Interpretable ensemble learning model using intratumoral and peritumoral multi-sequence MR-radiomics predicts high-grade cervical cancer with lymphovascular space invasion

Abstract

Objective To develop an interpretable ensemble learning model for predicting dual positivity of pathological grading and lymphovascular space invasion (HGVI) in cervical cancer (CC) by integrating multiple machine learning algorithms, fusion strategies, and multi-sequence MR radiomics. Methods A total of 242 CC patients who underwent preoperative MRI were retrospectively enrolled and randomly divided into a training set (n = 169) and a test set (n = 73). Volumes of interest were manually delineated on intratumoral and peritumoral (3 mm and 5 mm) regions across three MR sequences of apparent diffusion coefficient (ADC), T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI). Six machine learning classifiers were employed, and the optimal base models were subsequently integrated via mean, weighted averaging, and logistic regression-based stacking (ENMLR). Model performance was evaluated using the area under the curve (AUC), confidence intervals (CI), and decision curve analysis (DCA). Interpretability was examined through SHAP, correlation, and restricted cubic spline (RCS) analyses. Results Compared with the eight models, the ENMLR achieved superior performance in the training set, which was robustly validated in the test set via 1,000 Bootstrap resamples (AUC: 0.858, 95% CI: 0.762–0.932). In the test set, the ENMLR demonstrated good accuracy (0.7260), a Brier score of 0.1523, and enhanced clinical net benefit across a wide threshold range (10–90%). Correlation analyses revealed positive agreement between intratumoral and peritumoral radiomics (Spearman’s r: 0.15, 0.21, 0.41, and 0.29; all P < 0.05). RCS analysis further identified a significant nonlinear relationship between ADC sequence-derived intratumoral radiomics scores and an increased incidence of HGVI double positivity (Non-linear P < 0.05; Overall P < 0.05), with a notable inflection point. SHAP analysis identified ADC_peritumoral_3mm and the Xgboost classifier as the top contributors to the fusion model. Conclusions By integrating diverse learning algorithms with multi-sequence MR radiomics (intratumoral and peritumoral) and serum biomarkers, the ENMLR demonstrates robust discriminative performance for predicting HGVI double positivity in CC. Nonetheless, further multicenter prospective studies are warranted to validate these findings.