Can MRI radiomics predict neck metastasis at initial diagnosis in patients with squamous cell carcinoma of the tongue?

Abstract

Objectives: This study aimed to predict the risk of neck metastasis in patients with tongue squamous cell carcinoma by performing radiomics analysis of preoperative magnetic resonance (MR) images.

Methods: In total, 143 patients with primary tongue cancer were enrolled and divided into training and validation groups. The presence of neck metastases was assessed after at least 6-month of follow-up. Using fat-suppressed T2-weighted images, two observers manually set the volume of interest at the tumor site using the 3D Slicer software and extracted 107 image features. The analysis was based on significant differences between patient groups in the presence or absence of neck metastases by the Mann-Whitney U test and good agreement with intra- and inter-observer intra-class correlation coefficients exceeding 0.9. In addition, two characteristics that were determined to be very useful for diagnosing neck metastases were selected by receiver operating characteristic analysis and evaluated for goodness of fit in the validation data.

Results: Neck metastases were identified in 57 of 143 patients and divided into 121 training and 22 validation datasets. Using the combined criteria of Major Axis Length of 3D-shape features and Joint Entropy of the gray-level co-occurrence matrix, neck metastases were identified in 80.7% of the cases; the validation data predicted neck metastases in 80% of the cases.

Conclusions: MR-imaging texture analysis of the primary tumor helps predict neck metastasis in patients with tongue cancer. The proposed criteria are simple yet useful for identifying groups of patients who may require neck dissection.