Classification of salivary gland tumors in optical coherence tomography images based on deep learning

Abstract

Abstract The diagnosis of oral diseases mainly relies on visual examination by doctors with clinical experience. Histopathological examination is still the gold standard of oral disease diagnosis, but it is invasive and time-consuming. In recent years, optical coherence tomography (OCT) has played an important role in the field of biomedicine with its unique advantages of non-invasiveness, high resolution, real-time and three-dimensional imaging, which can be well applied to the imaging of oral lesions. In this paper, four deep learning (DL) models including LeNet-9, VGG-16, ResNet-18 and ResNet-50 were used to classify oral tumors including two benign and two malignant salivary gland tumors (SGTs), which were imaged by our home-made swept-source OCT. The results indicated that ResNet-18 has the best classification performance, with accuracy, precision, recall (sensitivity), F1 score and specificity all above 98%. Then, we analyzed the visualization process of DL and explored how the DL model extracts features. It is demonstrated that the DL model has a good clinical auxiliary role in the classification of SGTs.