Machine learning based on magnetic resonance imaging and clinical parameters helps predict mesenchymal-epithelial transition factor expression in oral tongue squamous cell carcinoma: a pilot study

Abstract

Objectives: This study aimed to develop machine learning models to predict phosphorylated mesenchymal-epithelial transition factor (p-MET) expression in oral tongue squamous cell carcinoma (OTSCC) using magnetic resonance imaging (MRI)-derived texture features and clinical features.

Methods: Thirty-four patients with OTSCC were retrospectively collected. Texture features were derived from preoperative MR images, including T2WI, apparent diffusion coefficient mapping, and contrast-enhanced (ce)-T1WI. Dimension reduction was performed consecutively with reproducibility analysis and an information gain algorithm. Five machine learning methods-AdaBoost, logistic regression (LR), naïve Bayes (NB), random forest (RF), and support vector machine (SVM)-were adopted to create models predicting p-MET expression. Their performance was assessed with fivefold cross-validation.

Results: In total, 22 and 12 cases showed low and high p-MET expression, respectively. After dimension reduction, 3 texture features (ADC-Minimum, ce-T1WI-Imc2, and ce-T1WI-DependenceVariance) and 2 clinical features (depth of invasion [DOI] and T-stage) were selected with good reproducibility and best correlation with p-MET expression levels. The RF model yielded the best overall performance, correctly classifying p-MET expression status in 87.5% of OTSCCs with an area under the receiver operating characteristic curve of 0.875.

Conclusion: Differences in p-MET expression in OTSCCs can be noninvasively reflected in MRI-based texture features and clinical parameters. Machine learning can potentially predict biomarker expression levels, such as MET, in patients with OTSCC.