Detection of Cavities from Oral Images using Convolutional Neural Networks

The task of automatic dental cavity detection from oral images has been recently revolutionized due to several works that uses convolutional neural networks (CNNs). One of the challenges that faces this research is the ability of these systems to perform well on images collected from varied sources. Hence, this work proposes a CNN and also studies its ability to perform on images collected from various sources. For this, a custom-made CNN called Dental-Net is proposed to automatically detect cavities in the teeth from oral photographic images. A dataset with a total of 609 diverse images consisting of teeth with and without cavities, shot from varied positions and orientations are collected from a variety of sources and by web scraping. The images from the dataset are resized, converted to grayscale, followed by normalization and on-the-fly data augmentation. The dataset is then used to train the Dental-Net model. In this study, the performance of four pre-trained models namely VGG16, MobileNetV2, InceptionV3, and ResNetlS have also been evaluated. The results show that the Dental-Net model achieves the highest accuracy of 94.25% and 91.09% on the training and the validation set respectively. Comprehensive analysis of this study reveals positive results that can be improved in the future and be implemented on a commercial scale.