Optimization methods for deep neural networks classifying OCT images to detect dental caries

Dental caries are common chronic infectious oral diseases affecting most teenagers and adults worldwide. Optical coherence tomography (OCT) has been studied extensively for the detection of early carious lesions. Deep learning techniques are a rapidly emerging new area of biomedical research and have yielded impressive results in diagnosis and prediction in the field of oral radiology. Deep learning models particularly deep convolutional neural networks (CNN) can be employed along with OCT imaging system to more accurately identify early dental caries. In this work, after OCT data acquisition, data augmentation was performed to obtain a large amount of training data in order to effectively learn, where collection of such training data is often expensive and laborious. For the backpropagation process, seven optimization methods, namely Adadelta, AdaGrad, Adam, AdaMax, Nadam, RMSProp, and Stochastic Gradient Descent (SGD) were utilized to improve the accuracy of a CNN classifier for diagnosing dental caries. In this study, 75% of the data were utilized for training and 25% for testing. The diagnostic accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and receiver operating characteristic (ROC) curve were calculated for detection and diagnostic performance of the deep CNN algorithm. This study highlighted the performance of various optimization methods for deep CNN models with OCT images to detect dental caries.