs. IVIM model, Z = 3.724, P < 0.001; combined model vs. DCE-MRI model, Z = 3.691, P < 0.001). CONCLUSION The nomogram combining clinicopathological and multimodal MRI parameters can accurately predict HIF-1α expression non-invasively and assist personalized breast cancer therapy.