Nodal-based radiomics analysis for identifying cervical lymph node metastasis at levels I and II in patients with oral squamous cell carcinoma using contrast-enhanced computed tomography.

Abstract

Discriminating metastatic from benign cervical lymph nodes (LNs) in oral squamous cell carcinoma (OSCC) patients using pretreatment computed tomography (CT) has been controversial. This study aimed to investigate whether CT-based texture analysis with machine learning can accurately identify cervical lymph node metastasis in OSCC patients. Twenty-three patients (with 201 cervical LNs [150 benign, 51 metastatic] at levels I-V) who underwent preoperative contrast-enhanced CT and subsequent cervical neck dissection were enrolled. Histopathologically proven LNs were randomly divided into the training cohort (70%; n = 141, at levels I-V) and validation cohort (30%; n = 60, at level I/II). Twenty-five texture features and the nodal size of targeted LNs were analyzed on the CT scans. The nodal-based sensitivities, specificities, diagnostic accuracy rates, and the area under the curves (AUCs) of the receiver operating characteristic curves of combined features using a support vector machine (SVM) at levels I/II, I, and II were evaluated and compared with two radiologists and a dentist (readers). In the validation cohort, the AUCs (0.820 at level I/II, 0.820 at level I, and 0.930 at level II, respectively) of the radiomics approach were superior to three readers (0.798-0.816, 0.773-0.798, and 0.825-0.865, respectively). The best models were more specific at levels I/II and I and accurate at each level than each of the readers (p < .05). Machine learning-based analysis with contrast-enhanced CT can be used to noninvasively differentiate between benign and metastatic cervical LNs in OSCC patients. • The best algorithm in the validation cohort can noninvasively differentiate between benign and metastatic cervical LNs at levels I/II, I, and II. • The AUCs of the model at each level were superior to those of multireaders. • Significant differences in the specificities at level I/II and I and diagnostic accuracy rates at each level between the model and multireaders were found.