Room: AAPM ePoster Library
Purpose: develop non-invasive grading models of parotid gland cancer (PGC) malignancy based on radiomic features in preoperative magnetic resonance (pMR) images using five conventional machine learning (cML) and five deep convolutional neural network (DCNN).
Methods: We retrospectively collected the pMR images of 39 PGC patients with 20 high- and 19 intermediate- plus low-malignancy grades. A total of 829 two-dimensional PGC images were cropped from T1- and T2-weighted pMR images. The PGC images were divided into 70% for a training dataset and 30% for a test dataset. High- versus intermediate- plus low-malignancy grades were predicted using radiomic MR features with five cML (logistic regression (LR), random forest (RF), k-nearest neighbors (k-NN), support vector machine (SVM), and artificial neural network (ANN)) and five pre-trained DCNNs (AlexNet, GoogLeNet, VGG-16, ResNet-101, and DenseNet-201). A total of 972 hand-crafted radiomic features were extracted from PGC images, and then radiomic biomarkers were obtained by a least absolute shrinkage and selection operator (LASSO) in the training dataset for five cML models. Last three layers of the pre-trained DCNNs were replaced with new three layers for the prediction task, and they are fine-tuned. The prediction performances of the 10 models were evaluated by the accuracy and the area under the receiver operating characteristic curve (AUC).
Results: The VGG-16-based model demonstrated the best prediction performance (accuracy: 85.4%, AUC: 0.906) among the 10 models. Moreover, all DCNN-based malignancy grading models showed the higher accuracy than the conventional histological diagnosis approach by fine needle aspiration cytology (FNAC) of 73.7% in our hospital.
Conclusion: The VGG-16-based model could be feasible for noninvasively grading of PGC malignancy using pMR images.