Preoperative and Non-Invasive Approach for Radiomic Biomarker-Based Prediction of Malignancy Grades in Patients with Parotid Gland Cancer in Magnetic Resonance Images

We have investigated a non-invasive approach for predicting parotid gland cancer (PGC) malignancy grade based on radiomic biomarkers in preoperative magnetic resonance (pMR) images using six conventional machine learning (cML) and five deep learning (DL) algorithms. 39 patients were divided into 70% (27 patients) for a training dataset and 30% (12 patients) for a test dataset. A total of 972 hand-crafted features were extracted from cancer regions on the twodimensional T1- and T2-weighted pMR images, and then hand-crafted biomarkers were obtained by a least absolute shrinkage and selection operator (LASSO) in the training dataset for six cML models. Five DL models were constructed by transfer learning of pre-trained DL architectures, i.e. AlexNet, GoogLeNet, VGG-16, ResNet-101, and DenseNet-201. Highgrade versus intermediate- plus low-grades malignant PGCs was predicted using the eleven prediction models for the test dataset. The VGG-16-based DL model demonstrated a highest accuracy of 85.4% among the eleven models for the test dataset, which was a higher accuracy than the histological diagnostic accuracy of 79.5% using fine needle aspiration cytology (FNAC). The MR-based DL approach could be feasible for preoperatively and non-invasively predicting the grades of PGC malignancy.