Application of radiomics combined with machine learning algorithms for preoperative prediction of perineural invasion in oral squamous cell carcinoma
10.12016/j.issn.2096-1456.202660032
- Author:
MENG Xiangze
1
;
YUAN Ying
2
;
YANG Xi
1
Author Information
1. Department of Oromaxillofacial Head and Neck Oncology, Shanghai Jiao Tong University School of Medicine, College of Stomatology, Shanghai Jiao Tong University, National Center for Stomatology, National Clinical Research Center for Oral Diseases, Shanghai Key Laboratory of Stomatology, Shanghai Research Institute of Stomatology
2. Department of Radiology, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine
- Publication Type:Journal Article
- Keywords:
oral squamous cell carcinoma;
perineural invasion;
contrast-enhanced computed tomography;
ra⁃diomics;
machine learning;
LightGBM;
preoperative prediction;
calibration curve;
decision curve analysis
- From:
Journal of Prevention and Treatment for Stomatological Diseases
2026;34(5):456-470
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the value of contrast-enhanced computed tomography (CT) radiomics combined with machine learning algorithms in the preoperative prediction of perineural invasion (PNI) in oral squamous cell carcinoma (OSCC), aiming to provide evidence for assisting clinical treatment decision-making.
Methods:This study was approved by the Ethics Committee of the Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine. A total of 250 OSCC patients confirmed by postoperative pathology were included, comprising 128 PNI-positive and 122 PNI-negative cases. The dataset was randomly divided into training (n=175), validation (n=38), and independent testing (n=37) sets in a ratio of 7:1.5:1.5. Regions of interest were delineated on preoperative images, and radiomic features were extracted. After dimensionality reduction and feature selection using methods like Least Absolute Shrinkage and Selection Operator (LASSO) regression, multiple machine learning models, including support vector machine (SVM), random forest, Light gradient boosting machine (LightGBM), and a Stacking ensemble model, were constructed. Model performance was evaluated using metrics such as the area under the receiver operating characteristic curve (AUC), sensitivity, calibration curves, and decision curve analysis (DCA). Model interpretability was analyzed using Shapley additive explanations (SHAP) and grouped permutation feature importance analysis.
Results : Among the 250 samples analyzed, the LightGBM model based on radiomics demonstrated the best performance on the independent test set, with an AUC of 0.781, outperforming models like SVM (AUC = 0.730) and Random Forest (AUC = 0.691), as well as clinical models (AUCs ranging 0.549-0.711). The LightGBM model showed good calibration (Brier score 0.198), and DCA indicated high clinical net benefit over a wide threshold probability range. Paired DeLong tests revealed no statistically significant differences in AUC between the ensemble (Stacking) model and the corresponding best-performing radiomics-based model. SHAP analysis and grouped permutation feature importance analysis further indicated that the primary discriminative information for the model came from radiomic texture features.
Conclusion :The LightGBM model based on contrast-enhanced CT radiomics demonstrated good discriminative ability for preoperative prediction of PNI in OSCC. In the independent test set, it achieved the highest AUC. This model holds promise as a non-invasive auxiliary tool for preoperative risk assessment. Given the limited sample size of the independent test set, these results require further validation in larger cohorts and external datasets.
