MRI-based radiomics combined with clinical indicators for preoperative prediction of cervical lymph node metastasis in oral squamous cell carcinoma

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Objective To develop and validate a nomogram that integrates magnetic resonance imaging (MRI) radiomic features with clinical indicators, including the platelet-to-lymphocyte ratio (PLR) and depth of invasion (DOI), for the preoperative prediction of cervical lymph node metastasis (LNM) in oral squamous cell carcinoma (OSCC). Methods This retrospective study included 148 OSCC patients, randomly split 7:3 into training cohort ( n = 103) and test cohort ( n = 45). Two head and neck radiologists delineated tumor volumes of interest (VOIs)on T2-weighted imaging (T2WI ) and gadolinium-enhanced T1-weighted imaging(Gd-T1WI).Radiomic features were extracted using the Pyradiomics package. Feature selection was performed using the Mann-Whitney U test, recursive feature elimination (RFE), and least absolute shrinkage and selection operator (LASSO) regression to construct a radiomic score (Radscore). Independent clinical predictors of LNM were identified via univariate and multivariate logistic regression to build a clinical model. A combined predictive nomogram was subsequently developed by integrating the Radscore with these independent clinical predictors. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, calibration curves, and decision curve analysis (DCA). Results Multivariate analysis identified PLR ≥ 111.385 and DOI (measured on Gd-T1WI) as independent predictors of LNM (both P < 0.05). Fourteen key features were selected from 2,048 initial features extracted from MRI sequences to construct the radiomic model. The areas under the ROC curve (AUC) for the clinical model were 0.760 (training) and 0.694 (test), and those for the radiomic model were 0.840 (training) and 0.725 (test). The combined model (incorporating Radscore, PLR, and DOI) demonstrated superior predictive performance, with AUCs of 0.892 (training) and 0.804 (test), significantly outperforming the single models (DeLong test, P < 0.05). The calibration curve indicated good agreement between predicted and observed LNM probabilities. DCA showed a higher net clinical benefit for the combined model than for the single models across a wide range of threshold probabilities. Conclusion The nomogram combining MRI radiomic features, PLR, and DOI effectively predicts cervical LNM in OSCC patients preoperatively, providing a valuable tool for individualized treatment planning.