Radiomics analysis of intraoral ultrasonographic images for prediction of late cervical lymph node metastasis in patients with tongue cancer: influence of marginal region.

Abstract

Objectives: To investigate the predictability of late cervical lymph node metastasis using radiomics analysis of ultrasonographic images of tongue cancer.

Methods: We selected 128 patients with tongue cancer who underwent intraoral ultrasonography at the pre-treatment, 35 of whom had late cervical lymph node metastasis. Radiomics analysis was used to extract and quantify the image features. Segmentations were performed on three regions: the hypoechoic region with a 3-mm margin (0 + 3-mm margin), the hypoechoic region alone (0-mm margin), and the 3-mm region surrounding the hypoechoic area (3-mm margin). Support vector machine (SVM) and neural network (NNT) were used as the machine learning models, and sensitivity, specificity, and area under the curve (AUC) from the receiver operating characteristic curves were determined for diagnostic performances.

Results: The AUC values in the test group were 0.893, 0.929, and 0.679 for the SVM models with 0 + 3-, 0-, and 3-mm margins, respectively. The AUC values in the test group were 0.905, 0.952, and 0.821 for the NNT models with 0 + 3-, 0-, and 3-mm margins, respectively.

Conclusions: Radiomics analysis and machine learning models using ultrasonographic images of pretreated tongue cancer with a hypoechoic area (0-mm margin) could be the best models to predict late cervical lymph node metastasis.

Advances in knowledge: This study makes a significant contribution to the tongue cancer treatment because radiomics analysis and machine learning models using ultrasonographic images of before the primary treatment for the tongue cancer could predict late cervical lymph node metastasis with high accuracy.